See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/321543275

Centrifugal Pump Cavitation Detection Using Machine Learning Algorithm

Technique

Conference Paper · June 2018

DOI: 10.1109/EEEIC.2018.8494594

CITATIONS READS

19 7,629

6 authors, including:

Nabanita Dutta Arun Shankar V.K.

VIT University VIT University
21 PUBLICATIONS 73 CITATIONS 33 PUBLICATIONS 375 CITATIONS

SEE PROFILE SEE PROFILE

Umashankar Subramaniam P. Sanjeevikumar

Prince Sultan University Aarhus University
374 PUBLICATIONS 3,361 CITATIONS 954 PUBLICATIONS 10,035 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Investigations of Interleaved DC-DC Boost converter for Solar PV Applications View project

Power Quality Survey @ Traction Substation, tiruvalam and Valathoor, Southern Railway View project

All content following this page was uploaded by Zbigniew Leonowicz on 18 July 2018.

The user has requested enhancement of the downloaded file.

 Centrifugal Pump Cavitation Detection Using
Machine Learning Algorithm Technique
Nabanita Dutta S.Umashankar Arun Shankar V.K.,
Dept. of Energy & Power Electronics, Dept. of Energy & Power Electronics, Dept. of Energy & Power Electronics,
SELECT, VIT University, Vellore, India SELECT, VIT University, Vellore, India SELECT, VIT University, Vellore, India
nabajhilikbarbi@gmail.com umashankar.s@vit.ac.in arunshankarvk@vit.ac.in
Sanjeevikumar Padmanaban Zbigniew Leonowicz Patrick Wheeler
Dept. of Energy Technology, Faculty of Electrical Engg., Dept. of Electrical & Electronics Engg/,
Aalborg University, Esbjerg, Denmark. Wroclaw University of Technology, Poland. Nottingham University, United Kingdom.
san@et.aau.dk leonowicz@ieee.org pat.wheeler@nottingham.ac.uk

Abstract—Cavitation is one of the major disadvantages in

pumping system, which enhance to form bubbles in the pipeline
and it reduces the efficiency of the pump. So it should be
identified and take the preventive measure. Machine Learning is
a fast and computational method which can easily detect any
faults in the pumping system. Still now lots of work has been
done on a detection of fault in the pumping system, but mainly
those work has done based on vibration details and variation of
speed. The paper presents how by the help of machine learning
algorithm by varying the speed and pressure cavitation can be
identified. It is the comparative study between how the vibration
and speed together affects the cavitation result and variation of
speed and pressure affects the cavitation. Support Vector
Machine is one of the classification methods in machine learning Fig. 1. Common Faults in Pumping System.
algorithm where it can be easily classified the cavitation problem.
So this paper analyses how the method of SVM can more industry, mainly centrifugal pumps are working. In general
efficiently detect the cavitation problem with the centrifugal pump is two types, Rotodynamic type pump, and a positive
water pump. displacement pump. On the basis of case study rotodynamic
pump works as 73% and positive displacement pump work as
Keywords—Centrifugal pump, faults, SVM, machine learning 27%. The centrifugal pump is the part of roto-dynamic type
algorithm, cavitation. pump. Among all the rotodynamic pumps the use of a
centrifugal pump is around 16%. The pumping system
I. INTRODUCTION performance depends on the Head, and Flow rate of the pump.
The most essential part of the creature is water. Without In a rated speed for a constant flow rate Head value is same.
water, it will be very difficult for the existence of any creature. For different speed Head value will change as the flow rate will
Humans use water for various purposes like domestic purpose, change. Impeller diameter also has the vital role in the system
cultivating, power generation and so on. On an average around performance curve. However, centrifugal pump can operate in
150 liters of water has been consumed by an individual in India a wide range of flow rate with respect to Head value. The
on every day. India has 17% of the total world population, but pumping system can be either a closed loop or open loop
it has only 4% of fresh water resources. So the demand of system. For open loop the pumping system static head (Hst)
water is increasing every day. The statistics is only for remains constant and the dynamic head varies with respect to
household purpose. When it comes through the power the flow rate. The performance curve is based on head value,
generation, irrigation and commercial purpose the demand is speed and Flow rate. The header value, flow rate and speed
too high. Agriculture is backbone of India and demand of water depend on the affinity law other pump. The affinity law is
in agriculture is more. Pumping system is used in agriculture employed in the pumping system based the hypothesis that the
for irrigation purpose, for cultivation of the crops. Pumping pump’s efficiency is maintained constant, though there is a
system is a mechanical device which moves the fluid in a change in the pump speed.
mechanical way. The main work of the pump is to suck the
water from the ground and discharge the water. It is a cyclic Hsys = Hst + Hdyn (1)
process. The process can be done by both open loop system H dyn = K .Q 2
(2)
and closed loop system. Pumps can be classified into many
Q1 N1
groups according to the fluid displacement method it can be = (3)
categorized like centrifugal pump, impulse pump, gravity Q2 N2
pumps, velocity pumps, stream pumps, valve less pumps. All
these pumps are driven by an induction motor mainly. In

978-1-5386-5186-5/18/$31.00 ©2018 IEEE

 Fig. 2. Types of Machine Learning Algorithm

2 pressure in the pump is called net positive suction head

H 1  N1  available. When the net positive suction head required value
=  (4)
H2  N2  exceeds the net positive suction head available it will show
3 cavitation in the system. It can be taken the value of net
P1  N1  positive suction head in constant speed and variable speed also.
=  (5)
It can be calculated net positive suction head from constant
P2  N 2 
speed and flow rate, from impeller erosion, from static
The cavitation is the creation of bubbles is the pipeline or in pressure, from vibration details etc.
impeller. Mainly there are five causes for the creation of
cavitation like failure of pump housing, destruction of impeller, II. RELATED WORK
excessive vibration, higher than necessary power consumption, A novel method is used to detect cavitation by using
decreased flow and pressure. There are mainly five types of frequency converter as a tool [1]. When the internal parts of the
cavitation in the pumping system like vaporization, turbulence, pumps are break down then pump performance also decorates.
vane syndrome, internal re-circulation, air aspiration cavitation. Using of sensor make the high costing in the system. For two
There are several method are for detecting incipient different pumping system this method is applicable for
cavitation. But an algorithm which should be reliable and has detecting high flow cavitation.
idea to detect and classify the fault is very rare. So the machine
learning algorithm this new concept has been introduced to
detect the fault easily. Various types of machine learning
algorithms are available. Among those Support Vector
Machine, Artificial Neural Network, K-Nearest Neighbor
Method, Empirical Mode of Decomposition are the common
algorithm for detecting the pumping system faults such as
cavitation, sludge, water hammering etc.
Fig. 1 presents the common fault of the pumping system.
and the Fig. 2 present the types of Machine Learning
Algorithm. The pump can rotate at any speed. But Every Pump Fig. 3. Schematic diagram of Test rig software
has a rated speed. So Beyond that speed if the speed can be
increased Flow rate will be high and vacuum will be created. Empirical mode of decomposition and Intrinsic mode of
For this reason air bubbles will be created and cavitation will Function has been used for improving the vibration data for
occur. So to prevent cavitation mainly Speed should be cavitation detection in hybrid feature or three deferent pumping
reduced, then flow rate also will be reduced and cavitation will system vibration data have been noted for model based
be prevented. The speed depends on pressure, flow rate and system[2]. The generalized neural regression network
Head value of the pump. Pump manufacturers provide one algorithm is used for fault classification which accuracy is
pressure in the suction side of the pump exceeds the vapour more than 97.5% high than artificial neural network.10 fold
pressure to prevent the pump from cavitation. This is known as cross validation has been done for getting accurate result which
net positive suction head required. This is the base of pump is 100% accurate than all the results of artificial neural
design and manufacturing process. It varies with pump network. By selecting this feature system can get multi-layer
rotational speed. At zero speed it does not tend to zero. It will perception. Here Best IMFs has been detected for identifying
be considered that NPSHr will vary with the variable speed. the cavitation. In centrifugal pump back propagation learning
Even if in the low speed also there will be some value of algorithm, multilayer perception neural network has been used
NPSHr. On the other hand pressure exceeds the vapour to detect the system cavitation after a particular suction
head[3]. Here Fast Fourier transformation and artificial neural
network has been used for analyzing variance, skewness.
Results will highlight the 100% accuracy in the neuron layer of
artificial neural network. In variable speed vibration data have
been identified to detect the cavitation [4].The probability of
occurring cavitation depends upon the rotational speed and
sound signal.6 blade motor has been used here. Each blade
vibration details has been noted and each blade passing the
frequency which creates vibration septra.NI92432 data
acquisition amplifier card based system used here where
provision for measure pressure in different speed, temperature Fig. 5. a) SVM Classification for pattern class[10]
transducer is also used for measure temperature in different
speed. Labview software is used to note down the flow rate,
sound level, pressure etc. In different speed NPSHa and NPSHr
value has been noted down. If NPSHa > NPSHr then there will
be no cavitation otherwise it will show cavitation. In their
pump after 2100 RPM cavitation will show as it is the rated
speed. Here Blade passing frequency is used to detect the
rotational speed and vibration septum is used to for vibration
spectrum.
Fig. 3 presents the block diagram of Test rig software. Fig. 5. b) Boundary SVM [10]
Cavitation is very harmful to any water turbine. So to prevent
water turbine from cavitation on line monitoring system is used
[5]. In Kaplan water turbine also vibration details has been
noted down. The lab view 8.0 software is used to acquire the
signal. Sixth order Butterworth filter is used to extract the
features of the velocity and make it in order. Model based
cavitation indicator is used here to observe the data. Fuzzy
neural network is presenting the classification problem. Quasi
synchronous vibration is another vibration process which can
easily identify the vibration details and the process in a way
which can be categorized the fault in a proper way
[6].Vaporization process, flash break also done to get more Fig. 6. Accuracy Rate of Machine Learning Algorithm
accurate result. Piezoelectric acceleration signal meter is used
to measure the signal details. Mono-block centrifugal pump is
used many cases [8]. If cavitation occurs in Mono-block pump
if creates lots of problem. So continuous monitoring is needed.
Machine learning method will be the best method for
continuous assessment. So Artificial neural network is used
here to detect the fault.c4.5 decision tree is used to classify the
fault.

Fig. 7. Block Diagram of SVM Algorithm

Fuzzy interface system formed with rough set is popular to

test data with high accuracy level compared with old statistical
data [10]. It is the multilayer perception based on classifier that
helps to reduce the centrifugal pump breakdown. Here rather
than fuzzy logic proximal support vector machine is used for
its high memory power and complexity. Here pattern based
algorithm is used to classify each group of fault. Winger
distribution is used to analysis vibration details. Rough set
generate statistical details for good feature extraction. Load
torque signature analysis is one of the method to detect
cavitation in centrifugal pump.[12] It is the computer based
tool process where vibration details are identified by fast
Fourier transformation method. This is the mathematical
approach by which it can be easily detected the faulty condition
Fig. 4. Flow chart of Monoblock centrifugal pump diagnosis system by 10 fold cross vibration analysis. Some industry use butterfly
valve for pipeline water flow. [16].It is very sensitive device
and when cavitation occurs noise is created. So to prevent on-faulty conditions. The parameter block indicates which
cavitation and to identify the fault support vector machine is parameters are taking, such as voltage current, speed, flux,
used to control the valves by trained data. This classification pressure, etc. here speed, pressure, current, the flow rate value
accuracy rate is higher than artificial neural network. are taken. SVM allows to plot up to n dimension. So by curve
fitting tool, scaling can be done for the value also if needed.
Fig. 4 presents the flow chart of Mono-block centrifugal For checking the accuracy of the system, it has been taken
pump, Fig. 5 and 6 presents the SVM pattern, Fig. 7 presents some sample data up to 500 data in one group. SVM allows the
the accuracy rate of the algorithm of machine learning, Fig. 8 formation of the model according to the trained parameter. By
represents the block diagram of the general SVM algorithm using test data can be checked for the output and verify the
and Fig. 9 represent the flow chart of EMD algorithm. algorithm is working. By using a prediction method it can be
predicted the output of the proposed algorithm. The output data
will be received by the based on training data. It is the
polynomial function which is used in the algorithm.
IV. PROPOSED WORK PROCEDURE
In the paper the speed and vibration analysis details have
used to detect the cavitation in the centrifugal pump. But the
system is very complex. Here no machine learning process has
been used. The total system is a test based hardware process.
On the basis of net positive suction head available and net
positive suction head require it conclude that after which speed
in the pipe cavitation is seen.24 bit data acquisition card based
amplifier system is used. The formula used here,
NPSH A = NPSH R + 0.6 (6)
When the value of NPSHR will be higher than the value of
NPSHa then the system will show the cavitation. Here it had
made the process with six different speed and after 2100 RPM
Fig. 8. Flowchart on EMD Algorithm their system cavitation as 2100 RPM is their rated speed.
NPSHa Value can be identified from atmospheric pressure,
III. PROPOSED ALGORITHM gauge pressure, friction loss and static head. Static head (Hs) is
The proposed algorithm comprises of Support vector the head difference between suction head and discharge head.
machine which is based on automation technology and the part
of artificial neural network. It can store previous data, store Total Head (H) = Static head Dynamic Head (Hd)
data and classify the data also. A brief description of machine Dynamic head can be calculated from the system friction
learning algorithm and the method of classification process will based on Darcy Weisbach equation.
be discussed with the block diagram shown in Fig. 10.
kv 2
HD = (7)
2g
When Q is the formula for identifying the flow rate and
area is the cross sectional area can be defined by and v is the
velocity of the pump. It can be written the set of formulas
which are the base of identifying the dynamic head of the
pump.

π D2
A= (8)
4
Q
Fig. 9. Block diagram of implementation of SVM
v= (9)
A
In the proposed system first the data have to from the K = K fittings + K pipe (10)
system parameter, and then feature should be extracted to
verify that margin is compatible or not, if not then margin fL
should be scaled down. Then a sample of trained data will be
K pipe = (11)
D
taken for data acquisition and verify the algorithm is working f can be identified by Colebrook White equation. The
or not. Then the test data will be loaded according to the equation depends on the material of the pipe and internal
system parameters are needed. Then from the test data can be coatings.
classified the faulty condition. A margin will show as a hyper-
line and it creates the separation between faulty condition and
 0 .2 5 V. MODIFIED PROPOSED SYSTEM RESULT
f =
k 5 .7 4 2 (12)
[lo g { + 0 .9
}]
3 .7 × D Re
So Net Positive Suction Head Available is equal to
Atmospheric pressure(Converted to head Value)+ Static
head±Surface Pressure head-Vapor Pressure Head-Loss of the
pipe fittings (converted to head value).Net Positive suction
head required will be calculated from flow rate and speed of
the pump.
Fig. 11. Pump Performance Curve Fig. 12. Pump Performance Curve
TABLE I. SAMPLE DATA OF THE WORK
Rotational NPSHA NPSHR Cavitation Cavitation
Speed (RPM) (m) (m) Status Strength (%)
1476 0.754 0.40 Free NA
1644 0.754 0.48 Free NA
1932 0.754 0.66 Free NA
2190 0.754 0.84 Cavitation 10.24
2466 0.754 1.80 Cavitation 58.10 Fig. 13. Pump NPSHr Curve Fig. 14. Pump NPSHa and NPSHr
2682 0.754 1.27 Cavitation 40.63 curve

V. MODIFICATED PROPOSED WORK

If a modification has been done by machine learning
algorithm and use SVM algorithm in the point pump whose
rated speed is 2800 RPM it can be used the same procedure but
it will be modified in the angle of machine learning
background. So it can easily detect the flow rate and the speed
of the pump from that it can identify the cavitation fault.
Fig. 1. Classification of fault Fig16 Classification of fault (readings
(readings taken from ref 4, follows table 2)
follows table1)
Here all the data are put in the system after 2800 RPM it
will show the cavitation fault. The main advantage is no need
for manual control of flow rate because it depends on the speed
value so if in the master follower pump if 3 pumps will be run
simultaneously it can get a different flow rate in different
speed. From that it can be calculated net positive suction head
required and net positive suction head available.Then it is
possible to plot pump performance curve and pump NPSHr
curve, and using a support vector machine algorithm for the
trained data can be classified as the cavitation fault.

TABLE II. MODIFIED DATA FOR SVM ALGORITHM

Fig. 10. Algorithm for the modified proposed work Rotational NPSHA NPSHR Cavitation Cavitation
Speed(RPM) (m)(taken (m) Status Strength
Fig. 11 represent the algorithm for the modified work. as Ref 4) (%)
500 0.754 0.25 Free NA
Step1: Take the data from the real time or sample data from 550 0.754 0.65 Free NA
the pumping system. 1400 0.754 0.75 Free NA
1772 0.754 0.95 Cavitation 20.63
Step2: Determine the atmospheric pressure (Patm), dynamic
2200 0.754 1.8 Cavitation 58.11
head (hd), and Static head (hs), vapour pressure (pv), gauge 2690 0.754 2 Cavitation 62.30
pressure (pga), friction head (hl). 2778 0.754 2.5 Cavitation 69.84
2910 0.754 4 Cavitation 81.15
Step3: Calculate the value of net positive suction head
2934 0.754 7 Cavitation 89.22
available and required.
Step 4: The system will check whether the net positive Table 1 describes the NPSHa and NPSHr value, cavitation
suction head available is greater than required value. status and its strength in different speed. From that it can be
notified that after 2100 RPM cavitation fault occurs and its
Step 5: If NPSHa > NPSHr then it will indicate no highest cavitation strength is 40.63%. Table 2 is the modified
cavitation in the system otherwise it will indicate the cavitation work of table1.Here it is notified that after 2800 RPM
fault. cavitation occurs and its highest cavitation strength is 89.22%.
 Fig. 12 presents the Head and Flow rate value of the technique,” Measurement, vol. 108, pp. 9–17, 2017.
pumping system of both the sample work and modified [3] S. Farokhzad, H. Ahmadi, A. Jaefari, M. R. A. A. Abad, and M.
work.Pump1 curve presents sample work value and pump 2 R. Kohan, “Artificial neural network based classification of
curve presents modified work value. Those curves are called a faults in centrifugal water pump,” J. Vibroengineering, vol. 14,
pump performance curve. Similarly Fig. 13 presents Pump no. 4, pp. 1734–1744, 2012.
NPSHr curve for both sample and modified work.pump1 result [4] A. Kotb and A. M. Abdulaziz, “Cavitation Detection in
presents sample work data, pump 2 results present modified Variable Speed Pump by Analyzing the Acoustic and Vibration
work data. Fig. 14 presents NPSHa and NPSHr curve of both Spectrums,” no. October, pp. 706–716, 2015.
[5] S. Lahdelma and E. K. Juuso, “Vibration analysis of cavitation
sample and modified data.Pump1 result similarly present
in Kaplan water turbines,” IFAC Proc. Vol., vol. 17, no. 1
sample work data and modified work data represented by pump PART 1, pp. 13420–13425, 2008.
2 data. Fig. 15 is the classification result of the cavitation fault [6] Y. NI, “Detection of Cavitation in Centrifugal Pump By
of sample work and Fig. 16 is the classification result of Vibration Methods,” Chinese J. Mech. Eng. (English Ed., vol.
cavitation fault of modified work. Automation Technology and 21, no. 5, p. 46, 2008.
Artificial Intelligence are now ruling the world. Every day new [7] Arun Shankar V.K., Umashankar S., Paramasivam S., Norbert
algorithm has been developed to enrich the software and H. “Real time simulation of Variable Speed Parallel Pumping
making the system easier, less complex. Machine learning is a system”, International Conference on Applied Energy,
new phase in the world of artificial intelligence by which it is ICAE2017, Cardiff, UK.
possible to detect any fault in the system. In pumping sector [8] N. R. Sakthivel, V. Sugumaran, and S. Babudevasenapati,
previously manually detection of fault and hardware based “Vibration based fault diagnosis of monoblock centrifugal
system were reported. Now without hampering the system pump using decision tree,” Expert Syst. Appl., vol. 37, no. 6,
procedure, working it is possible to detect faults in the system. pp. 4040–4049, 2010.
So many algorithms are formed like supervised, unsupervised, [9] Arun Shankar, V K, Umashankar S, Shanmugam Paramasivam,
a deep learning method which is more detected and easy to Sanjeevikumar P, “A Comparison of PI, Adaptive PI, and
detect the fault in the system. Fuzzy based speed control of Induction motor for Centrifugal
Pumping System”, The 19th IEEE International Conference on
VII. CONCLUSION Industrial Technology, IEEE-ICIT'18, Lyon (France).
[10] N. R. Sakthivel, V. Sugumaran, and B. B. Nair, “Automatic
This paper is the modification of the hardware based work rule learning using roughset for fuzzy classifier in fault
in machine learning background, where the flow rate with categorization of mono-block centrifugal pump,” Appl. Soft
varying speed. After 2800 RPM the system is showing Comput. J., vol. 12, no. 1, pp. 196–203, 2012.
cavitation as rated speed of single pump is 2800 RPM. In 9 [11] V. K. Arun Shankar, Umashankar S, Sanjeevikumar P,”
different speed data have been checked which is giving proper Comparative Study of Photovoltaic Based Power Converter
pump Performance curve and NPSHr curve. The rated speed is Topologies for Pumping Applications”, IEEE Cencon 2017,
2100 RPM for sample data so it is notified that after 2100 RPM Malaysia.
cavitation occurs. The comparison curve of pump performance, [12] T. Mamo, I. Juran, and I. Shahrour, “Virtual DMA Municipal
which approximately follows the general pump performance Water Supply Pipeline Leak Detection and Classification Using
curve, and the comparison of pump NPSHr curve which also Advance Pattern Recognizer Multi-Class SVM,” J. Pattern
follows the general NPSHr curve. The comparison curve of Recognit. Res., vol. 9, no. 1, pp. 25–42, 2014.
NPSHa and NPSHr is described also. For pump1 rated speed is [13] V. K. Arun Shankar, Umashankar S., Paramasivam S.,
2100 RPM and pump 2 rated speed is 2800 RPM. At last it is “Investigations on Performance Evaluation of VFD fed PMSM
possible to identify the cavitation fault in the system. This using DTC Control Strategies for Pumping Applications”
method has been tested with a real time accurate value which is International Conference on Innovations in Power and
Advanced Computing Technologies (i-PACT), 2017.
showing the efficiency curve and best fitted method. All the
[14] M. M. Stopa, B. J. C. Filho, and C. B. Martinez, “Incipient
data are collected from the centrifugal point pump which has Detection of Cavitation Phenomenon in Centrifugal Pumps,”
0.75 HP power, 2000lph flow rate, 1.7 a current with 3 phase, vol. 50, no. 1, pp. 120–126, 2014.
416 v, 50 Hz AC Induction motor. [15] Arun Shankar V.Ka, Umashankar S., Sanjeevikumar P., Lucian
M.P., Viliam F., Vigna R., “Power Quality Performance
APPENDIX Analysis of grid tied PV fed Parallel Pumping System under
Parameters Parameters Normal and Vibrating Condition”, Applied Energy Symposium
K Loss coefficient L Length of the pump and Forum, REM2017, REM 2017, China.
F Friction coefficient Re Reynolds No [16] B. S. Yang, W. W. Hwang, M. H. Ko, and S. J. Lee,
D Pipe Diameter U Kinetic Viscosity “Cavitation detection of butterfly valve using support vector
A Cross Sectional Area machines,” J. Sound Vib., vol. 287, no. 1–2, pp. 25–43, 2005.
[17] V. K. Arun Shankar, Umashankar S., Paramasivam S., “Energy
REFERENCE Savings using VFDs in a Centrifugal Pumping System”,
[1] T. Ahonen, J. Tamminen, J. Ahola, and J. Kestilä, “Novel International Conference on Power, Circuit and Information
method for detecting cavitation in centrifugal pump with Technologies (ICPCIT), 2017.
frequency converter,” Insight - Non-Destructive Test. Cond. [18] A. Shankar, V. Kalaiselvan, U. Subramaniam, and P.
Monit., vol. 53, no. 8, pp. 439–449, 2011. Shanmugam, “A comprehensive review on energy efficiency
[2] R. Azizi, B. Attaran, A. Hajnayeb, A. Ghanbarzadeh, and M. enhancement initiatives in centrifugal pumping system,” Appl.
Changizian, “Improving accuracy of cavitation severity Energy, vol. 181, pp. 495–513, 2016.
detection in centrifugal pumps using a hybrid feature selection

View publication stats