Applied Acoustics 182 (2021) 108289

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

Application of Spectral Kurtosis on vibration signals for the detection

of cavitation in centrifugal pumps
G. Mousmoulis a,⇑, C. Yiakopoulos b, G. Aggidis c, I. Antoniadis b, I. Anagnostopoulos a
a
Laboratory of Hydraulic Turbomachines, School of Mechanical Engineering, National Technical University of Athens, Iroon Polytechniou 9, Zografou, Athens, 15780, Greece
b
Laboratory of Dynamics and Structures, School of Mechanical Engineering, National Technical University of Athens, Iroon Polytechniou 9, Zografou, Athens, 15780, Greece
c
Lancaster University Renewable Energy Group, Department of Engineering, Lancaster University, Bailrigg, Lancaster LA1 4YW, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: The detection of cavitation formation in hydraulic turbomachinery has been widely investigated due to
Received 21 January 2021 its significant impact on their steady and dynamic operation. The aim of this study is the application of
Received in revised form 17 May 2021 Spectral Kurtosis tool in order to effectively detect the impulsive shock waves generated during the
Accepted 2 July 2021
implosion of vapour bubbles. The methodology is applied and evaluated on the vibration signals obtained
Available online 13 July 2021
from two different semi-open impellers. The results under initial cavitating conditions show that the high
frequency implosions of vapour bubbles interact with the low frequency passing of the rotating blades
Keywords:
and compose part of the vibration signal. The application of the band pass filter, with central frequency
Centrifugal pumps
Condition monitoring
and bandwidth estimated from the Fast Kurtogram, to the original signal allows to extract this informa-
Cavitation detection tion both in time and in frequency domain, and to correlate the periodic impulsive behaviour with the
Flow visualisation blade passing frequency of the impeller. The present results support the establishment of a robust detec-
Spectral Kurtosis tion cavitation criterion in centrifugal pumps and show that Spectral Kurtosis is a useful tool for the pre-
Vibration measurements vention of related faults in centrifugal pumps.
Ó 2021 Elsevier Ltd. All rights reserved.

1. Introduction The aforementioned unwanted impacts on the performance of cen-

trifugal pumps has led researchers and engineers to try to develop
The phenomenon of vaporization of liquids at macroscopically techniques for early cavitation detection.
isothermal conditions due to decrease of static pressure below The first method used for the detection of two-phase flow for-
vapour pressure value is called cavitation. Centrifugal pumps are mations in centrifugal pumps, proposes the monitoring of the total
machines that convert mechanical to hydraulic power with the head of the machine and identifies cavitation when the latter drops
use of a rotating impeller with hydrodynamically-shaped blades. more than 3% [4,5]. This drop is due to the development of vapour
The acceleration of the flow and the low pressure conditions pre- formations in the impeller that cover a substantial part of the flow
vailing at the entrance of the rotating impeller can create such local path and start to block the flow, thus reducing the capability of the
conditions that favour the formation of vapour cavities [1]. The impeller to transfer power to the working medium [6–9]. The main
area covered by the two-phase flow can modify the hydrodynamic advantage of this technique is that it constitutes a general detec-
shape of the blades and create flow recirculation and separation tion cavitation criterion, since the head drop is a common symp-
regions that eventually decrease the performance and hydraulic tom of cavitation presence in all types of centrifugal pumps.
efficiency of the pump [2]. In addition, the vapour cavities are While the 3% drop can be identified in the laboratory using rela-
shifted downstream by the flow towards higher pressure areas, tively low cost and precision measuring devices.
where they are instantly imploded, creating high energy micro jets However, application of this method in real operating condi-
and impulsive pressure waves that stress and erode the adjacent tions is not the most efficient approach to detect vapour phased
metallic surfaces [3]. High levels of noise and vibration can be cre- formations, because cavitation even in the earlier stages, before
ated if the phenomenon is intense and extensive in the impeller. the total head drop, can cause significant wear of the impeller sur-
faces, as well as noise and vibration [1–4,10]. As a consequence,
⇑ Corresponding author. additional tools have been developed for the monitoring of hydrau-
E-mail addresses: mousmoulisgeorge@mail.ntua.gr (G. Mousmoulis), chryiako@
lic turbomachines to identify the presence of cavitation, based on
central.ntua.gr (C. Yiakopoulos), g.aggidis@lancaster.ac.uk (G. Aggidis), antogian@ noise and vibration measurements [7,9,11–40]. Signal processing
central.ntua.gr (I. Antoniadis), anagno@fluid.mech.ntua.gr (I. Anagnostopoulos).

https://doi.org/10.1016/j.apacoust.2021.108289
0003-682X/Ó 2021 Elsevier Ltd. All rights reserved.
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

tools are then applied to discern and detect the sound and the the pump response vary periodically with time. When cavitation
mechanical oscillations generated during the implosion of vapour occurs, the cyclic components begin to be masked by the cavitation
cavities. There are two main challenges in the application of these impulsive noise. To overcome these difficulties, a cyclostationary-
tools; i) the extraction of cavitation characteristics from the mea- based analysis [48] could be employed to detect the signature of
sured signal that are usually masked by the noise and the vibra- cyclostationarity at the shaft rotating frequency and the blade pass
tions of the pump and its motor, and ii) the utilisation of these frequency (BPF) and to extract the components related to the mod-
characteristics towards the formation of a general detection crite- ulation mechanism of the flow-induced effects in a centrifugal
rion that would be independent from the individual structural and pump.
operational features of each installation. Towards this direction, the application of wavelet analysis has
The most common signal processing technique in the literature been proposed with the use of microphones [35] and accelerome-
that has been applied on the noise and vibration signals, deals with ters [36,37], where the original signal is decomposed to a number
the use of the Fourier Transformation or with the calculation of of levels in order to bring out the cavitation signal in time. In addi-
Power Spectral Density (PSD), in order to study the frequency spec- tion, Sun et al. [38] applied cyclic spectral analysis on the vibration
trum (FS) of the measured data. The collapse and implosion of the signal and concluded that the onset and development of cavitation
vapour bubbles result in shock pressure waves that excite a wide is correlated with the quadruple of the Rotational Frequency (RF)
range of frequencies and make discernible the development of of the pump motor. Moreover, the study of the signal’s envelope
the phenomenon in the spectrum [9,11–20]. In order to quantify in frequency domain has been proposed by some authors
these excitations, some authors calculate the power content in dif- [19,33,39,40], in order to examine the possible modulation of the
ferent bands of the spectrum and plot the results as function of frequencies excited by cavitation noise and vibration. In the pre-
cavitation development [4,11,14,18–34]. Interestingly, the major- sent study, the frequency spectrum of the envelope is referred
ity of the published works notice an increase of the noise and for brevity as envelope spectrum (ES). These studies focus on cav-
vibration power content in these signals close to the visual incep- itation formations in the runner and the draft tube of Francis tur-
tion of the phenomenon, and show that this approach can be used bines, and manage to associate cavitation with discrete low
towards the prompt detection of cavitation before total head frequency components of the ES. With this approach it is possible
collapse. to extract the detection information needed to form a common cri-
On the other hand, however, the application of the above signal terion. However, the aforementioned studies do not provide a tool
processing method in several different pumps and test configura- that identifies the frequency band that is used for the signal’s
tions, reveals its weakness to provide a safe cavitation detection demodulation, and as previously discussed, their results deviate
tool with general applicability. More specifically, comparison of significantly among the different testing configurations.
the noise and vibration spectra in various cases shows the excita- In order to efficiently define the characteristics of the band pass
tion of different frequencies from cavitation development [9,11– filter used to unmask cavitation noise and vibration, as well as to
20]. Such behaviour highlights the increased randomness of the efficiently extract cavitation information from cyclic frequencies
phenomenon itself, as well as the additional randomness intro- and resonant frequency fields, the present study proposes the
duced by the interaction of the bubble implosion with the struc- use of the Spectral Kurtosis tool (SKT) [49,50]. The SKT is imple-
tural and operational characteristics of the machine and the mented here for the first time in vibration measurements for the
configuration. Hence it is not possible to define a common fre- detection of two-phase flow formations in centrifugal pumps. More
quency band in all cases, at which the integration of the noise specifically, two accelerometers are located on the casing of a
and vibration power could be used as a general method for prompt radial flow centrifugal pump and the cavitation characteristic
detection of cavitation in hydraulic turbomachines. curves are derived in order to check the capabilities of the pro-
A second limitation of this detection approach is that the posed detection methodology. The tested pump uses two semi
increase of noise and vibration under cavitating conditions open impellers with different blade geometrical characteristics,
requires the knowledge of the level of noise and vibration in the so as to investigate their effect on the applicability of the method-
same frequency ranges when the machine operates under healthy, ology. The pump casing is made of Plexiglas and hence the pres-
non-cavitating conditions. This information can be obtained during ence and the extent of vapour formations in the pump can be
controlled laboratory measurements or for pumps equipped with visualised. The present study justifies the selection of SKT after
predictive maintenance monitoring equipment, but not for the vast carefully examining the transient behaviour of the statistical
majority of hydraulic turbomachines in operation. moments, as well as the FS in the measured data and highlights
One additional difficulty and limitation of the above method the effect of cavitation on the vibration time series (TS) pattern,
arises from the use of rolling element bearings that are found to as well as on the ES under different suction conditions.
excite wide range of frequencies when operate under faulty condi- The description of the experimental set up and testing proce-
tions, in a similar manner to the cavitation phenomenon [41–47]. dure is presented at first (Section 2), and the mathematical formu-
These similarities in the FS are due to the impulsive behaviour that lation of the detection methodology follows (Section 3). Then, the
both these mechanisms exhibit. As a result, additional characteris- cavitation test results (Section 4) and the appropriate processing of
tics have to be found in the measured signal so as to discern the the vibration signals (Section 5) are analysed and discussed, in
faults related to rolling element bearings from cavitation. order to identify the effect of cavitation on the vibration signal,
The extraction of the above conclusions from the literature to justify the selection and to evaluate the detection capabilities
reveals the additional effort towards the formation of a well- of the proposed method.
rounded cavitation detection criterion that will not change in dif-
ferent machines, as well as it will allow the differentiation of the
various impulsive fault mechanisms. In addition, the random 2. Experimental set up
impulsive noise sources, generated by cavitation bubbles collaps-
ing in the pump, cause complex modulation effects. Thus, the 2.1. The pump test laboratory
vibratory response of a pump under cavitation is not periodic,
but it inherently exhibits cyclostationary characteristics, as it is The test configuration (Fig. 1) is installed in the Engineering
generated by a hidden periodic mechanism due to the rotation of Department of Lancaster University. UK, and includes the radial
shaft and impellers. This means that the statistical moments of centrifugal pump (i), the main tank (vii), the valves (vi, viii), and
2
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Fig. 1. Laboratory test rig configuration; i) pump, ii) accelerometer B01, iii) accelerometer A01, iv) discharge pressure transducer, v) flow meter, vi) discharge valve, vii) water
tank, viii) suction valve, ix) suction pressure transducer, x) 90° elbow, xi) stroboscope.

the installed sensors for measuring the vibration (ii, iii) and the onset appears at first at its left side. For this reason, the flow mon-
pump performance (iv, v, ix). A stroboscope (xi) enables to visu- itoring area is located at this side, as depicted in the blue area of
alise the development of vapour bubbles in the pump. The pump Fig. 2b. The camera used is a DSLR type with 4752
3168 pixels,
is driven by an induction motor with 3 horsepower rated output CMOS image sensor and max shutter speed 0.25 msec.
at constant rotational speed, n = 1800 rpm, while the water tem- Two piezoelectric accelerometers, B01 (Fig. 1, ii) and A01 (Fig. 1,
perature during the test was T = 23 °C. iii), are mounted on the casing of the pump, the first closer to the
Fig. 1 highlights the main advantage of this test rig for the study two-phase flow area and the second on the volute. The frequency
and visual inspection of cavitation inception and development; the range, [fmin-fmax] that each sensor is able to measure is provided
volute, the front casing of the pump and all the piping system are by the manufacturer and its value is given for both accelerometers
made from Plexiglas material. In this way it is possible to visualise in Table 2. However, the analog signal of the sensors passes
the flow cavitating conditions in the pump impeller and the suc- through an analog low pass filter (LPF) before its digitization in
tion pipe system. In addition, the pump does not have its own the MCC USB-1608G DAQ device, which practically decreases the
bearings but all the mechanical and hydraulic loads are transferred upper limit of their linear frequency range. More specifically, the
to the bearings of the motor. This facilitates the detection of cavi- cutoff frequency, fcutoff, of the LPFs used in the present study is
tation phenomenon, since the noise and vibration emitted from the 10 kHz and 5 kHz for the LPF connected with the B01 and A01,
bearings is more difficult to surpass the cavitation noise and vibra- respectively (Table 2), but the filter roll-off is only 12 dB/octave
tion in the impeller. (Table 2). The analog signal of both sensors is digitised with a
The tests include the investigation of two impellers with differ- sampling frequency, fs, of 20 kHz (Table 2). This means that com-
ent geometric design, one with curved main and splitter blades ponents from the range 10–20 kHz for B01 will be attenuated by
(impeller No. 1, Fig. 2a) and a second with straight, radial blades 3–13 dB in the aliased range 10 down to 0 kHz, and an accelerom-
(impeller No. 2, Fig. 2b). Both are unshrouded, semi-open impel- eter resonance of 30 kHz would be attenuated by a maximum of
lers, made from aluminum, and their geometrical characteristics 19 dB. Components from the range 15–20 kHz (including the
are given in Table 1, where R is the radius (Fig. 2a,c), b the blade accelerometer resonance, Table 2) for A01 will be attenuated by
angle (Fig. 2a), b is the blade height (Fig. 2c), e the blade width at 19–24 dB in the aliased range 5 down to 0 kHz.
the impeller inlet and outlet periphery (Fig. 2a) and ttc the thick- However, suspicious changes in resonance frequencies from the
ness of the radial tip clearance (Fig. 2c). The subscripts 1, 2 corre- onset of cavitation, which might be due to aliasing, are not evident
spond to the inlet and the outlet of the impeller, respectively. for either case in the spectra or kurtograms of Figs. 9 and 11. Even
The elbow pipe upstream of the pump inlet (Fig. 1(x)) intro- if there were some aliasing it would only affect the apparent reso-
duces additional losses and causes a non-axisymmetric flow field nance frequencies excited, and not change the demodulated envel-
at the pump inlet and in the rotating impeller, where cavitation ope signals on which most of the results are based. In future work

Fig. 2. a) Impeller No. 1, b) impeller No. 2, c) meridional view of the centrifugal pump.

3
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Table 1
Design data of the impellers tested.

Imp. No. Inlet/Outlet Blades No. Splitter Blades. R [mm] b [deg] b [mm] e [mm] ttc [mm]
1 1 6 6 33 65 7.5 1.1 1
2 82.5 25 3 10.1
2 1 12 0 33 85 7.6 2.8 0.9
2 82.5 90 3.1 5.9

Table 2
Accelerometer, LPF and acquisition characteristics.

Sensor fs [kHz] fcutoff [kHz] tacq [sec] [fmin-fmax] [Hz] Resonant Freq. [kHz] Filter roll-off [dB/oct.]
B01 20 10 30 0.2–15000 30 12
A01 20 5 30 0.2–10000 20 12

!
it will be ensured that potential aliasing is eliminated. Finally, the 1 8Q 2 1 1
time duration, tacq, of the digitization of both vibration signals is 30 Htot ¼ ðpa  pe Þ þ 2  þz ð3Þ
qg p g da 4 de 4
sec (Table 2).
where q is the water density, g the gravitational acceleration, d the
2.2. Testing procedure, performance equations and uncertainty pipe diameter and z the vertical distance between the discharge (a)
estimation and suction (e) static pressure measuring points. In the present
study, the Hsat value is taken equal to 0.29 mcw that corresponds
The measurements aim at the derivation of the centrifugal to 23 °C (the water temperature during the tests), z = 0.305 m,
pump cavitation characteristic curves, NPSH-Htot, in their non- and da = 0.041 m and de = 0.047 m. Finally, the Q, NPSH and Htot val-
dimensional form, i.e. r-W, where r is the Thoma cavitation factor ues are non-dimensionalised using the flow coefficient U (Eq. (4)),
and W the head coefficient. These curves illustrate the variation of the Thoma cavitation factor, r (Eq. (5)) and the head coefficient W
the total head of the pump, Htot (W) when the available net positive (Eq. (6)), respectively:
suction head of the establishment, NPSH (r) decreases. Four curves
c2n 1 Q
at four different flowrates (Q) are derived for each impeller exam- U¼ ¼ ð4Þ
ined, following the open sump with throttle valve procedure that is
u2 xR2 ð2pR2  zs e2 Þb2
provided from ISO 9906:2012 [51]. During this procedure, the level
NPSH
of static pressure in the testing configuration is progressively r¼ ð5Þ
Htot
reduced by throttling the suction valve. The static pressure level
of the test facility is described from the NPSH available expression
2gHtot
that is given in Eq. (1), where HE is the pressure at the surface of the W¼ ð6Þ
u22
tank (Fig. 1, vii) in meters of water column, and DHz,e the height
difference between the pump suction (e) (Fig. 1) and the water where cn is the radial component of the absolute flow velocity, u is
level in the tank. The term feQ2 represents the hydraulic losses in the tangential velocity of the impeller (Fig. 2a), and zs the number of
the suction pipe, and Hsat the water saturation pressure expressed impeller blades.
in meters of water column (mwc). During the time span of 2 min the macroscopic characteristics
(position, shape and extent) of the cavitation area through the
NPSH ¼ Hoe  Hsat ¼ HE  DHz;e  fe Q 2  Hsat ð1Þ
camera remain almost unchanged. As a result, the experimental
In practice, the implementation of the open sump with the throt- photographs presented in this article are representative of the
tle valve procedure increases the losses term (feQ2) of Eq. (1), and visually observed characteristics that the vapour phase exhibits
this in turn reduces the flowrate. For this reason, the discharge valve during the 2 min of each measurement, even though the phe-
has to be properly regulated so as to restore the flowrate to its initial nomenon microscopically exhibits transient features. It should be
value and derive the r-W curves under constant flowrate conditions. mentioned here that flow visualisation is used as a tool that con-
This procedure introduces a limitation to the minimum attainable firms macroscopically the existence or absence of cavities, and its
value of r, when the discharge valve reaches its fully open position, results are subject to camera limitations. As a consequence,
because after that any further throttling of the suction valve reduces micro-cavities may exist microscopically even before the optical
the flowrate and cannot be compensated. Thus the r variation range visual inception but the resolution of the camera could not depict
becomes smaller at higher flowrates, where the testing starts with them. All tests are performed three times in order to ensure the
the discharge valve in a more open position. repeatability of measurements of the flow quantities, vibration
After 30 sec from the adjustment of the valves, the flow is sta- and cavitation depiction.
bilised and the flow conditions are considered as steady state. The total relative uncertainty, F(u,t), of each measured (Q, p, T,
From that point, the flowrate, Q and the static pressure, p, at the n) flow and operating variable u is calculated according Eq. (7)
suction (e) and the discharge (a) of the pump (Fig. 1) are recorded [52,53]:
for a time period of 2 min, while the vibration signals are measured
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
h i2ffi
for 30 sec (Table 2). The measuring of the pressure and flowrate 2 2 2 dðu;r Þ
data allows for the calculation of NPSH and Htot values, with the F ðu; t Þ ¼  F ðu; sÞ þ F ðu; r Þ ¼  F ðu; sÞ þ u ¼ :::
use of Eqs. (2) and (3), respectively: rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  2
tSu
::: ¼  F ðu; sÞ2 þ u
p 8Q 2
NPSH ¼ Hoe  Hsat ¼ e þ  Hsat ð2Þ
qg g p2 de 4 ð7Þ

4
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

D E
where F(u,s) is the systematic relative uncertainty, F(u,r) the ran- jY w ðkP; f Þj
4n

dom relative uncertainty, d(u,r) the absolute random uncertainty, ^ Y ðf Þ, nD

K E o2  2
k
ð9Þ
2n
Su the standard deviation and t the Student parameter [52]. F(u,r) jY w ðkP; f Þj
k
and d(u,r) are associated with the Student’s t distribution and with
the uncertainty at the 95% confidence level, and they express the where Yw(kP,f)xn is the xn-th order empirical spectral moment of the
errors of repeated measurements obtained under unchanged oper- STFT of a sampled process Y(n), w(n) is the analysis window, P is the
ating conditions. Finally, errors of the measured flow and operating temporal step and h. . .ik is the time averaged operator over index k,
variables are propagated in the calculations of NPSH, Htot, U, r and which denotes the number of temporal steps. It is noted that the
W. The levels of total relative uncertainty of the above calculated mathematical definition of Spectral Kurtosis (Eq. (9)) is based on
quantities, are calculated based on the general formula (Eq. (8)) cumulants. This justifies the subtraction of 2 in the Eq. (9), which
for error propagation. Accoding to this formula, the absolute error makes Spectral Kurtosis value to be equal to 0 for Gaussian signals.
of a quantity u derived from equation u = f(x,y,z), where x,y,z are Furthermore, in order to efficiently scan the FS and identify the fil-
variables with individual uncertainties, is given by: ter characteristics, the Spectral Kurtosis is calculated for different
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi frequency resolutions and all the results are plotted as function of
 2  2  2
@u @u @u frequency. This calculation can be accomplished with the use of Fast
dðu; t Þ ¼ dðx; tÞ þ dðy; t Þ þ dðz; t Þ ð8Þ
@x @y @z Kurtogram (FK) that is introduced by Antoni [50], where the FS
decomposition is made with the use of Eq. (10):
In Table 3, the total relative uncertainty of all measured and cal-
culated quantities is provided. The uncertainties of density q, grav- df ¼ 2ðDL1Þ ; DL ¼ 0; :::; 7 ð10Þ
ity acceleration g, vertical distance z, impeller radius and suction
where df is the frequency resolution and DL is the Decomposition
and discharge pipe diameters are ignored. In general, the uncer-
Level. The final outcome of the calculations is a two-dimensional
tainty values presented in Table 3 are low and ensure the reliability
map of the value of Spectral Kurtosis (Fig. 3), calculated by Eq.
of the conclusions drawn from the cavitation tests on this particu-
(9), as function of frequency versus frequency resolution. The loca-
lar test configuration.
tion and the maximum value of the Spectral Kurtosis for each DL
changes as the original signal is divided into more and more seg-
3. Signal processing of the vibration data ments. The Maximum Spectral Kurtosis (MSK) can be finally
obtained for a certain frequency range and DL (Fig. 3).
In this section, the theoretical background of the SKT and the The FK designates the frequency and frequency resolution of the
demodulation technique used in the dynamic measurements of MSK value, which is information of crucial importance regarding
the present study for the detection of cavitation, are provided. It the detection of non-stationarities. More specifically, those quanti-
is noted that the implementation of the signal processing tools is ties correspond to the central frequency, fc and the bandwidth of
made in MATLAB proprietary programming language. the band pass filter to be used for the isolation of the impulsive
behaviour hidden in the sampled signal. In the case of Fig. 3, the
3.1. The Spectral Kurtosis tool MSK = 4 corresponds to a band pass filter with fc = 8 kHz and
1.25 kHz bandwidth (DL = 3.6).
The aim behind the development of the SKT is the extraction of
information related with a non-stationary process from a TS that 3.2. Envelope analysis
otherwise presents Gaussian characteristics. The applicability of
such a tool on the field of machine diagnostics lies in the fact that The present study investigates the envelope characteristics in
usually noise and vibration signals measured for the detection of a the frequency domain towards the detection of cavitation in
fault contain information from multiple parts of the machine and hydraulic turbomachines. The signal processing approach adopted
resemble stationary random signals, such as white noise. At the here is a typical demodulation or enveloping procedure, which is
same time, information related to a possible fault that exhibits also used for the detection of rolling bearing faults that present
highly impulsive non-Gaussian behaviour may be hidden in the impulsive characteristics, similarly to cavitation [58–61]. The
original signal, and hence its extraction would be of critical impor- demodulation procedure is carried out in the frequency domain
tance to effectively detect this fault. of the discretised filtered signal, xf(n), so as to avoid the special
The use of higher-order statistics for the pattern recognition of manipulations introduced by the convolution operation in the time
non-stationary signals was first introduced by Dwyer [54–57] and domain. More specifically, the demodulated time signal and the ES
dealt with underwater noise measurements. During the statistical are calculated in the following successive steps [62]:
analysis of noise signals, Dwyer observed that skewness and kurto-
sis values deviate from the values of Gaussian distribution. Dwyer i. The fast Fourier transform algorithm is used for the calculation
also perceived that the representation of these statistical moments of the discrete Fourier transform of the filtered signal, Xf(k), as
in the FS and, especially, the depiction of Spectral Kurtosis value expressed by Eq. (11):
could provide additional information for the source of non-
stationarity. The complete mathematical formulation for the STFT
1 X
N1

estimation of Spectral Kurtosis, Kb Y(f), is given by Antoni [49]: Xqf ðkÞ ¼ xf ðnÞej2pkn=N ; k ¼ 0; :::; N  1 ð11Þ
N n¼0

Table 3
Total relative uncertainty of measured and calculated variables.

Measured Variables Calculated Variables

u Q pe pa T n U W r Htot NPSH
±F(u,t) [%] 0.6 0.3 0.3 1 0.1 0.6 0.6 0.8 0.6 0.6

5
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Fig. 3. Fast Kurtogram at the vibration signal of a centrifugal pump under cavitating
conditions.

where N is the length of the sample, calculated by Eq. (12).

The multipliers 103 appears because fs is given in kHz
Fig. 4. Cavitation r-W characteristic curves for the two impellers at different U
(Table 2).
values.
N ¼ t acq  f s  103 ð12Þ
tograph in the middle). The photographs provided here correspond
to one specific flowrate, but they can represent sufficiently the
ii. The H(m) vector is created, according to Eq. (13): changes observed in the shape and the extent of the vapour area,
8 during the testing of all loading conditions.
< 1 for m ¼ 1; ðN=2Þ þ 1
> For all cases tested the total head values (W) remain unaffected
HðmÞ ¼ 2 for m ¼ 2; 3; :::; ðN=2Þ ð13Þ by the reduction of r value (Fig. 4), even after the formation of the
>
: first visually observed cavities at rVI (blue markers in Fig. 4 and
0 for m ¼ ðN=2Þ þ 2; ::; N
Figs. 5a and 6a). Furthermore, for both impellers and all flowrates
examined, the first small bubbles (blue circles drawn on Figs. 5 and
iii. The Hadamard product of H(m) and Xf(k) vectors is then 6) always appear at the suction side of the leading edge of the
calculated: blades, where the minimum static pressure is developed that can
drop below the water saturation pressure and inititate local evap-
Z ðmÞ ¼ HðmÞ  X f ðkÞ ð14Þ oration [2,4,63]. Although, the reduction of the cavitation factor
value at intermediate cavitation conditions results in increased
iv. The N-point inverse fast Fourier transform of Z(m) is the ana- number of bubbles (Figs. 5b and 6b), the vapour area does not
lytic signal, As(n), and its amplitude is the demodulated time influences the power transfer to water, since W remains constant,
signal, i.e. the envelope of the filtered TS, ed(n), which is calcu- as when the pump operates under non-cavitating conditions
lated by Eq. (15): (Fig. 4).
  However, reducing the suction available pressure at rFC (black
XN  colour markers in Fig. 4) results in deterioration of the pump total
 þj2pmn=N 
ed ðnÞ ¼ jAs ðnÞj ¼  Z ðmÞe  ð15Þ head and confirms the general shape of the r-W curve that can be
m¼1 
found in the literature [2,4,64]. The drop of W, presented in Fig. 4,
results from the creation of a large attached cavitation vapour bub-
v. Finally, the ES can be calculated from Eq. (16): ble at the inlet of the impeller that reduces drastically the flow pas-
sage area between the adjasent blades, as well as spoils the
1XN1 hydrodynamic performance of the blade (Figs. 5 and 6). Only the
ed ðkÞ ¼ ed ðnÞej2pkn=N ; k ¼ 0; :::; N  1 ð16Þ head of the largest flowrate of Impeller No. 1 (Fig. 4a) does not
N n¼0
show this behavior, due to the inability of the testing method to
further reduce r at higher flowrates, as discussed in Section 2.2.
4. Cavitation r-W curves and flow visualisation results Another observation from Fig. 4 is that for both impellers, an
increase of the flowrate results in higher values of both rFC and
The results of the r-W cavitation characteristic curves of the rVI. This is because at higher flowrates in a specific impeller, the
two impellers at four different flowrates are given in Fig. 4. It is static pressures at the suction side of the pump reduce, hence
noted that the blue and black color markers on these curves indi- favoring the vapour formation. Finally, vapour formations during
cate the visual inception (rVI) of the first vapour bubble and the rFC conditions at impeller No. 2 extend to a larger part of the flow-
total head drop (rFC) conditions, respectively. The discussion of path (Fig. 6c) compared to impeller No. 1 (Fig. 5c) for two reasons,
r-W curves is accompanied by the presentation of the flow visual- related to their geometrical characteristics; i) the larger number of
isation results, which are given in Figs. 5 and 6, for the two impel- blades (12, instead of 6) and ii) the higher inlet blade angle (85°
lers tested. These figures illustrate the development of cavitation instead of 65°). More specifically, the larger number of blades
visually obtained characteristics under rVI (left photograph), rFC reduces the net flow passage area and accelerates the flow at the
(right photograph) and intermediate cavitating conditions (pho- impeller inlet. On the other hand, the even higher inlet blade angle

6
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Fig. 5. Flow visualization results for impeller No. 1, at U = 0.027.

Fig. 6. Flow visualization results for impeller No. 2, at U = 0.046.

of impeller No. 2 than the relative flow incidence angle, increases passing frequency (TBPF). The latter is calculated only for impeller
the flow impact losses and local recirculation at the suction side No.1 that contains a number of splitter blades. The values of RF,
of its blades, thus making the impeller more prone to cavitation. BPF and TBPF are given for each impeller in Table 4, along with
their corresponding periodicities RP, BPP and TBPP.
The vibration TS during non-cavitating conditions (Fig. 7a) are
5. Vibration measurement results
mainly formed by a synthesis of sinusoidal components that
include the passing of the rotating blades, and its trend for impeller
The results of the vibration data, which are obtained from the
Νo. 1 is depicted in this graph. Here, the signal pattern agrees with
accelerometers on the pump casing (Fig. 1), are initially examined
the expected vibration signature, which in hydraulic turbomachin-
in the FS and in time, along with the behaviour of their statistical
ery is usually modulated by RF, BPF and TBPF [66,67] and should be
moments. The FS of the measured time signals are calculated with
a kind of sinusoidal, especially in the present pump that does not
the use of Welch function [65]. In addition, the present study
have embedded bearings (zoom in area of Fig. 7a).
focuses on the application of the SKT and on the analysis of the fil-
The passing from non-cavitating to visual inception conditions,
tered signals that result after the implementation of the band pass
at rVI (Fig. 7b), does not change the signal pattern compared to
filters calculated from the FK. Finally, the filtered signals are fur-
Fig. 7a, except of a slight increase in the vibration amplitude,
ther processed, so as to study the ES and to reveal the special char-
which, however, is difficult to correlate with cavitation. Under such
acteristics introduced after the development of cavitation in the
conditions, only few bubbles develop in the impeller and cover a
centrifugal pump. It should be noted that the cavitation tests
very small part of the flowpath (Figs. 5a and 6a). Consequently,
include the investigation of four different flowrates for each impel-
the overal vibration of the machine overrides that of pressure wave
ler (Fig. 4), each of which consists of an average of ten operating
noise.
points (OPs), while every point includes the vibration data of two
At intermediate cavitation conditions (Fig. 7c) and at head
accelerometers. Consequently, in order to reduce the number of
breakdown, rFC (Fig. 7d) some special characteristics appear in
the presented data, only part of the latter that adequately repre-
the vibration TS. When the pump operates at r = 0.30, the data
sents the vibration pattern of the impellers is selected for discus-
include areas of small amplitude that make the signal have a den-
sion. Similarly, regarding the raw vibration data and their
ser pattern (blue circles in Fig. 7c). Furthermore, at full cavitation
statistical properties, the results of one impeller are presented,
conditions, where vapour bubbles cover a large part of the impeller
since they do not differentiate significantly when the blade geom-
flowpath (Figs. 5c and 6c), the vibration signature exhibits
etry changes.
impulses of fine timespan with increased amplitude (blue ellipses
in Fig. 7d), which are presumably associated with the implosion of
5.1. Detection based on time and frequency domain analysis vapour bubbles.
Although the analysis of the raw measured data of Fig. 7 pro-
Fig. 7 shows four raw vibration signals, obtained from sensor vides some first indications that reveal the presence of cavitation,
A01 under different suction conditions of impeller No.1; i) single- they cannot constitute a prompt cavitation detection criterion.
phase (Fig. 7a), ii) cavitation inception, rVI (Fig. 7b), iii) intermedi- The reason is that these indications are masked by the machine
ate cavitation (Fig. 7c), and iv) head breakdown, rFC (Fig. 7d). The noise and cannot become easily observable at the early stages of
TSs of Fig. 7 are focused in a short, indicative timespan (0.08 cavitation development. As a result, further analysis of the signal
sec), in order to illustrate details such as the signal amplitude or properties must be carried out, in order to unmask and quantify
possible bursts and impulses. Moreover, this time span is large new information related to the cavitation mechanism, and its
enough to include the period of the main low frequency compo- interaction with the flow conditions and the machine components.
nents created by the pump, such as the rotating frequency of the An additional conclusion extracted from the TSs of Fig. 7, is that,
shaft (RF), the blade passing frequency (BPF) and the total blade regardless of the suction conditions, the combination of processes
7
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Fig. 7. Time series under various suction conditions for impeller No. 1, sensor A01, U = 0.027.

Table 4
Frequencies and periodicities excited by the pump.

Impeller No. RF [Hz] RP [s] BPF [Hz] BPP [s] TBPF [Hz] TBPP [s]
1 30 0.033 180 0.0055 360 0.0027
2 30 0.033 360 0.0027 – –

that compose the measured signals exhibit stationary characteris- the normal distribution value (kurtosis = 3). The closer the pump
tics in terms of their mean value, since the latter remains constant operates to the total head drop (rFC) point the stronger peaks are
and does not change during the sampling time. In order to examine generated due to bubble implosion, making kurtosis to take values
the development of possible non-stationarities in the remaining higher than 3. This conclusion is of crucial importance since it
statistical properties, the time-varying patterns of the standard reveals the hidden non-stationarities resulting from two-phase
deviation, skewness and kurtosis need to be calculated. flow development in the pump, which are not observable in the
A practical way to calculate the time-varying behavior of the TSs of Fig. 7. The correlation of non-stationarities with the kurtosis
moments is through the splitting of the measured data in sampled value of the vibration data agrees with the results presented from
groups that include the same number of raw data. In the present Dwyer [54–57] and justifies the selection of kurtosis-related tools,
study, the signals are put in groups of 1000 raw data. For each such as the SKT presented in Section 5.1, for the detection of
group, the standard deviation, the skewness and the kurtosis are cavitation.
calculated, and in this way it is possible to illustrate their pattern The raw vibration data are also used to investigate the differ-
in time that is now expressed in number of samples. After this pro- ences observed in the FS, due to cavitation development. The anal-
cessing, it is possible to examine whether their values remain con- ysis makes use of semi-log plots (Fig. 9), where only the y-axis of
stant (the moments describe a stationary process) or deviate in the power distribution is shown in logarithmic scale. This approach
time (the moments describe a non-stationary process). The results is used by several researchers [11,12,21,26,29], in order to high-
that correspond to impeller No.2, sensor B01, at U = 0.061 are light the differences observed in the low amplitude area of the
shown in Fig. 8a,b,c for the standard deviation, skewness and kur- vibration spectrum. The results for an indicative flowrate are given
tosis, respectively. in Fig. 9 and include the spectra of various cavitating conditions for
As can be seen in Fig. 8a, the standard deviation value does not both impellers tested.
change as function of the number of samples (time) during con- In these graphs it is clear that cavitation excites wide areas of the
stant suction conditions. Differences are only observable between FS, regardless of the particular impeller and sensor mounting posi-
the different r values examined, where the standard deviation tion. The excitations produced by the two-phase flow development
value increases with the development of cavitation in the pump. (Fig. 9) are strongly related with the mechanism of cavitation, and
The latter depicts the effect of the impulses on the measured data, more specifically, with the shock pressure waves created during
shown in Fig. 7c,d which increase the vibration amplitude range the bubble implosion. It is important to underline that the excita-
and, concequently, affect the standard deviation value. The skew- tions due to cavitation development do not refer to the increase of
ness value in Fig. 8b does not deviate neither as function of time discrete frequency components, which appear in the high frequency
nor under different cavitating conditions. range even before the cavitation development, at r > rVI (green line
On the contrary, the graph of Fig. 8c, reflects the highly tran- in Fig. 9). This is concluded after careful examination of the spectrum
sient behavior of kurtosis, when the impeller operates under inter- and the observation of the existence of green peaks in the majority of
mediate (r = 0.39) and heavy cavitating (r = 0.12) conditions. More the high frequency discrete components (Fig. 9).
specifically, for non-cavitating (r = 0.67) and visual inception (rVI) In general, the changes in the characteristics of FS due to
conditions, the kurtosis values remain unchangeable and close to cavitation development are similar with those of the published
8
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Fig. 8. a) Standard deviation, b) skewness and c) kurtosis values of the grouped data under various suction conditions for impeller No. 2, sensor B01, U = 0.061.

Fig. 9. Welch estimator of power spectral density under various suction conditions for a) impeller No. 1, sensor A01, U = 0.054, and b) impeller No. 2, sensor B01, U = 0.091.

literature for hydraulic turbomachines. Particularly the similarities machines, operating and suction conditions [11,12,18,19,21,
with the results of Escaler [18–19] spectra that refer to Francis 26,29,40].
hydroturbine prototypes of 10 MW, reveal the analogies of the cav- The reason for such differences is that the FS excitations
itation effect between the different types and size of hydaulic demonstrate mainly the symptom of the fault (cavitation develop-
machines. However, despite the decent agreement with the pub- ment) rather than the actual mechanism, which in case of cavita-
lished results, the intensity, the location and the range of excita- tion is the bubble implosion close to the rotating and stationary
tions differentiate among the various testing configurations, solid surfaces, and it is the same for every hydraulic turbomachine.

9
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

On the other hand, the symptom (high frequency excitation)

reveals the interaction of the fault mechanism with the pump
structural characteristics, and consequently, it varies among vari-
ous configurations. As a result, it is not possible to create a general
detection cavitation criterion that could be broadly used in hydrau-
lic turbomachinery applications.
Furthermore, as mentioned previously, there are faults related
with rolling element bearings, which at some stage of their evolu-
tion excite a wide range of high frequencies, up to 20 kHz, in a sim-
ilar manner to cavitation. As a result, the FS analysis may result in
incorrect conclusions regarding the determination of the type of
damage that the hydraulic machinery suffers from. A practical
way to distinguish the cavitation from bearing faults is to operate
the hydraulic turbomachine in lower (or higher) r values, to study
the changes on the FS. However, this procedure is not practical
when dealing with industrial hydraulic turbomachines and conse-
quently, the use of such tools can not provide a clear answer
regarding cavitation development.
Therefore, additional effort is needed for a well-rounded char-
acterization of cavitation that will include the identification of
the phenomenon signature in time, and its correlation with a range
of frequencies that will not change in different machines. In this
direction, the results presented in this section, especially those that
refer to kurtosis and to excitations of the frequency vibration spec-
trum, are further investigated, since they justify the use of SKT for
cavitation detection and will be thoroughly discussed in the next
section. Fig. 10. MSK, DL and fc values as a function of r, for a) impeller No. 1, sensor A01,
U = 0.040 and b) impeller No. 2, sensor B01, U = 0.076.
5.2. Detection based on the application of Spectral Kurtosis tool

The presentation of the results obtained by the application of to those under non-cavitating conditions. Nevertheless, since other
SKT is divided in two parts. The first refers to the results of the faults that exhibit impulsive behaviour, such as those related with
FK and includes the calculations of the MSK, DL and band pass filter bearings, are also expected to excite wide ranges of the spectrum
central frequency, fc, values. The second part presents the results [42,43,68] justifies that such a drop of DL value occurs also during
after the application of the filtering and demodulation procedure, their development [50,69,70]. Consequently, the decrease of DL
i.e. the TS and the ES of the filtered signals. In both parts the results value could be a strong sign that an impulsive fault develops in
are presented as a function of the suction conditions at various the machine (bearing or cavitation related), however, in order to
flowrates. The study focuses mainly on the initial and intermediate identify the exact type of fault further analysis on the signal is
cavitating conditions (rVI > r > rFC), where the effective and reli- required.
able detection of two-phase flow formation is a challenge. Finally, in order to provide an interpretation of the FK quantita-
tive results presented in Fig. 10, the corresponding FK charts for
5.2.1. Fast Kurtogram results OPs of a particular flowrate and sensor for each impeller are
The results of the MSK, DL and fc values as function of r for a selected to be drawn in Fig. 11. The left FK chart corresponds to
specific flowrate and sensor of each impeller are drawn in normal, and the right to cavitating flow conditions, for which the
Fig. 10. The blue dashed line symbolizes the rVI OP, where two- systematic shifting to lower DL and fc values can be seen for both
phase flow area becomes visible in the impeller. According to these cases. The area with the MSK that is selected to be filtered after
graphs, the MSK value does not exhibit the same trend for both the FK calculations, is highlighted with a red rectangle in Fig. 11.
impellers, after the onset of cavitation. More specifically, the
MSK value of impeller No.1, remains relatively unchanged as the 5.2.2. Filtered time signals
r drops (Fig. 10a), whereas for impeller No.2 there is an abrupt The filters obtained from the application of the FK tool are
increase of MSK at rVI (Fig. 10b), which, at smaller r values exhi- applied in the raw time signals. These are the same TS as used
bits a scattered behaviour, but keeps clearly higher values then by the FK tool for the calculation of the band pass filter character-
under single-phase conditions (r > rVI). This behaviour could be istics, which are different for every OP (Fig. 10). The filtering proce-
the result of the mechanism discussed in [7,9,11,14] that deals dure is evaluated based on the results of the TS and those of the ES.
with the noise absorption from the vapour phase at low r values. The reason for presenting the filtered results in the time domain is
This could affect in a similar manner the non-stationary character- to examine their characteristics and investigate the possibility to
istics of a vibration signal and, consequently, may reduce the MSK distinguish cavitation from other possible faults.
values. The filtered time signals for the two impellers are presented in
On the other hand, the DL and fc curves presented also in Fig. 10, Figs. 12 and 13, during early and intermediate cavitating condi-
appear to decrease at lower r values, no matter the flowrate, the tions, for rVI > r > rFC. The results during non-cavitating conditions
sensor and the impeller under examination. Especially for the DL (r > rVI) are also included for the lowest flowrate for comparison
value, its drop during cavitating conditions seems to be in agree- with the corresponding TS under cavitation. The timespan, dt, of
ment with the conclusions made in Fig. 9, where the wide range the filtered series is kept equal to the plots of Fig. 7, so as to include
excitations from cavitation are shown. These excitations are the all the characteristic frequencies excited by the centrifugal pump.
results of vapour bubble implosion and drive the FK to identify The filtered signals when the pump operates under single-
the maximum kurtosis value in larger frequency ranges compared phase conditions (Figs. 12a and 13a) exhibit the minimum ampli-
10
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Fig. 11. Comparison of FK between a,c) normal and b,d) cavitating conditions for a,b) impeller No. 1, sensor A01, U = 0.040 and c,d) impeller No. 2, sensor B01, U = 0.076.

tude values compared to cavitating conditions (Figs. 12b-e and considering the significant amount of noise and vibration sources
13b-e), and do not present any distinctive characteristics. In prac- removed from the band pass filter.
tice, at such conditions the application of the filter made from the The amplitude of the filtered signals of Figs. 12b-e and 13b-e,
FK tool does not reveal a particular pattern in the filtered signal, does not present a clear dependence on the flowrate. The signals
since for r > rVI the machine is expected to operate at no fault are formed by two main parts: one with low amplitude between
conditions. the impulses, and one that includes the impulse amplitudes that
The first important observation regarding the filtered vibration range from [0.4–3] and [1.5–9], for the impellers No. 1 and 2,
time signals is that their pattern changes under cavitating condi- respectively. The largest values for both impellers are obtained
tions, for both impellers (Figs. 12b-e and 13b-e). The application for the highest flowrate (Figs. 12e and 13e).
of band pass filter calculated from the FK, reveals the existence An additional observation regarding the TS presented in the fil-
of clear bursts in the filtered signal. Comparison of the filtered sig- tered signals of Figs. 12 and 13, is that the areas where the bursts
nal of Fig. 12b with its corresponding original time series presented are formed consist of random and periodic impulses. The latter fol-
in Fig. 7c, confirms the successful application of the filter, since the low mainly the blade passing frequency (BPF) of the impeller, and
impulses are not observable in the raw measured vibration data. they are indicated with the blue dashed lines. On the other hand,
This observation is of paramount importance, since it reveals the impulses that are aperiodic are marked with red ellipses and indi-
impulsive pattern of the cavitation mechanism that is masked in cate the random character of bubble implosion along with some
the original raw signal by various other noises of the pump. The distortion of the signal after the implementation of the filtering
decrease of the amplitude after filtering the signal is justified, process. The enveloping process that is presented in the next sec-

11
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Fig. 12. Filtered time series for impeller No. 1 a) sensor A01, U = 0.027, r = 0.43, b) sensor A01, U = 0.027, r = 0.30, c) sensor B01, U = 0.040, r = 0.30, d) sensor A01, U = 0.054,
r = 0.26, e) sensor B01, U = 0.067, r = 0.58.

Fig. 13. Filtered time series for impeller No. 2, a) sensor Β01, U = 0.046, r = 0.52, b) sensor Β01, U = 0.046, r = 0.36, c) sensor Α01, U = 0.061, r = 0.39, d) sensor Β01,
U = 0.076, r = 0.36, e) sensor Α01, U = 0.091, r = 0.39.

12
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

tion, attempts and succeeds to bring out the frequency that modu- the clear appearance of these differences in the ES makes them uti-
lates the bubble implosion. lizable for cavitation detection by an automated system.
The correlation of the aforementioned impulses with the cavita- The BPF modulation resembles the envelope characteristics,
tion phenomenon is further validated by studying the TS at rFC presented also in [19,39,40] during the testing of reaction hydro-
(Fig. 14). At such operating conditions the cavitation area is turbines. More specifically, in these studies the vortex rope cavita-
expanded and covers large part of the flow path between the tion in the draft tube is associated with the 0.27RF discrete
blades, as depicted in Fig. 5c and 6c, and dominates the vibration frequency component, while the cavitation in the runner has been
signal. In the plots of Fig. 14 the periodicity with the blade passing associated with the blade and guide vane passing frequencies.
frequency and, simultaneously, the impulsive behavior of TS can be However, in those cases the characteristics of the filters were taken
clearly observed. from an examination of the measured signal spectrum, an
In addition, the bursts of Fig. 14 are more clear and frequent approach that has been proven to be complicated in the present
than those presented in Figs. 12 and 13. Especially for the impeller study (Fig. 9), but also from the FS of similar studies [11,12,16].
No. 2 (Fig. 14b), the effect of passing of each of the 12 blades is The successful implementation of Spectral Kurtosis method with
highlighted. The impulsive shape of the bursts of these plots the FK tool overcomes the aforementioned difficulties and achieves
resembles the shape of the filtered vibration signal, presented in to define the filter’s characteristics efficiently, in cases where it
[71,72], which corresponds to rolling element bearing fault. There- would be impossible to define them otherwise.
fore, in order to be able to discriminate the bearing fault and the The comparison of the amplitudes from the same sensor
cavitation impulsive behaviour, the use of the periodicity of the between the different impellers, depicts a rise of vibration inten-
impulses is proposed, since cavitation follows the blade passing sity, from the lowest values of impeller No. 1 [105–4104,
frequency and bearing frequencies are not exact harmonics of shaft Fig. 15], to the maximum values of impeller No.2 [104–4103,
speeds. However, as shown in the impulses of Figs. 12 and 13, dur- Fig. 16]. Therefore, this is the result of the geometric differences
ing the initial stages of cavitation development the BPF periodicity of the two impellers. More specifically, the amplitudes are reduced
is not always so clear, and hence, the ES is also used to extract the for impeller No. 1 (Fig. 15) due to: i) the use of splitter blades that
aforementioned information from the filtered signal. increase the cross section area of the flow path between the blades
and, consequently, reduces the flow velocity at the impeller inlet,
and ii) the smaller blade inlet angle that reduces the inlet flow
shock losses. In addition, in the majority of the ES of both impel-
5.2.3. Envelope spectra of the filtered signals
lers, the RF component can be observed (Figs. 15a-c and 16) indi-
The ES of the filtered vibration signals is calculated from Eq.
cating the amplitude modulation of both the BPF and the
(16), and is presented in Figs. 15 and 16, for three different r values
resonances from the shaft rotation.
and four flowrates. The green colour corresponds to non-cavitating
To sum up, the appearance of the passing frequency of the
conditions (r > rVI), while the red and black to initial (rVI) and
impeller blades is clear after the inception of cavitation. This beha-
intermediate cavitating conditions (rVI > r > rFC), respectively.
viour is observed in the filtered time series of Section 5.2.2 and it is
These plots include the spectra of both accelerometers, A01 and
further confirmed after their demodulation. Therefore, the identifi-
B01, used in the cavitation tests.
cation of the impulses in the filtered signal, which are produced
The results of the ES demonstrate the applicability of the digital
due to implosion of the vapour bubbles, and then modulated by
band pass filter, achieved after the calculation of MSK value in the
the blade passing frequency, can be used to detect cavitation from
FK and the decomposition of the filtered signal. More specifically,
the initial and intermediate stages of its development, but also
no matter the impeller, the sensor, and the flowrate examined
they highlight the physical mechanism of the fault, which is the
(Figs. 15 and 16), the ES exhibits significant differences when com-
implosion of the vapour bubbles. More specifically, the bubbles
paring non-cavitating with cavitating conditions. These differences
are formed at the suction side of the impeller blade (Figs. 5 and
are systematically related with the appearance of strong blade
6) and their implosion is followed by the creation of shock pressure
passing frequency components (BPF for impeller No.2, Fig. 16 &
waves. The waves hit the neighbouring solid surfaces, which for
TBPF for impeller No. 1, Fig. 15), when the pump operates at inter-
the impeller are the rotating blades. The latter rotate and due to
mediate cavitating conditions (rVI > r > rFC).
the machine’s non-uniformities, create blade passing frequency
The above behavior is critical, because it not only differentiates
noise that modulates the shock pressure waves. Consequently
the envelope’s characteristics after cavitation development, but
the raw vibration signal measured on the solid boundaries of the
also associates these characteristics with the blade passing fre-
pump includes the above information, along with additional noise
quency of the impeller, and further confirms the results of the fil-
from the pump-motor system. The application of the procedure
tered signal analysed in the previous Section 5.2.2. Furthermore,

Fig. 14. Filtered time series for a) impeller No. 1, sensor Β01, U = 0.027, r = 0.05, b) impeller No. 2, sensor A01, U = 0.046, r = 0.09.

13
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Φ=0.027, sensor A01

σ=0.43
(green colour)

σ=0.30
(red colour)

σ=0.23
(black colour)
Φ=0.040, sensor B01

σ=0.48
(green colour)

σ=0.30
(red colour)

σ=0.21
(black colour)
Φ=0.054, sensor A01

σ=0.64
(green colour)

σ=0.46
(red colour)

σ=0.26
(black colour)
Φ=0.067, sensor B01

σ=0.67
(green colour)

σ=0.58
(red colour)

σ=0.35
(black colour)

Fig. 15. Frequency domain of the envelope of the vibration signal of impeller No. 1.

described in this section achieves the unmasking of the noise SKT. The research focuses on the fault detection at the initial and
sources and reveals the interaction of cavitation with the rotation intermediate stages of cavitation development, and the following
of the blades. conclusions are drawn:
The combination of the systematic characteristics in the filtered
signal and the spectrum of the envelope with the physical meaning  The examination of the raw signals under heavy cavitating con-
of the fault mechanism create a unique detection criterion. To our ditions reveals the presence of impulses that increase the vibra-
knowledge, there have not been reported any fault mechanisms tion amplitude range. This is a typical symptom of fault in
related with centrifugal pumps and reaction hydroturbines, rotating machinery, while in the controlled tests of this study
excluding cavitation, which exhibit impulsive behaviour modu- the impulses appear only after passing into fully developed cav-
lated by BPP. In the published literature there are mechanical itation operating conditions.
faults, related with rolling element bearings in the hydraulic  The analysis of the statistical properties of the vibration signals
machines, where the pattern of the fault mechanism is strongly shows (or reveals or demonstrates) the increase of kurtosis to
impulsive, however, its periodicity does not follow the BPP but values higher than 3 when cavitation develops. This observation
the kinematic bearing frequencies [69,71]. Finally, is should be highlights the effect of vapour bubble implosion on the tran-
noted that the discussed signal processing methodology could be sient behaviour of the kurtosis of the population.
easily incorporated in algorithms already used in the industry  From the examination of the frequency spectrum in the semi-
towards the fault detection in rotating machines, without increas- log plots it is concluded that cavitation excites wide areas of
ing the cost of the vibration monitoring set-up. the frequency spectrum, regardless of the particular impeller
and sensor mounting position. However, the deviation of the
excitation range and location in the spectrum at different suc-
6. Conclusions tion conditions, as well as the strong similarities with the vibra-
tion spectrum of faulty bearings, question significantly the use
In this study, the Spectral Kurtosis methodology is applied for of such processing method for the efficient detection of
the detection of cavitation in centrifugal pump impellers. Two cavitation.
semi-open radial impellers are tested and their cavitation curves  The results of the SKT reveal the capability of the method to cor-
for various flowrates are derived. At each operating point, the relate vapour formation in the impeller with the appearance of
vibration signal from two sensors at different positions is recorded BPF at the frequency spectrum of the envelope, and also to iden-
and processed in time, in frequency domain and with the use of tify the impulsive signature of cavitation in the filtered time

14
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Φ=0.046, sensor B01

σ=0.52
(green colour)

σ=0.36
(red colour)

σ=0.22
(black colour)
Φ=0.061, sensor A01

σ=0.62
(green colour)

σ=0.39
(red colour)

σ=0.26
(black colour)
Φ=0.076, sensor B01

σ=0.64
(green colour)

σ=0.43
(red colour)

σ=0.36
(black colour)
Φ=0.091, sensor A01

σ=0.72
(green colour)

σ=0.39
(red colour)

σ=0.34
(black colour)

Fig. 16. Frequency domain of the envelope of the vibration signal of impeller No. 2.

domain. The filtered signal in time is formed by random and Further tests and measurements in centrifugal pumps and
periodic impulses, which are masked from various noise hydroturbines of various specific speed, size and suction character-
sources in the original signal. Nevertheless, the successful istics are required in order to evaluate the effectiveness and the
implementation of SK method allows their extraction. The generality of this new methodology for early cavitation detection.
aforementioned behavior of ES and TS is accompanied by the
drop of signal Decomposition Level, observed in all examined
cases. This drop confirms, from the filter construction perspec- CRediT authorship contribution statement
tive, the excitation of wider frequency bands after the develop-
ment of cavitation. Georgios Mousmoulis: Conceptualization, Data curation, For-
mal analysis, Funding acquisition, Investigation, Methodology,
The appearance of the aforementioned characteristics in the TS Resources, Software, Validation, Visualization, Writing - original
and the ES are systematic for both impellers tested, no matter the draft. Christos Yiakopoulos: Conceptualization, Formal analysis,
flowrate conditions, and demonstrates the modulation of the fil- Methodology, Supervision, Writing - review & editing. George
tered signal from the periodicity of the rotating blades. The combi- Aggidis: Data curation, Supervision, Writing - review & editing.
nation of the identification of the bursts that appear in the filtered Ioannis Antoniadis: Supervision, Writing - review & editing. Ioan-
signal, which are periodic with the BPF, along with the appearance nis Anagnostopoulos: Conceptualization, Formal analysis, Fund-
of BPF discrete frequency in the envelope spectrum, constitutes a ing acquisition, Investigation, Methodology, Project
general and robust cavitation detection criterion that can be used administration, Supervision, Writing - original draft, Writing -
in centrifugal pumps. This behaviour of the filtered vibration signal review & editing.
is unique, since in the motor-coupling-pump system only the bear-
ings could excite impulses, however they are modulated by differ-
ent frequencies. Moreover, the fault mechanism (vapour bubble Declaration of Competing Interest
implosion) and the motor-pump system is analogous to that of
the hydroturbine-generator, and this enables the application of The authors declare that they have no known competing finan-
the same criterion towards the detection of cavitation in cial interests or personal relationships that could have appeared
hydroturbines. to influence the work reported in this paper.

15
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

Acknowledgements [28] Dong L, Zhao Y, Dai C. Detection of inception cavitation in centrifugal pump by
fluid-borne noise diagnostic. Shock Vib 2019;2019:1–15. https://doi.org/
10.1155/2019/9641478.
This research has been co-financed by the European Union and [29] Neill GD, Reuben RL, Sandford PM, Brown ER, Steel JA. Detection of incipient
Greek national funds through the Operational Program Competi- cavitation in pumps using acoustic emission. Proc Instit Mech Eng Part E: J
Process Mech Eng 1997;211(1997):267–77. https://doi.org/10.1243/
tiveness, Entrepreneurship and Innovation, under the call
0954408971529737.
RESEARCH – CREATE – INNOVATE (project code:T1EDK-01334. [30] Al-Hashmi SA. Spectrum analysis of acoustic signals for cavitation detection.
2018–2021). In: ISIEA 2012–2012 IEEE Symp. Ind. Electron. Appl. p. 321–5. https://doi.org/
10.1109/ISIEA.2012.6496652.
[31] Cencîc T, Hoĉevar M, Ŝirok B. Study of erosive cavitation detection in pump
References mode of pump-storage hydropower plant prototype. J Fluids Eng Trans ASME
2014;136. https://doi.org/10.1115/1.4026476.
[1] Li SC. Cavitation of hydraulic machinery, vol. 1. World Scientific; 2000. [32] Schmidt H, Kirschner O, Riedelbauch S, Necker J, Kopf E, Rieg M, et al. Influence
[2] Brennen CE. Hydrodynamics of pumps. Cambridge University Press; 2011. of the vibro-acoustic sensor position on cavitation detection in a Kaplan
[3] Brennen CE. Fundamentals of Multiphase Flow. Cambridge University Press; turbine. IOP Conf Ser Earth Environ Sci 2014;22(5):052006. https://doi.org/
2005. 10.1088/1755-1315/22/5/052006.
[4] Gülich JF. Centrifugal pumps, vol. 2. Berlin: Springer; 2008. [33] Bourdon P, Simoneau R, Dorey JM. Accelerometer and pit counting detection of
[5] Schiavello B, Visser FC. Pump cavitation - Various NPSHR criteria, NPSHA cavitation erosion on a laboratory jet and a large Francis turbine. In: XVII IAHR
margins, and impeller life expectancy. In 25th Int. Pump Users Symp.; 2009, p. Symposium. p. 599–615.
113–143. https://doi.org/10.21423/R1XM30. [34] Escaler X, Ekanger JV, Francke HH, Kjeldsen M, Nielsen TK. Detection of Draft
[6] Zhang YL, Yuan SQ, Zhang JF, Feng YN, Lu JX. Numerical investigation of the Tube Surge and Erosive Blade Cavitation in a Full-Scale Francis Turbine. J Fluids
effects of splitter blades on the cavitation performance of a centrifugal pump. Eng 2015. https://doi.org/10.1115/1.4027541.
IOP Conf Ser Earth Environ Sci 2014;22(5):052003. https://doi.org/10.1088/ [35] Hongbin Z, Hui L. Diagnosis of Preliminary Cavitation In Pumps By Wavelet
1755-1315/22/5/052003. Analysis, Energy Resource and Power. Engineering 2010:394–8.
[7] Mousmoulis G, Karlsen-Davies N, Aggidis G, Anagnostopoulos J, Papantonis D. [36] Hajnayeb A, Azizi R, Ghanbarzadeh A, Changizian M. Vibration-based
Experimental analysis of the onset and development of cavitation in a cavitation detection in centrifugal pumps. Diagnostyka 2017;18:77–83.
centrifugal pump. J Phys Conf Ser 2017;813:012044. https://doi.org/10.1088/ [37] Azizi R, Hajnayeb A, Ghanbarzadeh A, Changizian M. Cavitation Severity
1742-6596/813/1/012044. Detection in Centrifugal Pumps. In: International Congress on Technical
[8] Liang D, Yuqi Z, Cui D, Yong W. Research on cavitation acoustic characteristics Diagnostic. p. 47–55. https://doi.org/10.1007/978-3-319-62042-8_4.
of centrifugal pump based on fluid-acoustic field coupling method. Adv Mech [38] Sun H, Yuan S, Luo Y. Cyclic Spectral Analysis of Vibration Signals for
Eng 2018;10:1–13. https://doi.org/10.1177/1687814018773665. Centrifugal Pump Fault Characterization. IEEE Sens J 2018;18(7):2925–33.
[9] Mousmoulis G, Karlsen-Davies N, Aggidis G, Anagnostopoulos I, Papantonis D. https://doi.org/10.1109/JSEN.2018.2804908.
Experimental analysis of cavitation in a centrifugal pump using acoustic [39] Escaler X, Egusquiza E, Mebarki T, Avellan F, Farhat M. Field Assessment of
emission, vibration measurements and flow visualization. Eur J Mech B/Fluids Cavitation Detection Methods in Hydropower Plants. In Proc. Hydraul. Mach.
2019;75:300–11. https://doi.org/10.1016/j.euromechflu.2018.10.015. Syst. 21 St IAHR Symp. Sept. 9; 2002, p. 12.
[10] Brennen CE. Cavitation and bubble dynamics. Cambridge University Press; [40] Escaler X, Egusquiza E, Farhat M, Avellan F, Coussirat M. Detection of
2014. cavitation in hydraulic turbines. Mech Syst Signal Process 2006;20
[11] Cernetic J, Prezelj J, Chudina M. Use of noise and vibration signal for detection (4):983–1007. https://doi.org/10.1016/j.ymssp.2004.08.006.
and monitoring of cavitation in kinetic pumps. J Acoust Soc Am 2008. [41] McFadden PD, Smith JD. Vibration monitoring of rolling element bearings by
[12] Čudina M, Prezelj J. Detection of cavitation in operation of kinetic pumps. Use the high-frequency resonance technique - a review. Tribol Int 1984;17
of discrete frequency tone in audible spectra. Appl Acoust 2009;70(4):540–6. (1):3–10. https://doi.org/10.1016/0301-679X(84)90076-8.
https://doi.org/10.1016/j.apacoust.2008.07.005. [42] Prashad H, Ghosh M, Biswas S. Diagnostic Monitoring of Rolling-Element
[13] Ganeriwala SN, Kanakasabai V. Using vibration signatures analysis to detect Bearings by High-Frequency Resonance Technique. ASLE Trans 1985;28
cavitation in centrifugal pumps, Rotating Machinery, Structural Health (4):439–48. https://doi.org/10.1080/05698198508981641.
Monitoring, Shock and Vibration, Volume 5. Shock 2011;5:307–24. https:// [43] Saidi L, Ben Ali J, Benbouzid M, Bechhoefer E. The use of spectral kurtosis as a
doi.org/10.1007/978-1-4419-9428-8. trend parameter for localized bearing fault diagnosis in induction machines.
[14] Zhang N, Yang M, Gao Bo, Li Z. Vibration characteristics induced by cavitation ISA Trans 2016;63:436–47. https://doi.org/10.1016/j.isatra.2016.02.019.
in a centrifugal pump with slope volute. Shock Vib 2015;2015:1–10. https:// [44] Yiakopoulos C, Gryllias K, Antoniadis I. Spectral effects of the slip motion in
doi.org/10.1155/2015/294980. defective rolling element bearings. Adv Vibr Eng 2005:199–211.
[15] Abdulaziz AM, Kotb A. Detection of pump cavitation by vibration signature. [45] Xiao L, Zhang X, Lu S, Xia T, Xi L. A novel weak-fault detection technique for
Aust J Mech Eng 2017;15(2):103–10. https://doi.org/10.1080/ rolling element bearing based on vibrational resonance. J Sound Vib
14484846.2015.1093261. 2019;438:490–505. https://doi.org/10.1016/j.jsv.2018.09.039.
[16] Lu J, Liu X, Zeng Y, Zhu B, Hu Bo, Yuan S, et al. Detection of the flow state for a [46] Lu S, Zheng P, Liu Y, Cao Z, Yang H, Wang Q. Sound-aided vibration weak signal
centrifugal pump based on vibration. Energies 2019;12(16):3066. https://doi. enhancement for bearing fault detection by using adaptive stochastic
org/10.3390/en12163066. resonance. J Sound Vib 2019;449:18–29. https://doi.org/10.1016/j.
[17] Al-Obaidi AR. Investigation of effect of pump rotational speed on performance jsv.2019.02.028.
and detection of cavitation within a centrifugal pump using vibration analysis. [47] Xiao L, Tang J, Zhang X, Xia T. Weak fault detection in rotating machineries by
Heliyon 2019;5(6):e01910. https://doi.org/10.1016/j.heliyon.2019.e01910. using vibrational resonance and coupled varying-stable nonlinear systems. J
[18] Escaler X, Egusquiza E, Farhat M, Avellan F. Vibration cavitation detection Sound Vib 2020;478:115355. https://doi.org/10.1016/j.jsv.2020.115355.
using onboard measurements. Fifth International Symposium on Cavitation, [48] Antoni J. Cyclic spectral analysis in practice. Mech Syst Signal Process 2007;21
2003. (2):597–630. https://doi.org/10.1016/j.ymssp.2006.08.007.
[19] Escaler X, Egusquiza E, Farhat M, Avellan F. Cavitation erosion prediction in [49] Antoni J. The spectral kurtosis: A useful tool for characterising non-stationary
hydro turbines from onboard vibrations. 22nd IAHR Symposium on Hydraulic signals. Mech Syst Signal Process 2006;20(2):282–307. https://doi.org/
Machinery and Systems, 2004. 10.1016/j.ymssp.2004.09.001.
[20] Wu K, Xing Y, Chu N, Wu P, Cao L, Wu D. A carrier wave extraction method for [50] Antoni J. Fast computation of the kurtogram for the detection of transient
cavitation characterization based on time synchronous average and time- faults. Mech Syst Signal Process 2007;21(1):108–24. https://doi.org/10.1016/j.
frequency analysis. J Sound Vib 2020;489:115682. https://doi.org/10.1016/j. ymssp.2005.12.002.
jsv.2020.115682. [51] International Organization for Standardization. Rotodynamic pumps:
[21] Chudina M. Noise as an indicator of cavitation in a centrifugal pump. Acoust Hydraulic performance acceptance tests: Grades 1, 2 and 3, ISO 9906:2012.
Phys 2003;49(4):463–74. Geneva: International Organization for Standardization; 2012.
[22] Sloteman DP. Cavitation in high energy pumps-Detection and assessment of [52] IEC 60193. Hydraulic turbines, storage pumps and pump turbines Model
damage potential. In Proceedings of the 23rd International Pump Users acceptance tests, (1999–2011).
Symposium, 2007. [53] Farrance I, Badrick T, Frenkel R. Uncertainty in measurement: A review of the
[23] Gopalakrishnan S. Modern cavitation criteria for centrifugal pumps. procedures for determining uncertainty in measurement and its use in
Proceedings of the 2nd International Pump Symposium, 1985. deriving the biological variation of the estimated glomerular filtration rate.
[24] McNulty PJ, Pearsall IS. Cavitation inception in pumps. J Fluids Eng Pract Lab Med 2018;12:1–13. https://doi.org/10.1016/j.PLABM.2018.e00097.
1982;104:99–104. https://doi.org/10.1115/1.3240865. [54] Dwyer RF. FRAM 2 Single Channel Ambient Noise Statistics. In 101st Meeting
[25] Alfayez L, Mba D. Detection of incipient cavitation and the best efficiency point of the Acoust. Soc. of Am.; 1981.
of a 2.2 MW centrifugal pump using Acoustic Emission. J Acoust Emiss 2004. [55] Dwyer RF. Detection of non-Gaussian signals by frequency domain kurtosis
[26] Čudina M, Prezelj J. Detection of cavitation in situ operation of kinetic pumps: estimation. IEEE International Conference on Acoustics, Speech, and Signal
Effect of cavitation on the characteristic discrete frequency component. Appl Processing, vol. 8. p. 607–10. https://doi.org/10.1109/ICASSP.1983.1172264.
Acoust 2009;70(9):1175–82. https://doi.org/10.1016/j.apacoust.2009.04.001. [56] Dwyer RF. A technique for improving detection and estimation of signals
[27] Cernetič J, Chudina M. Cavitation noise phenomena in centrifugal pumps. In contaminated by under ice noise. J Acoust Soc Am 1983;74(1):124–30. https://
5th Congress of Alps-Adria Acoustics Association; 2012. doi.org/10.1121/1.389729.

16
G. Mousmoulis, C. Yiakopoulos, G. Aggidis et al. Applied Acoustics 182 (2021) 108289

[57] Dwyer R. Use of the kurtosis statistic in the frequency domain as an aid in [65] Welch P. The Use of Fast Fourier Transform for the Estimation of Power
detecting random signals. IEEE J Oceanic Eng 1984;9(2):85–92. Spectra: A Method Based on Time Averaging Over Short, Modified
[58] Randall RB, Antoni J. Rolling element bearing diagnostics-A tutorial. Mech Syst Periodograms. IEEE Trans Audio Electroacoust 1967;15(2):70–3. https://doi.
Signal Process 2011;25(2):485–520. https://doi.org/10.1016/j. org/10.1109/TAU.1967.1161901.
ymssp.2010.07.017. [66] Taylor JI. The vibration analysis handbook; 2005.
[59] Randall RB. Vibration-based condition monitoring: industrial, aerospace and [67] Scheffer C. Pump condition monitoring through vibration analysis. In: Pumps:
automotive applications. John Wiley & Sons; 2011. Maintenance, Design, and Reliability Conference. p. 1135–46.
[60] Yiakopoulos CT, Gryllias KC, Antoniadis IA. Rolling element bearing fault [68] Randall RB. Noise and vibration data analysis. Handbook of Noise and
detection in industrial environments based on a K-means clustering approach. Vibration Control. John Wiley & Sons; 2007.
Expert Syst Appl 2011;38(3):2888–911. https://doi.org/10.1016/j. [69] Zhang X, Kang J, Xiao L, Zhao J, Teng H. A new improved Kurtogram and its
eswa.2010.08.083. application to bearing fault diagnosis. Shock Vib 2015;2015:1–22. https://doi.
[61] Gryllias KC, Antoniadis IA. A Support Vector Machine approach based on org/10.1155/2015/385412.
physical model training for rolling element bearing fault detection in [70] Li Li, Li M, Chen F, Zhang D, Zi B, Cui G, et al. A Fast Kurtogram Demodulation
industrial environments. Eng Appl Artif Intell 2012;25(2):326–44. https:// Method in Rolling Bearing Fault Diagnosis. MATEC Web Conf 2016;77:01003.
doi.org/10.1016/j.engappai.2011.09.010. https://doi.org/10.1051/matecconf/20167701003.
[62] Marple L. Computing the discrete-time ‘‘analytic” signal via FFT. IEEE Trans [71] Antoni J, Randall RB. The spectral kurtosis: Application to the vibratory
Signal Process 1999;47(9):2600–3. surveillance and diagnostics of rotating machines. Mech Syst Signal Process
[63] Cudina M. Pumps and pumping system noise and vibration prediction and 2006;20(2):308–31. https://doi.org/10.1016/j.ymssp.2004.09.002.
control. In: Crocker MJ, editor. Handbook of Noise and Vibration [72] Sawalhi N, Randall RB. Spectral kurtosis optimization for rolling element
Control. Hoboken, NJ, USA: John Wiley & Sons, Inc.; 2007. p. 897–909. bearings. In Proc. - 8th Int. Symp. Signal Process. Its Appl. ISSPA 2005, vol. 2;
https://doi.org/10.1002/9780470209707.ch73. 2005, p. 839–42. https://doi.org/10.1109/ISSPA.2005.1581069.
[64] Anderson HH. Centrifugal pumps and allied machinery. Elsevier; 1994.

17