Proceedings of the 2011 IEEE

International Conference on Mechatronics

April 13-15, 2011, Istanbul, Turkey
NIUt3!IOH bI_H3!utC /H3!ySIS IOt C!CC!IH_ L3VI!3!IOH
IH LCH!tIIu_3! uH]S uSIH_^ Cut3!^C!WOtKS
M.R. Nasiri, M.J. Mahjoob, H. Vahid-Alizadeh
NVA Research Center, Shool of Mechanical Engineering
Universit of Tehran, Tehran, Iran
nasiri.mech.eng@gmail.com, mmahjoob@ut.ac.ir, h.vahid.alizadeh@gmail.com
Abstract-Vibration analysis is applied to detect cavitation in a
centrifugal pump using a neural net system. The features
extracted from vibration signals are used as inputs to the neural
network. The output data of the system is set as 0, 0.5 and 1, for
normal condition, developed cavitation and fully developed
cavitation, respectively. Experiments are also conducted to
validate the developed model. The method provides an
intelligent system to be used in condition monitoring of
centrifugal pumps. Also the number of sensors and the best
sensor positions are studied.
Kewords- vibration signal processing; cavitation;
centrifugal pump; neural network
I. INTRODUCTION
Fault detection of engineering systems especially in
mechanical ones is a growing feld of research wit importt
industial applications. Centifgal pumps have vaious types
ad sizes wit vaiety of applications. Pumps are one of te
major components i maufactuing ad processing industres
such as oil, gas, petrochemical, power generation, food
processing, agiculture, and may oter industies.
I many cases, because of te critical role of pumps, te
incidence of any fault or failure may cause malfction or
even shutdown in process plants. When a defect is begining
to grow in a machine, te vibrtion levels and specta are
ofen afected. Therefore, the vibration sigatues of a
machine ca be used as a means to detect the defects in the
equipment [ 1].
Today, accelerometers ae used as te most suitable
vibration sensors due to teir large fequency and dynaic
rage. I addition, they are easy to install ad reliable for
measurement. Vibration sigals picked by accelerometers can
be prcessed in different ways to analyze diferent types of
data. They can then efectively be used to detect diferent
status (healty/faulty conditions) of runing machines. Major
techniques used in tis feld ae time domain, fequency
domain ad time-fequency domain techniques [2].
Cavitation is a complicated phenomenon. It is te main
destctive cause in te critical components of pumps such as
impeller. It creates oter adverse efects such as dop in te
pump head ad efciency, corosion ad undesirble
noise/vibration. However, detection of tis phenomenon in te
early stages is hadly possible [3].
978-1-61284-985-0/11/$26.00 2011 IEEE
632
Diagnosis of cavitation by vibration sigal has been
conducted using different data processing techniques. I te
time domain, the sigal rs tat indcates the signal energy
remains nearly constat, but peak signal levels increases
sigifcatly [4]. Thus te ratio of peak value to rs, Crest
Factor, ca be taken a measure to detect cavitation. I te
fequency domain [5], the advent and prgess of cavitation
can also boost up te high-fequency pa of te power
spectm of the sigal.
Diferent techniques based on aifcial intelligence have
been developed as alterative methods for fault detection in
engineering applications. Trubleshootig wit the help of
fzzy logic ad vibrtion signal processing has been
investigated [6]. The results were not always reliable. I
reference [7], fzzy logic decision tee is used reporting tat
te metod gives better results in detecting defects.
The metod presented here employs some selected
features of the vibration sigal (in time/fequency domain) as
te inputs of a multi layered neurl network. The network is
specifcally applied to detect cavitation status of centifgal
pumps. Experiments ae desiged and caried out which
support the results of te developed method.
We will frst review some theoretical backgroud on
cavitation in section II. The features obtained fom sigal
processing to detect cavitation are also discussed. I section
III, the test set up ad the data acquisition system are
presented. Section I includes the test results and
computational analysis of the experimental data. Section V
concludes te mauscript.
FigJ Test equipment
II. THEORY
Cavitation in fuid fow is a specifc two-phase
phenomenon wit phase chaging (evaporationcondensation)
due to decrease in pressure witout heating. Cavitation occurs
in a paicular location (M) ad temperature (T), whenever the
fuid pressure P reaches the saturation vapor pressure Pv(T);
namely in ( 1):
PM (T) : Pv(T) (1)
I order to have a qualitative expression of te aout of
cavitation development, two cases will be dstinguished:
developed ad flly developed cavitation. The forer is taken
as the rage between the occur ence of cavitation ad te
noral state in a fuid fow, i.e. te moments the frst bubbles
for until the state of stable cavitation where te hydaulics
perforance drops noticeably. I te latter (flly developed
cavitation), bubble foration maintains continuously; there is
a steady state in foration of bubbles. I tis condition, the
hydraulic perforace of te mache deteriorates
sigifcantly; te cavitation efect is no longer limited to the
foration of local bubbles.
The fault detection approach here is based on vibration
sigatue aalysis. The data type used as input is te
acceleration (sigal) record at different times. Time domain
sigal processing methods use only te time-tace of te
sigal for analysis. There is no oter tsforation applied
on te time data. I statistical and radom processes, the
different moments of te sigal, have special importace, and
widely used as features to discriminate data.
The fort moment is called Kurtosis.
N
(xi -f)
4
kurtosi
=
-i="I
0
4
(2)
Where in (2) Xi'S are the elements of the rndom vaiable X, I
is te mea value and 0 is te variance. According to tis
defmition, the Kurtosis value of a noral distibution is zero.
Kurtosis is a measure of whether the data are peaked or fat
relative to a noral distibution. That is, data sets wit hig
kuosis tend to have a distinct pea nea te mean, decline
rater rapidly, ad have heavy tails. Data sets wit low
kuosis tend to have a fat top near te mea rather tan a
shar pea. Crest-Factor is defmed as the ratio of the peak
sigal amplitude to the rs.
F
Pea level
Crest actor
=

RS
(3)
The purose in calculation of te Crest-Factor as
mentioned in (3) is to provide a quick idea of the amount of
impulses existing in te sigal. The time domain data only
offers limited inforation about te sigal. Hence, the signal
is usually tansfored fom one domain to anoter to reveal
some usefl inforation. The tansforation used fom time
domain to fequency domain is te Fouier transfor. The
Discrete Fouier Transfor (DFT) used in most computations
is defmed in (4) as follows.
-i211
k
x(/)= _.(-) N
(4)
Afer calculating te Discrete Fouier Transfor of a
sigal, a complex vector is obtained that contains inforation
about the fequency content of te sigal (magnitude ad
phase at diferent fequencies). The magnitude is used here as
a fequency domain signal in (5) as follows:
s(/)-x(/) (5)
Using statistical chaacteristics of this sigal S(K), two
features in te fequency domain are extacted:
1
K
}~_s(|)
K
k=l
1
K
2
F
2
~_,s,|)~:
K
-1
k=l
(6)
(7)
We will use the above features (6) ad (7) as te inputs to
a multi layered neurl network. It consists of an
interconnected group of artifcial neurons which processes
inforation using a connectionist approach to computation.
The neural network is used to estimate te extent of cavitation
based on the above featues. The details of te neurl network
used here will be explained in the following sections.
Tale 1: THREE SAMPLES of PROCESSED DATA
Crest Factor Kurtosis F
I
F
2
Outut RM
Radial Back Front Radial Back Front Radial Back Front Radial Back Front
sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor sensor
0 1800 0.1927 0.1507 0.1255 3.1892 3.5717 3.8117 0.0108 0.007 0.0027 0.0029 0.0012 0.0001
0.5 2000 0.2235 0.166 0.1304 2.6951 2.653 2.3661 0.0284 0.0254 0.0066 0.0198 0.0132 0.0002
1 1900 0.2094 0.1582 0.1681 2.4391 3.0765 2.3467 0.008 0.0089 0.0032 0.0014 0.0014 0.0001
633
Fig.2 Position of te sensors attached to te pump, fom lef to right, Front,
Radial ad Back.
III. EXPERIMENTAL WORK
The experimental appaatus used in te tests is
maufactued by Tecquipment of Englad (Fig. 1) ad includes
a centifgal pup mouted on a hydaulic table. Water fows
out of the t troug a valve ad enters the pump through a
contol valve. The pup outut fow enters a separated tank to
measure the fow rate. There are two contol valves: suction
contol valve (valve A) located before pump, ad te outfow
contol valve (valve B) located afer te pup. Two Boudon
maometers show te input and outut pressures of pump. A
digital tachometer indicates te pump rotation speed, RPM.
The impeller is made of corosion resistat bronze. It has 6
blades wit entance angle of 75 degrees ad exit agle of 30
degrees. The maximum angular velocit of the pup is 2900
rm. The maximum speed canot be more tan 2300 rm,
because of stong cavitation. The power of pup diving motor
is 2.2 KW. The motor maximum speed (pump not engaged) is
3000 rm ad wit a nomal electic curent is 15 A.
I ou tests, thee DJ accelerometers are used to record
te vibration sigals. Two of tem ae mounted on te font
and back plate of te pump casing (axial direction) and the
tird is attached in te radial direction. (Fig. 2)
The vibration data-logger and analyzer used to record the
sigals is B&K 3560C along wit Pulse sofwae, recording
sigals simultaeously in six channels. The upper fequency
limit of te system is 25 Kz. The maximum sampling
fequency is set about 2.56 times te maimum fequency of
RPM
C",.F,ct. {
Radial
Back
Front
KWO',{
Radial
Back
Front
{
Radial
FI
Back
Front
{
Radal
F
l
Back
Front

% *
.
',
.

_
_
...... '
. ,". f
o

.
0) (Developed)
~
1 (Ful Developed)

Pe Le
.
Hyperolc

Tangent Sigoid
Hyperolc
Tangent Sigoid
Fig.3 Neual network
te input data. The device also has an interal low pass flter
to eliminate fequencies above half of te sampling fequency
to prevent Aliasing. The hadware is accompanied by te
sofware 'Pulse Labshop' for basic vibration analyses. Duing
te data captue, te sampling fequency was set at 10 Kz,
terefore te maximum meaningfl fequency was
10/2.56=4Kz. The time duation for taing the time-series
data was 10 seconds. Cavitation is made in the pump at
specifed speed of pumps for several different outlet-valve
conditions. The frst situation of te valve was fuly open.
Cavitation is confmed by observation of pump fow through
its transparent wall with stroboscope. I each opening-status
of te outut valve, the input valve was closed gradually till
observing te cavitation bubbles. Few seconds was needed to
have stable bubble foring where the fow rate of the pump
and input/outut pressures were recorded. In the last stage, the
data recordng was activated and te signals (accelerometers
and tachometer) were recorded for 10 seconds.
j

_
_

r_
_

_
_
_

_
_

B
=

_
O
_

j
. """"" L>O

B |B
_
.B
i 0.7
.

0
i
:
0.4
U
0.1
-0.1
-0.2
0 50
g

100 150 200

Train Data (Neural Netork Input)
Fig.4 The sae result of experiment ad network
634
250
.
I
.
Lr3O3
LBr.B.3
7B
300 350
10
8
..
6
0
t:
Q
4

2
0
radial back
9.375
front
0.2
radial-back radial-front
o
back-front radial-back-front
Fig.5 The efect of number ad positions of the sensors on the neural network results.
This operation has been done for diferent speeds of pump,
fom 1300 rm to 2000 rm, and at diferent opening-states of
te outut valve. It should be noted tat closing te input valve
to establish cavitation, is associated with fow rate reduction.
Therefore to have two compaable sets of data, the faulty
condition has to be created frst. Then wit its fow rate at
paicula speed, normal condition is created. Consequently
there will be two series of data at the same fow rate ad speed.
I. EXPERIMENTAL RESULTS A ANALYSIS
As it was mentioned in te previous sections, we have
obtained three sets (time taces) of data fom sensors. Table 1
shows tree samples of processed data achieved by using the
proposed featues.
As shown in te Fig.3 the tree processed series of sensors
outut plus te speed of the pup ae considered as te iputs
of a neural network. The neural network consists of two hidden
layers. The frst layer contains 7 neurons wit Hyperbolic
Tagent Sigmoid tasfer fnction.
The second layer consists of two neurons wit the same
fctions, and te third layer includes one neuron wit Pure
Line transfer fnction. The outut is one of the nubers: 0
(zer, noral condition), 0.5 (developed cavitation) ad I
(flly developed cavitation).
Totally we have 320 series of processed data, in three
conditions namely noral, developed and flly developed
cavitation. 270 data sets ae used for training te network ad
50 data as the test date. The feed forwad back prpagation
metod is selected as te taining stategy. The results of te
experment in three cavitation conditions are illustated wit
dashed red line and te network output is show wit solid
blue line in Fig.4. The results of laboratory tests agree with te
ones predicted by te artifcial neurl network. This shows that
the network is a reliable model to estimate the pup condition
in te tree cases defned in tis case study.
At the end, the network eror based on location of sensors
installed on te pup and te numbers of tem are compared
in FigA. This fgue ca be considered as a proposed method
for selecting te number and positions of te sensors.
635
v. CONCLUSION
A method was presented to automate the cavitation fault
detection in centifgal pups using vibration signatue
aalysis. A neual network was used to identif tree
healty/faulty conditions naely noral, moderate ad flly
developed cavitation. Experiments were also conducted to
evaluate te presented model. The results were satisfactory ad
te fault status of the pup could be identifed automatically
be te computer which was fed wit te vibration sigals fom
te pump casing. As shown in Fig.5, if just one sensor is
available it should be used in radial position, ad also if two
sensors are available they should be used in rdial ad back
positions ad if tree sensors are available te cavitation will
be detected witout any eror by placing the sensors in radial,
back ad font positions.
REFERENCES
[ 1] P. Peck, J. Burows, "on-line condition monitoring of rotating equipment
using neural network", ISA Trasactions 33 (1994) 159-164.
[2] L. Alfayaz, D. Mba, G.Dyson, "The application of acoustic emission for
detecting incipient cavitation ad the best efciency point of a 60KW
mono block centifgal pump", NDT&E Interational 38 (2005) 354-
358.
[3] Mc.Nulty,P.J ad Peasall,I.S, "Cavitation Inception in pumps.",Joual
of fuids engineering,VoU 04,99-104.1982
[4] S.C Li, " Cavitation of hydraulic Machinery" ,London Imperial College
Press.2000
[5] Beyer,G.Smit, "Acoustic emission chaacteristics of a modal facis
turbine under cavitating ad non-cavitating condition", Proceeding
interational symposium on cavitation noise,ASME,(1982) 87-92
[6] J. Wag, H. Hu, "Vibration-based fault diagosis of pump using fz
technique", Measurement 39 (2006) 176-185.
[7] N.R. Saivel, V. Sugua, S. Babudevasenapati, "Vibration based
fault diagosis of monoblock centifgal pump using decision tree",
Expert Systems with Applications 37 (2010) 4040-049