Professional Documents
Culture Documents
Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust
a r t i c l e i n f o a b s t r a c t
Article history: When a liquid drop impacts on a liquid surface, bubbles can be entrained in the liquid by several hydro-
Received 19 July 2021 dynamic mechanisms, generating an acoustic signal in the liquid and in the air. Each of the five mecha-
Received in revised form 12 April 2022 nisms reported displays a typical acoustic signature, both in the liquid and in the air. Even if the
Accepted 7 June 2022
differences between these signatures are subtle, it is possible to distinguish them. A detailed study of
a labeled database made of acoustic signals from different bubble mechanisms allows to extract relevant
descriptors of the signatures. A classification method is then proposed, using a KNN algorithm and shows
Keywords:
great performances to classify the acoustic signals. Thus, the method can be used to study the bubble
Drop impact
Bubble entrainment
entrainment hydrodyanmics during drop impacts, only from the acoustic signals and thus without the
Classification help of a high-speed camera.
Acoustics Ó 2022 Elsevier Ltd. All rights reserved.
Hydrodynamics
Signal processing
1. Introduction ciated with a typical bubble entrainment mechanism, like the reg-
ular entrainment and its pinch off mechanism [11,12,10,13–18], the
Fast hydrodynamic events can generate a sound, like bubble Mesler entrainment and its air film entrapment [19–21] or even
bursting [1] or liquid boiling [2]. This sound can carry important the irregular entrainment regime and its bubble entrapment by a
information about the hydrodynamic events, and thus can be used secondary droplet [22,23], caused by a zip-like flow mechanism
to better understand their unfolding [3]. Indeed, even if the devel- described in detail in [24]. The identification of the bubble entrain-
opment of high-speed cameras allows today to observe in detail ment mechanism involved during a drop impact can thus lead
fast mechanisms, the implementation of this type of measure is directly to the impact initial conditions when these are not easily
costly, not always possible and the generated data are heavy. On accessible, like in the case of rain studies [12]. Even if acoustic
the contrary, acoustic measures are easy to set up and the gener- methods for bubble detection already exist [25,26], they are not
ated data are easy to store. Moreover, the use of machine learning able to identify the mechanism of the bubble formation.
algorithms allows great performances for the classification of pas- Recently, it was shown that for the irregular entrainment regime,
sive acoustic emissions [4,5]. The detection and recognition of the addition of surfactant to the water bath leads to different bub-
hydrodynamic events thanks to their emitted acoustic signature ble entrainment mechanisms during the Rayleigh jet collapse [27].
can therefore be useful for environmental [6], geological [7] or Some of these mechanisms are similar to those observed usually
even structure damage studies [8]. for different impact regimes. More specifically, five mechanisms
A liquid drop impacting a liquid surface is a fast hydrodynamic are reported, including 4 different bubble entrainment mecha-
event that produces a recognizable sound. This sound is mostly due nisms and one bubble excitation mechanism. Each mechanism
to bubble vibrations after their entrapment in the liquid [9]. This exhibits a typical acoustic signature. However, these acoustic sig-
entrapment can happen at different times after the impact and in natures can not be classified automatically by using simple signal
different ways, depending on the drop size and its height of fall features. Thereafter, we call family the set of signals produced by
[10]. The different highlighted impact regimes are generally asso- the same mechanism.
In this paper, we propose an in-depth study of the acoustic sig-
⇑ Corresponding author. natures of the different hydrodynamic mechanisms, reported in
E-mail address: laurent.simon@univ-lemans.fr (L. Simon). [27], induced by a drop impact on a liquid pool of various solutions,
https://doi.org/10.1016/j.apacoust.2022.108868
0003-682X/Ó 2022 Elsevier Ltd. All rights reserved.
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
for the impact regime with bubble entrapment by a secondary dro- The acoustic and video acquisitions are triggered when the pas-
plet described in [24]. This first step will allow to develop an algo- sage of the drop is detected by an XCSource TE174 (f) infrared sen-
rithm capable of identifying the hydrodynamic mechanism sor, situated just under the hose and connected to an Arduino
involved from the acoustic signal alone. In this regard, a more Mega card. The origin of time is set to the instant of initial impact
detailed analysis of the common features of the acoustic signal of the drop falling on the liquid surface.
inside the same family is presented. Signal processing tools are The acoustic signal is recorded underwater with a Brüel&Kjaer
then used to extract these typical features from any acoustic signal, hydrophone (Type 8103) (g), connected to a conditioning amplifier
in order to obtain a relevant learning base for a classification algo- Brüel&Kjaer (Type 2692-0S4). The hydrophone is positioned about
rithm. This learning base is then tested with the K-nearest neigh- 50 mm from the point of impact.
bors (KNN) classification method. The encouraging results The acoustic signal is recorded in the air with a G.R.A.S. Type
obtained are finally analyzed to highlight the possible improve- 40BP 1/4 in. microphone (h), connected to a conditioning G.R.A.S.
ment of the algorithm. Type 26AC 1/4 in. preamplifier, linked to a conditioning amplifier
Brüel&Kjaer Type 2690. The microphone is positioned about
80 mm just above the point of impact.
All signals are recorded with a Picoscope 4262 digital oscillo-
2. Materials and methods
scope with a sampling rate of 500 kHz.
A video of the interface is recorded by a high-speed camera Pho-
The measurements used for the algorithm training base come
tron FastCam SAX2 (i) equipped with a Sigma Macro 105 mm lens
from configurations with different kinds of solutions. However,
(j). A 50 W RS Pro LED lamp is used to backlight the experiment (k).
the measurement set-up shown in Fig. 1 is the same for any solu-
The video acquisition is made at a sample rate of 20 000 fps.
tion type.
The all set-up is supported by a metallic structure (l) located in
A drop is generated by a PhD Ultra syringe pump (a) equipped
a semi-anechoic room with a cut-off frequency of 100 Hz. The tem-
with a 60 mL syringe (b) connected to an Interchim Teflon cylindri-
perature is quite constant around 19 C.
cal hose (c). The interior and exterior diameters of the hose are
Each signal is denoised with a spectral subtraction algorithm to
1 mm and 1.5 mm respectively. The flow rate is set at 10 mL/h.
reduce the background noise [28,29].
The drop is made of distilled water or the same solution as in the
The different acoustic events are detected with a Python algo-
tank. So the drop diameter is between 2.60 and 3.70 0.05 mm
rithm based on a local energy maxima detection. It uses a sliding
because of the surface tension of the different solutions. The height
window along the signal to calculate the local energy. The local
of the drop released is fixed to 740 mm from the liquid surface, so
maxima of energy are then detected with the find peaks function
the measured speed of the impact, deduced by the footage, is
of Scipy [30]. These maxima correspond to the initial impact and
between 3.32 and 3.40 0.05 m/s.
the beginning of each acoustic event. The acoustic events are then
The drop falls into a Plexiglas tank (d) with external dimensions
extracted with a window that starts 2 ms before and ends 10 ms
of 160 160 110 mm3, with 10 mm thick walls. It is filled until
after the detected starting time. Thus, each considered signal of
its maximum height. The tank rests on four blocks of solid polymer
acoustic event has the same length. Since no relationship is
foam (e) to reduce the transmission of external vibrations to the
observed between the acoustic events on the one hand and the liq-
tank.
uid properties on the other hand, the events extracted from all con-
The tank is filled with different aqueous solutions of sodium
figurations are gathered in a single database.
dodecyl sulfate (SDS), polyoxyethylene lauryl ether (Brij-35),
cetyltrimethylammonium bromide (CTAB), ethanol and glycerol,
all with various concentrations. For each concentration, measures
3. Acoustic signals of the different classes
are made with a drop of distilled water and a drop of bath fluid.
Fig. 2. Succession of the main hydrodynamic events after a water drop impact on a surface of an aqueous solution of SDS at 240 mg/L. Considering t ¼ 0 the time of the initial
impact, the images are taken at times t ¼ 1:3 ms (a), t ¼ 6:9 ms (b), t ¼ 80:4 ms (c), t ¼ 135:4 ms (d), t ¼ 141:6 ms (e), t ¼ 158:5 ms (f), t ¼ 168:3 ms (g) and t ¼ 212:4 ms
(h). SD1, SD2 and SD3 are respectively the first, second and third secondary droplets to be released from the jet.
measurements with various solutions, five different families are 4. Signal descriptors
identified [27]. They correspond to five different types of hydrody-
namic events involving the vibration of one or several bubbles, The selection of relevant descriptors to discriminate the differ-
shown in Fig. 3, and typical acoustic signals in air and in water ent classes is an essential part of the signal classification. Here, we
are given in Fig. 4, with the corresponding power spectra shown call class the output of the classification algorithm that may or not
in Fig. 5. As a vibrating bubble emits a damped sine at a frequency correspond to a so-called family. The main difficulty for the signals
depending on its size [32], the acoustic signatures are quite alike. considered in this study is the great variability of basic signal fea-
However, some differences are noticeable and allow their identifi- tures inside a same family. Indeed, as the bubble size and the con-
cation. The hydrodynamics of the different mechanisms, as well as dition of the bubble entrapments are various even for a same
their acoustic signature, are described in detail in [27]. A brief mechanism, the corresponding damped sines display various fre-
description of the different families is given below. quencies, amplitudes and damping times. It is thus not possible
to discriminate the signals only based on these basic features.
Mesler-like entrainment (MLE): a film of air is entrapped Another difficulty for the signals considered in this study is
between a liquid volume descending from the base of the Ray- their similarity. Indeed, as all the signals correspond to bubble
leigh jet and the rest of the bath (Fig. 3-a), breaking then into a vibration, they are made of one or several damped sines with var-
multitude of small bubbles. Therefore, when it is present, it is ious frequencies. Their differences are quite subtle, but discernible
always the first acoustic event visible on the signal. The hydro- by sight.
phone signal is made of several overlaid damped sines with A set of relevant descriptors is proposed. Their identification
high frequencies and low amplitudes, as shown in Fig. 5-a. Only comes from the observation of the temporal shape of the different
a very low amplitude signal (or no signal at all) is visible on the signals, as well as the hydrodynamic behaviours of both the bubble
microphone signal, as shown in Fig. 4-a. and the liquid surface during the acoustic emission. Indeed, the dif-
Zip-like flow (ZLF): a secondary droplet falls to an air cavity ferent hydrodynamic events responsible for the bubble excitation,
(Fig. 3-b), detaching a bubble in a zip-like flow mechanism. specific to each family and easily observable on the video, can lead
The resulting signal is a single damped sine both for the micro- to different signatures visible on the acoustic signals. For example,
phone and the hydrophone, as shown in Figs. 4-b and 5-b. In the liquid flows, mostly induced by the dynamics of the cavity just
addition, the microphone signal displays a particular pattern, above the bubble, lead to variations of the local hydrodynamic
made of a pressure decrease followed by a sharp rise, just before pressure, which are recorded by the hydrophone as low frequency
the beginning of the oscillations. This pattern is typical of the variations of the local mean pressure. The cavity dynamics being
ZLF airborne signature, as described in a previous study [24]. typical to the bubble excitation mechanism, i.e. to the considered
Bubble pinch off (BPO): a small volume of air detaches from family, these variations carry essential information for the signal
the bottom of a cavity by a pinch off mechanism (Fig. 3-c). classification.
The resulting signal is a single damped sine both for the micro- To calculate the descriptors needed for the classification, differ-
phone and the hydrophone, as shown in Figs. 4-c and 5-c. ent processes are applied to the signals. Thus, the hydrophone sig-
Donut splash (DS): a secondary droplet falls to a retracting cav- nals are decomposed into both an acoustic and an hydrodynamic
ity (Fig. 3-d), detaching a toroide shaped air volume, which then contributions, by filtering the low frequency component, corre-
breaks into several smaller bubbles. The corresponding acoustic sponding to variations of the local pressure due to the different liq-
signal is generally made of several overlaid damped sines, with uid flows, i.e. the hydrodynamic contribution. The low-pass filter
different frequencies, corresponding to the emission of the dif- used to extract the hydrodynamic pressure part is a FIR filter with
ferent bubbles (Fig. 5-d). A local rise of the mean pressure can a cutoff frequency of 2 kHz. The acoustic part is extracted by a FIR
also be observed on the hydrophone signal, due to the cavity band-pass filter, with cutoff frequencies of 2 and 80 kHz, as no
retraction, as shown in Fig. 4-d. damped sine with a frequency outside of this band is observed in
Touched bubble (TB): a bubble at rest in the liquid is reached our measurements. Thus, the hydrodynamic pressure variations
by a cavity (Fig. 3-e). As both begin to merge, the cavity rapidly are removed, as well as high-frequency noise, leaving only the
retracts, producing the bubble vibration. The hydrophone signal acoustic signal generated by the bubble. An example of an acoustic
shows a single damped sine (Fig. 5-e), generally with a rise of signature, decomposed into the hydrodynamic and the acoustic
the mean pressure due to the cavity retraction, as shown in contribution, is shown in Fig. 6. In the following, it is specified from
Fig. 4-e. The microphone signal generally does not show any which signal each descriptor is extracted (low-pass, band-pass or,
event. if no indication, the raw signal).
3
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
ðhÞ
The main frequency F main of the hydrophone signal is defined as
the frequency corresponding to the maximum of P xðhÞ ðFÞ, the power
BP
ðhÞ
spectrum of xBP ½n, defined as
!2
X
N1
T s xBP ½nej2pFnT s
ðhÞ
PxðhÞ ðFÞ ¼ ; ð1Þ
BP
n¼0
ðmÞ
The microphone signal standard deviation xstd is defined as:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N1
u1 X
¼t
ðmÞ
ðxðmÞ ½n xðmÞ Þ ;
2
xstd ð2Þ
N n¼0
with xðmÞ ½n the microphone signal and xðmÞ its mean.
This descriptor is linked to the signal amplitude and its shape.
The microphone signal amplitude depends on the bubble ampli-
tude of the signal in water, as well as on its frequency and the dis-
tance of the bubble from the interface [24]. Thus, this descriptor is
expected to display high values for ZLF, medium values for BPO and
DS, and low values for MLE and BT.
4
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
0.02
0 0 0
-0.02
10 5 2 c)
0 0 0
-10 -5 -2
0.1 0.4 0.1
0 0 0
2
-0.4
8 d)
4 4
0 0 0
P xðhÞ ðFÞ are detected in the same way as for the microphone signal, N1
X 2
ðhÞ
with the same parameters. ExðhÞ ¼ T s xBP ½n ; ð4Þ
BP
n¼0
As multiple bubbles are created for the MLE and the DS families,
a large number of peaks is expected to appear on the power spec-
X
N1
ðhÞ 2
ExðhÞ ¼ T s x ½n ; ð5Þ
trum. However, the number of peaks for MLE is expected to be n¼0
much higher than for the microphone signal, as a lot of small bub-
bles are generated deep under the surface so their acoustic emis- with N the number of points of the signal and T s the sampling
sion is not measured in the air. The other families are expected period.
to show a low number of peaks as only one bubble is involved. This descriptor is expected to display values close to one for ZLF
and BPO families, as the bubbles created in these cases emit an
4.5. Part of the acoustic energy in the total hydrophone signal high amplitude signal and the variations of hydrodynamic pressure
are quite low. It is however expected to be quite low for the other
The energy of the different hydrophone signals may be calcu- families.
lated. We respectively note ExðhÞ and ExðhÞ the energies of the raw
BP
hydrophone signal and the band-pass filtered one. We then define 4.6. Filtered hydrophone signal amplitude ratio
RE the part of the acoustic energy in the whole energy of the hydro-
phone signal, as: The filtered hydrophone signal amplitude ratio FsarðhÞ is defined
as the ratio of the maximum of amplitude of the low-pass filtered
ExðhÞ ðhÞ
RE ¼ BP
; ð3Þ hydrophone signal xLP ½n over the maximum of amplitude of the
ExðhÞ ðhÞ
band-pass filtered hydrophone signal xBP ½n:
5
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
10 -10 10 -10
10 -15
-15 10 -20
10 10 -20 a)
10 -10 -10 10 -10
10
10 -15
10 -10 10 -10
10 -10
10 -15 10 -15 b)
)
2
10 -5
10 -5 -10
10
10 -10
Power spectrum ( (Pa.s)
10 -10
10 -10 10 -10 10 -10
ðhÞ
maxðxLP ½nÞ xðmÞ ½n:
FsarðhÞ ¼ ðhÞ
: ð7Þ
maxðxBP ½nÞ
minðxðmÞ ½nÞ
MMSRðmÞ ¼ j j: ð8Þ
This descriptor is mainly linked to the dynamics of the cavity. maxðxðmÞ ½nÞ
Indeed, each of the excitation mechanisms of a bubble presented
As the microphone signal for all families is close to a damped
in Section 3 involves a cavity. This cavity dynamics is responsible
sine, with no variations of the mean pressure, the value of this ratio
for variations of the hydrodynamic pressure [27]. As this cavity
is expected to be close to one, except for the ZLF family. Indeed, for
develops, the hydrodynamic pressure decreases. Then, when the
this family, the microphone signal displays a particular pattern,
cavity retracts, the pressure increases to a maximum. The maximum
just before the beginning of the oscillations. This pattern is shown
of pressure can display an high amplitude, especially for a high
in [24] and is typical of the ZLF family. It is made of a decrease in
retraction speed. Thus, this descriptor compares the contribution
pressure, followed immediately by a sharp and high peak of pres-
of the hydrodynamic pressure to the contribution of the acoustic
sure. The particularity of this pattern is its asymmetry. The pres-
pressure in the signal. This ratio is expected to be low for ZLF and
sure peak is systematically of higher amplitude than the
BPO, as the oscillations are generally of high amplitude and almost
minimum of pressure. It is also higher than the following damped
no variation of hydrodynamic pressure is observed, and rather high
sine amplitude. Thus, the min–max microphone signal ratio for the
for DS and TB.
ZLF family is expected to be lower than one.
4.7. Min–max microphone signal ratio 4.8. Mean time of the band-pass filtered hydrophone signal
This descriptor is defined as the ratio of the minimum of pres- This descriptor can be considered as the center of gravity in
ðhÞ
sure over the maximum of pressure of the microphone signal time of xBP ½n. It is calculated as:
6
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
6
4
2
0
4
Pressure (Pa)
-2
5
-5
0 2 4 6 8 10 12
Time (ms)
Fig. 6. Typical hydrophonic signal of a DS event (top), with the hydrodynamic contribution extracted with a low-pass filter (middle) and the acoustic contribution extracted
with a band-pass filter (bottom).
base. It is chosen to keep 80% of the original database for the learn- The number of peaks both on the microphone and on the hydro-
ing base and 20% for the test base. Before separation, the data are phone power spectra is directly linked to the number of generated
shuffled to have signals from all the different configurations in bubbles, but also to the energy of these bubbles. Indeed, as the
both bases. It is also verified that all the classes are represented peak detection threshold depends on the amplitude of the maxi-
with a same rate in both bases. mum of the power spectrum, an high maximum prevents the
detection of other peaks (which can be due to other bubbles, noise
or resonances). As shown in Fig. 7-b. the number of peaks on the
6. Results and discussion hydrophone power spectrum allows to separate MLE and DS fam-
ilies, as the number of peaks on the microphone power spectrum
6.1. Analysis of the descriptors calculated for the learning database rather allows to separate ZLF and DS families.
The part of the acoustic energy in the total hydrophone signal
In a first time, the different descriptors are calculated for every and the filtered hydrophone signal amplitude ratio are both related
signals of the learning database, to verify the expected tendencies to the hydrodynamic contribution in the hydrophone signal. The
reported in Section 4. The results are presented in Fig. 7. The results displayed in Fig. 7-c show rather clustered values, with
descriptors are represented by pairs related to a similar feature. no distinct tendency. However, as expected, both descriptors allow
The main frequency of the hydrophone signal and the micro- to separate ZLF and DS families quite well.
phone signal standard deviation are descriptors mainly linked to Finally, the min–max microphone signal ratio and the mean
the bubble features. As shown in Fig. 7-a. both descriptors show time of the band-pass filtered hydrophone signal are both related
tendencies in agreement with the expectations reported in Sec- to the temporal shape of the signal. As expected and shown in
tion 4. However, the two descriptors are not sufficient to separate Fig. 7-d, the min–max microphone signal ratio allows to separate
all the families. Even if the ZLF and MLE families are quite distinct quite well the ZLF family from the other families. However, the
from each other, the BPO and the DS families display however var- mean time of the band-pass filtered hydrophone signal displays
ious values for these descriptors, as the bubbles involved are of more scattered values than expected for all families.
various sizes and positions under the interface.
a) b)
Number of peaks - hydrophone
10 2
10 1
MLE
5 50 0 1
ZLF 10 10
k BPO Number of peaks - microphone
DS
TB
c) d)
Fig. 7. 2D representations of the signal descriptors described in Section 4 for the signals of the learning database.
8
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
6.2. Results of the classification on the test database ZLF 0% 78% 0% 22% 0%
True label
algorithm is proposed. The classification method is first tested on
the test database extracted from the original labeled database. As BPO 0% 0% 88% 12% 0%
each point is labeled, the class attributed by the algorithm can be
directly compared to the true family of the point, allowing to esti-
mate the success rate of the algorithm. A success rate of 82% is
obtained for a value of k ¼ 4 which is the optimal value as shown DS 0% 0% 0% 100% 0%
in Fig. 8. The associated confusion matrix is shown in Fig. 9. It
allows to estimate which classes are often confused with others.
Four main problems can be observed.
TB 50% 0% 0% 50% 0%
The ‘zip-like flow’ class can be confused with the ‘donut splash’
class. This confusion was expected, because the two events are MLE ZLF BPO DS TB
hydrodynamically close to each other: both consist of a cavity
Predicted label
closed by an impacting secondary droplet, the difference being
that the secondary droplet impacts the cavity later in the case of Fig. 9. Confusion matrix from the test of classification using a KNN method.
‘donut splash’ than for ‘zip-like flow’. Thus, no physical limit is
defined between the two phenomena, leading to a probable
‘transition zone’ for which the mechanism is a mix between
The ‘touched bubble’ family is not successfully identified at all.
the two phenomena, and so the acoustic signal displays similar
This is probably due to the very low number of representative
features.
signals of this family into the learning database. This can also
The ‘bubble pinch off’ class can be confused with the ‘donut
be due to the fact that no feature specific to this family was
splash’ class. This is due to the high dispersion showed by most
found. However, the very low number of representative of this
of the descriptors for both classes. Indeed, both mechanisms
family in the test database (only two) is probably not relevant.
lead to the entrainment of bubbles of various sizes and with
various dynamics.
The algorithm is then tested on a set of measurements with the
The ‘Mesler-like entrainment’ class can be confused with the
same configuration, including one measurement of acoustic signals
‘zip-like flow’ class. This happens when large bubbles detach
synchronized with a video and 50 purely acoustic measurements.
from the air film induced in the ‘Mesler-like entrainment’
The configuration corresponding to this set of measurements is a
mechanism.
water drop impacting on a water tank with 5% ethanol. The acous-
tic signal with an associated video is shown in Fig. 10-a. The three
acoustic events detected are highlighted and identified thanks to
85 the associated video. The first event, detected at 134 ms, only
appears on the hydrophone signal with a low amplitude. It is iden-
tified as an MLE mechanism. The second event, located at 151 ms,
80 is due to a ZLF mechanism. Lastly, the third one, arriving at 184 ms,
is linked to a BPO mechanism. The 50 acoustic signals, represented
overlaid in Fig. 10-b, show 3 packs of acoustic events at similar
75 times as for the reference signal. Indeed, as the studied impact
regime is quite reproducible, it is not surprising to have similar
acoustic signals for a same configuration. The events from each
Success rate (%)
9
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
Fig. 10. a) Microphone (top, red) and hydrophone (bottom, blue) signals from a water drop impact on a water tank with 5% ethanol. The acoustic events are detected and
identify with the associated video footage. b) Microphone (top) and hydrophone (bottom) overlaid signals from acoustic measurements of a water drop impact on a water
tank with 5% ethanol. c) Percentage of each class label attributed to the detected events by the algorithm.
family for this pack), 29% as DS and 4% as MLE. Once again, the mistake as this family is always observed to appear first or not at
expected family is mostly identified. However, the success rate is all in all of our measurements. This test highlights a common con-
lower than the one obtained on the test database. The second most fusion between BPO and MLE that is not shown on the confusion
attributed class is DS. As explained earlier, ZLF and DS have similar matrix. The same confusion is also observed when the algorithm
hydrodynamic mechanisms and can happen with similar condi- is tested on other data set. After analysis, this confusion seems to
tions. Thus, the high score of DS can be explained in two ways: happen when a small bubble is entrained with a pinch off mecha-
nism, leading to an acoustic signal with properties similar to MLE
the strong resemblance between ZLF and DS signals, leading to signals. For the studied regime, as it is known that MLE events
ZLF signals being confused with DS ones, as observed on the are only seen as a first event when it appears, this confusion can
confusion matrix of Fig. 9, be corrected by ignoring the MLE class for non first events. This
the presence of both ZLF and DS signals in the pack. Indeed, as correction leads to a better and more successful classification.
the two events come from the closing of a cavity by a secondary The test of the algorithm on a large number of acoustic signals
droplet, it is possible that the little variations between mea- for a same configuration shows that, even if some errors remain in
sures lead to the apparition of one or the other for a similar the classification, the expected family is mostly attributed inside a
time. group of same events. The repetition of several measures for a
same configuration, combined with a detection of the group of
Finally, the third pack shows a classification at 62% as MLE, 25% acoustic events happening at similar time (therefore correspond-
as BPO (the expected family) and 13% as DS. In this case, the ing to the same family of event) and the determination of the class
mostly identified class does not correspond to the expected family. of each group by choosing the class attributed in majority inside it,
Unlike the precedent case, the identification of a MLE is clearly a can help improving the performance of the algorithm.
10
G. Gillot, L. Simon, J.-M. Génevaux et al. Applied Acoustics 196 (2022) 108868
11