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

net/publication/277008293

The detection of cavitation in hydraulic machines by use of ultrasonic signal

analysis

Conference Paper  in  IOP Conference Series Earth and Environmental Science · September 2014
DOI: 10.1088/1755-1315/22/5/052005

CITATIONS READS

2 907

5 authors, including:

Peter Gruber
Lucerne University of Applied Sciences and Arts
68 PUBLICATIONS   316 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Peter Gruber on 21 May 2015.

The user has requested enhancement of the downloaded file.

  
 
The detection of cavitation in hydraulic machines by use
of ultrasonic signal analysis.

P. Gruber1,2, P. Odermatt, M. Etterlin, T. Lerch, M. Frei

HSLU T&A, Fluid Mechanics and Hydro Machines, Horw, Switzerland, 2Rittmeyer AG,
Baar, Switzerland
peter.gruber@hslu.ch
M. Farhat
EPFL-LMH, Laboratory for Hydraulic Machines, Lausanne, Switzerland

Abstract. This presentation describes an experimental approach for the detection of

cavitation in hydraulic machines by use of ultrasonic signal analysis. Instead of using
the high frequency pulses (typically 1MHz) only for transit time measurement different
other signal characteristics are extracted from the individual signals and its correlation
function with reference signals in order to gain knowledge of the water conditions. As
the pulse repetition rate is high (typically 100Hz), statistical parameters can be extracted
of the signals. The idea is to find patterns in the parameters by a classifier that can
distinguish between the different water states. This classification scheme has been
applied to different cavitation sections: a sphere in a water flow in circular tube at the
HSLU in Lucerne, a NACA profile in a cavitation tunnel and a Francis model test
turbine both at LMH in Lausanne. From the signal raw data several statistical
parameters in the time and frequency domain as well as from the correlation function
with reference signals have been determined. As classifiers two methods were used:
neural feed forward networks and decision trees. For both classification methods
realizations with lowest complexity as possible are of special interest. It is shown that
three signal characteristics, two from the signal itself and one from the correlation
function are in many cases sufficient for the detection capability. The final goal is to
combine these results with operating point, vibration, acoustic emission and dynamic
pressure information such that a distinction between dangerous and not dangerous
cavitation is possible.

1. Introduction
Cavitation is generated if the static pressure in a fluid falls beneath the evaporating pressure of
the fluid under constant water temperature. Cavitation bubbles emerge around germs in the
fluid. These germs weaken locally the adhesive forces in the fluid which makes cavitation
easier. The germs might be impurities, trapped gases or other tiny cavities. The necessary
reduction of the static pressure is due to local pressure fluctuations or to an increase of the fluid
velocity.
The cavitation bubbles which implode near the surfaces of the mechanical components of the
machine, generate a micro-jet. This leads to high pressure and velocity peaks, which cause an
abrasive and damaging effect on the components. Big cavitation bubbles or vapour regions do
not generate an abrasive effect, they disturb however the flow field, can cause flow separation
and therefore reduce the efficiency of the machine. Both effects are unwanted. In a hydraulic
machine cavitation might occur at a variety of locations. Figure 1 lists different types of
cavitation in a hydraulic machine, Avellan [1] describes the fact that different dangerous and not
dangerous cavitation cannot be distinguished from operating point data as it is shown in the hill
chart of Figure 2. Classical monitoring methods based on frequency content of pressure, noise
and vibrational signals are described in Escaler et al. [2].
1
To whom any correspondence should be addressed.
 Figure 1. Different type of cavitation

Figure 2. Hill chart for Francis turbine: regions of different types of cavitation
1: leading edge cavitation suction side, 2: leading edge cavitation pressure side
3: interblade cavitation, 4: ring swirl cavitation

Here a new approach is applied. The deterioration the ultrasonic signals due to the various
cavitation effects is exploited in a statistical way.

2. Measurements
Measurements were carried out at three objects by mounting the acoustic sensors in a clamp-on
fashion from the outside of the fluidized section of the installations (Müller [3], Gruber et al.
[4]):
- sphere in a vertical pipe of perspex at the hydraulic laboratory at the HSLU (Figure 3)
- different profiles in the cavitation channel of the EPFL-LMH laboratory (Figure 4)
- Francis model test turbine at the test rig at EPFL-LMH (Figure 5)

Figures 3 – 5 show the three installations tested and, Figures 6 - 8 the corresponding
acoustic path locations.
 Figure 3. Sphere Figure 4. Profile in cavitation tunnel Figure 5. Francis model test
turbine

       Figure 6. Acoustic path Figure 7. Acoustic path Figure 8. Acoustic path

sphere profile Francis turbine

The operating points of the three objects were chosen such that the different types of cavitation
belonging to each operating point was visible or was known to the test engineers. Therefore the
data could later be used for training the classifiers.

3. Extraction of statistical signal parameters (characteristics)

3.1 Statistical signal parameters for raw signal
For each operating points in each test section hundreds of signals (1MHz pulses) have been
recorded. If they are disturbed by not ideal water conditions, various quantities of the signal
change: amplitude (signal maximum), time of arrival of signal maximum, frequency content and
power. As the distortions do not occur continuously in time, the different signals recorded
sequentially experience different levels of distortions for the different water states. By that
histograms of the above quantities can be plotted. Figure 9 shows signal examples and Figure
10 histograms of groups of 100 signals. The signals are all preprocessed to remove outliers and
signals with very low signal level.

Signalanimation Signalanimation
0.8 0.15
Signal=9

0.6
0.1

0.4
0.05

0.2

0
A m p litu d e [V ]
A m p litu d e [ V ]

-0.05

-0.2

-0.1
-0.4

-0.15
-0.6

-0.8 -0.2
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
Zeit [s] Zeit [s]

Figure 9. Samples of undisturbed signal (left) and disturbed signal

Figure 10. Histogram of 100 measurements of amplitude: left undisturbed, right disturbed (draft
tube swirl)

3.2 Statistical signal parameters for correlation function

An interesting function to examine is the correlation of each incoming signal with a reference
signal that corresponds to a signal obtained in undisturbed conditions. In contrast to the signal
analysis of the previous section, the characteristics of two signals are used and compared to one
another if correlation is applied.

Figure 11: samples of correlation functions: upper left undisturbed signal, other signals:
different degree of disturbance

3.3 Selection of important statistical parameters

The selection of the most important statistical parameters for the neural network classifiers was
done in the following way:
- To begin with simple combinations and then increase complexity if needed
- For classifiers with only two inputs the selection has been done manually by looking to
simple conditions including thresholds by which most of the training data can be
 separated. Physical considerations were also used to support the selection process. That
means not all combinations have been considered. The selected combinations have then
been used for training the neural net. If several solutions were found, then no procedure
was followed to make out the winner. This approach does not guarantee that a solution
can be found even if one exists.
- For classifiers with three inputs the selection process gets more complex. For this study
the selection has been done similar to the first case. The starting point however were
combinations of two inputs which were partly successful and then adding another
candidate manually. This search was again not systematically and exhaustive. If the
training was successful, then maybe two or three alternatives have been tried out.
In case of a decision tree classifier, the inputs were chosen as for the neural network or they
were selected by an automated search.

4. Classification with neural networks (Hassoun [5])

In order to find a neural net the following steps were followed:
1) Data collection and preprocessing: The collected data have to be filtered and then
normalized (typically between [-1 1]). Generally it can be said, that the more data are
available the better the net can be trained. Basically the net can only classify what it was
trained for.
2) Choice of network structure: The structure including inputs and outputs has to be
selected with the knowledge acquired from the recorded data, the physics behind the
application and from other constraints. With the structure the number of parameters to
be trained is given.
3) Net configuration: The configuration determines the type of learning algorithm
(nonlinear optimization problem), the maximal number of iteration and the stopping
error criteria.
4) Initialisation of weights and bias: As the optimization problem is nonlinear the solution
is dependent on the initial conditions. The choice can be made between random initial
conditions and fixed initial conditions.
5) Training: The training data are split into three groups which can be chosen in % of the
whole set: the proper training data used for finding the net, validation data where
undertraining or overtraining can be checked (the net is still optimized with these data)
and test data for the final net. The optimization algorithm can also be chosen. Typically
the Marquardt-Levenberg algorithm (Press et al. [6]) is used for minimizing the mean
squared error function:
N N

  ( y  f ( p ))
1 1
MSE  ei2  i
2
N N i
i 1 i 1

y i  target or neural net output

f ( p i )  nonlinear mapping of neural net from inputs p to output y

6) Test: The net is tested as classifier. No more training is performed.

7) Application: if the test is successful, the net can be applied, otherwise one has to go
back to point 1) or 2).  
 
The Matlab neural network toolbox was used for the training. The neural network approach has
been tested at all three cavitation objects. In all situations the distinction between different water
states was possible. In the following only one example of the Francis turbine is given.
4.1 Example: Francis turbine (Test rig, LMH Lausanne), [7], [8], [9] and [10]
This is the most complex example with a neural net with three inputs, one hidden layer with
three nodes and four outputs, one for pure water, draft tube swirl cavitation, interblade vortex
cavitation and leading edge cavitation respetively. The first two inputs are the mean of the
amplitude and the coefficient of variation of the amplitude of the recorded signals, while the
third is a characteristic from the correlation function. Two choices of characteristics of the
correlation function which were successful are the ratio of mean of areas of right to left and the
ratio of the mean of centre of area gravity of right and left.

 
Figure 12. Definition of some parameters of the correlation function

Both are measures of the asymmetry of the correlation function. With each of the third input it
was possible to distinguish all four states. Table 1 gives the details of the experiment, while
Figure 13 shows the chosen network structure.

Francis Turbine (Test Rig, LMH Lausanne)

Experimental Setup
sensor mode direct path downstream, Clamp-On 500kHz
recorded data 5-10 x 100 samples for each state
Water state detection
classifier feed forward neural network
architecture 3 input neurons, 3 hidden neurons, 4 output neurons
number of weights/biases 30/10
input parameters i) the mean of the amplitude (sample size 100) Amean
ii) the coefficient of variation of the amplitude A,rel
iiia) ratio of mean of areas of right and left of correlation function
iiib) ratio of the mean of centres of area gravity of right and left of
correlation function
Correctly detected water (100%): W
draft tube swirl (100%): DTS
inter blade vortex cavitation (100%): IBVC
leading edge cavitation (100%): LEC
Table 1: Data for the Francis turbine experiments, chosen neural network structure and
classification results
  
Figure 13. Neural network structure for the Francis turbine experiments

5. Classification with decision trees (Breiman et al. [11], [12], Hand et al. [13])
A binary decision tree is a sequence of conditions factored into a tree-structured series of
branches as shown in Figure 14. Each node consists of a condition for one attribute or input
parameter of the data set. The result of the check of the condition is either Yes or No depending
on whether the attribute is above or under a certain threshold  . Each node leads therefore to
two outgoing branches. To know the order in which the attributes must be chosen to split the
data in two classes, a measure is needed that allows to compare the effect of the attributes and
choose then one above the other. One of these measures is called impurity and could be defined
as the amount of uncertainty present in the data. The input parameter which reduces the
impurity together with a threshold value for the condition in the node must be found by
optimization. Given the probability p of an input parameter data set belonging to one target
value, and 1-p the probability for the same data set belonging to the other target value, a
common impurity function used in applications is the Gini index criteria

J ( p )  p (1  p )

The input parameter for which J is minimized is then chosen as the node parameter with a
corresponding threshold  . The optimization of the Gini index tries to split the input parameter
data such that all the data of one input parameter for which the probability p is maximal is sent
to one of the branches while all the others are sent to the other branch. The minimization of the
Gini index leads therefore to pure nodes as best as possible.

5.1 Example: Francis turbine (Test rig, LMH Lausanne [14])

The same data set used before for the neural network classification is now used for the
cavitation classification with a decision tree. By doing this the performance of the two different
classification approaches can be compared. As the neural network used three input parameters
for a successful classification, it was tried to find a decision tree classifier with the same
dimensionality of the input parameters. The automated training algorithm was fed with a total of
34 parameters as input. Out of these 34 parameters the following three inputs are the minimum
number in order to classify all training data correctly. The first are exactly the same as for the
neural net: the mean of the amplitude Amean, and the coefficient of variation of the amplitude
A,rel. The third input parameter is related to the correlation function, it differs from the neural
net input, it is the standard deviation t_2pk_xcorr of the time shift of first (in amplitude) left side
peak from the centre of the correlation function as shown in Figure 12. The correlation function
parameter inputs: ratio of mean of areas of right and left of correlation function or the ratio of
the mean of the centres of area of gravity of right and left of the correlation function which were
chosen for the neural net was here not selected by the training algorithm. As training method the
classregtree algorithm from the Matlab statistical toolbox was used with the optimization
algorithm for the Gini index criteria.
The resulting classifier is summarized in Figure 14. The tree is lean and has only four nodes.
Interestingly enough the input parameter Amean was chosen twice as node parameter.

target parameter threshold

1 Amean 0.0350753 [V]

2 t_2pk_xcorr 5.1932 e-6 [s]
3 Amean 0.000746629 [V]
4 A,rel 55.3677 [%]
Figure 14: Used input parameters and thresholds per node and the simple decision tree structure
for the Francis turbine cavitation classification by a decision tree. The undefined state consists
of data, which were not associated with one of the four water states (e.g. noise only).

With the decision tree training data set, the tree structure and
each branch threshold is determined. Table 2 shows the
validation matrix, for which a much larger data set including
training data set is classified by the decision tree shown in
Figure 14.
The cavitation detection is correct except for one state which
is out of the diagonal where inter blade vortex cavitation is
classified as a draft tube swirl. This leads to a success rate of
98% compared to 100% with the neural network presented
above for the same data set. However, compared to the neural
network, the architecture of the decision tree is much simpler
Table 2: Validation matrix,
and more transparent. It can be easily implemented by simple
target versus classification
rules.

6. Comparison and conclusions

In order to compare the two classifier approaches in a fair way, the same procedure should be
used. In the case, this was not accomplished fully. For the neural net classifier no automated
selection procedure of the input parameters was applied compared to the decision trees
approach. In spite of this, some important conclusions can be drawn from the results of all the
experiments:

6.1 Neural net

 The neural feed forward net leads to a nonlinear mapping between inputs and target with
the typical quite general interconnected structure. This enhances the chances to find a
classifier which is able to distinguish the classes with high reliability. The solution
found shows a complicated structure and might therefore have small robustness if
applied to new data.
  The structure of the neural net must be specified before the training algorithm is applied.
Here a pragmatic approach has been used and the standard neural feed forward net was
selected. Physical rules of thumbs were only used for the number and the selection of
the input parameters. If however, more physical insight to the observed parameters of
the process to be classified is possible, then this knowledge should be used to define a
more sophisticated structure of the net. That implies the number and choice of input
parameters, the number of layers and the number of nodes per layer of the net. This
would also allow to build a neural net that could consist of several smaller neural nets in
a cascade.

 The number of weights and thresholds to be optimized gets soon pretty large with no
clear connections to the underlying physics. This makes it hard to interpret the resulting
function of the classifier.

 The high number of weights and thresholds leads to a nonlinear optimization problem of
high dimensionality. The choice of initial conditions for the optimization is therefore a
delicate issue because the found solution depends on them.

6.2 Decision tree

 In a decision tree approach only the binary node type but not the detailed tree structure is
given. The training algorithm can choose the number and the selection of input
parameters automatically and the detailed structure of the tree is part of the solution.

 The found decision tree is understandable, easy readable and implementable. The tree
creates after each branch a subspace, whereas a feed forward neural net tries to classify
all data in one space with the dimension of the numbers of different parameters. A
graphical interpretation is therefore easy.

 An input parameter can be used several times which is not possible with a single feed
forward neural network.

 From a decision tree we learn directly which input parameter can split best one class from
the other classes. Also the threshold for each parameter condition has a physical
meaning.

6.3 Conclusions
Both classifier methods are applicable and led to the required distinction of the different water
states. For all the experiments at least one classifier could be found with both methods. The
solutions however are not unique and are driven by the training data. The decision tree
approach is easier interpretable and therefore its acceptance is higher. The method provides the
user automatically with the most important input parameters. From this point of view, the
decision tree method could also be used for the choice of input parameters for a neural network
approach. A general automated search for the best inputs and structure of a neural net is huge
and could not be carried out. The most important input parameter found have a physical
meaning:

1. The attenuation observed in the mean of the signal amplitudes can be explained by the
concentration and size of particles or bubbles in the water.

2. The coefficient of variation of the signal amplitude and therefore also the standard
deviation can be interpreted by a concentration of cavitation or air bubbles. If there is
 only a small amount of bubbles, the standard deviation is small and increases with a
higher bubble concentration. If there are only a few bubbles, most of the sent ultrasound
signals are unaffected. So for the most part of the recorded signals, there is no
difference to a signal sent through clear water – consequently the few disturbed signals
by bubbles have only a small impact on the standard deviation. Interesting is the fact
that air filled bubbles influence the standard deviation much more than vapor filled
bubbles – this helps to distinct air bubbles from cavitation bubbles in the water.

3. On the other hand, the physical impact on the cross correlated signals is not so clear, but
with large measurement series and well defined boundary conditions, relations from
water states to signal interaction can be extracted experimentally. The important cross
correlation parameters describe the distortion of the shape of the signal.

Both classification methods will be applied in a next step to field experiments. An open issue is
the question of how generic a classifier can be specified and how much of the classifier has to
be trained on site. A special focus will also be given to the combination of the ultrasonic
classifier in combination with operating point information and analysis of vibrational,
hydrophone and noise signals. By merging the different sources of information by rule based or
decision tree methods the chance to be able to separate dangerous from not dangerous cavitation
will increase.

References
[1] Avellan F 2004 Introduction to cavitation in hydraulic machinery (The 6th International
Conference on Hydraulic Machinery and Hydrodynamics Timisoara, Romania)
[2] Escaler X, Egusquiza E, Farhat M, Avellan F, Coussirat M 2006 Detecton of cavitation in
hydraulic turbines (Mechanical Systems and Signal Processing 20) p.983-1007
[3] Müller C 2008 Untersuchung der Kavitation mit Ultraschall an zwei Prüfstrecken
(Bachelor Diplomarbeit HSLU Luzern)
[4] Gruber P, Roos D, Müller C, Staubli T 2011 Detection of damaging cavitation states by
means of ultrasonic signal parameter patterns (WIMRC 3rd International Cavitation
Forum, July 2011, Warwick, England)
[5] Hassoun M H 1995 Fundamentals of artificial neural networks (MIT Press)
[6] Press W, Teukolski S A, Vetterling W T, Flannery B P 1986 Numerical Recipes
(Cambridge University Press)
[7] Etterlin M 2012 Klassifizierung von Wasserzuständen mithilfe von Ultraschallsignalen
und neuronalen Netzen (Industrieprojekt, HSLU Luzern)
[8] Lerch T 2013 Klassifizierung von Kavitationszuständen mithilfe von Ultraschallsignalen
(Industrieprojekt, HSLU Luzern)
[9] Gruber P, Odermatt P, Etterlin M, Lerch T, Farhat M 2013 Cavitation Detection via
Ultrasonic Signal Characteristics (IAHR, 5th International Workshop on Cavitation
and Dynamic Problems in Hydraulic Machinery, Lausanne, Switzerland)
[10] Gruber P, Odermatt P, Etterlin M, Lerch T 2013 The detection of cavitation in hydraulic
machines by use of ultrasonic signal analysis (CTI-Report)
[11] Breiman L, Friedman J H, Olshen R A, Stone C J 1984 Classification and regression
trees (Chapman and Hall)
[12] Breiman L 1996 Technical Note: Some Properties of Splitting Criteria (Machine
Learning) p. 24, 41-47
[13] Hand D, Mannila H, Smyth P 2001: Data Mining (MIT Press)
[14] Frei M 2013 Klassifizierung von Kavitationszuständen mithilfe von Ultraschallsignalen
und regelbasierten Methoden (Industrieprojekt, HSLU Luzern)

View publication stats