Journal of Fluids and Structures 112 (2022) 103603

Contents lists available at ScienceDirect

Journal of Fluids and Structures

journal homepage: www.elsevier.com/locate/jfs

Proper orthogonal decomposition, dynamic mode

decomposition, wavelet and cross wavelet analysis of a
sloshing flow
∗
Tiziano Pagliaroli a , , Francesco Gambioli b , Francesco Saltari c ,
Jonathan Cooper d
a
Università Niccolò Cusano, Department of Engineering, Via Don Carlo Gnocchi 3, 00166 Rome, Italy
b
Airbus Operations Ltd, Loads & Aeroelastics Department, Filton, Bristol BS34 7PA, United Kingdom
c
University of Rome ‘‘La Sapienza’’, Department of Mechanical and Aerospace Engineering, Via di Eudossiana 18, 00184 Rome, Italy
d
University of Bristol, Department of Aerospace Engineering, University Walk, Bristol, BS8 1TR, United Kingdom

article info a b s t r a c t

Article history: Internal hydrodynamics and its coupling with structural dynamics are non-negligible
Received 3 November 2021 processes in the design phase of aerospace systems. An improved understanding of the
Received in revised form 9 March 2022 nature of this coupling would allow for greater flexibility in modeling and design of
Accepted 26 April 2022
such systems, and could lead eventually to the development of suitable active and/or
Available online 10 May 2022
passive control strategies for enhanced performances. In this manuscript we apply a
Keywords: number of data analysis techniques: proper orthogonal decomposition, dynamic mode
Wavelet decomposition and wavelet transform and their combination to time-resolved images of
Sloshing a liquid sloshing within an enclosure. We use these techniques to identify fluid-dynamic
POD modes in space and time and to verify their coupling with the structural dynamics of
DMD vibrating structures.
In particular we consider the transient case of a water tank mounted on a free
oscillating cantilever. As the acceleration amplitude decays, we observe and quantify the
transition from incoherent flow to standing waves. Our results show that the content
of the images is very informative and can be used for quantitative analysis. As the
main outcome, the hydrodynamic modes are identified using POD and DMD, and related
to known features of sloshing flow, such as the frequency of the first symmetric free
surface mode. Additionally we perform a comparison of wavelet transforms of POD time
coefficients and measured acceleration signals at the tank base. Viewing the latter as
the input and the former as the output of the fluid-dynamic system, we are able to
correlate the enhanced damping of the cantilever oscillation to the different regimes of
the hydrodynamic field.
© 2022 Elsevier Ltd. All rights reserved.

1. Introduction

The movement of the fluid free surface within its container, or sloshing, is of interests in various engineering fields. In
Aerospace Engineering, and particularly for modern satellites, the design is often driven towards lightweight structures,
high pointing accuracy and long life expectations (Konopka et al., 2019). These requirements result in a relatively large

∗ Corresponding author.
E-mail address: [email protected] (T. Pagliaroli).

https://doi.org/10.1016/j.jfluidstructs.2022.103603
0889-9746/© 2022 Elsevier Ltd. All rights reserved.
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

proportion of the overall weight being allocated to the liquid propellant. Typical space launchers also have a high ratio
of payload to propellant mass, 1:20 for the European Ariane 5 (Behruzi et al., 2006). Consequently, the liquid motion
can affect significantly the dynamic behavior of these systems, during in-orbit controlled maneuvers and gust encounters
in the atmospheric phase of the launch. Extensive analysis of the environment and the modeling of sloshing for space
applications is provided in Refs. Abramson (1966) and Dodge (2009).
Large passenger jets are designed to meet the regulatory requirements of Certification Specification 25 (CS25), as
defined by the airworthiness authorities — the European Aviation Safety Agency (EASA). The wings of such aircraft
are highly flexible structures, which deform when atmospheric turbulence or gusts are encountered or the aircraft is
maneuvered. Typical limit deformations can reach 20% of the wing span. Wings also house the fuel tanks, and generally
carry an amount of fuel comparable in weight to that of their structural components. For a typical wing of a single aisle
aircraft, with maximum take-off weight of 100 tonnes and 3000 nmi range, both structural and fuel weights are of the
order of 4 to 6 tonnes. In the design of civil aircraft, the study of fuel movement within tanks is of paramount importance
for the design of the fuel management control system, the evaluation of the handling characteristics of the aircraft and
ultimately the assessment of the structural response of the containment structure (Gambioli and Malan, 2017). Other
specialized applications are related to the transport of non-flammable liquids, such as water for fire-fighting purposes.
The fuel sloshing effects on the aeroelastic response of the aircraft structures have also been studied, with focus on the
changes of the flutter boundaries when compared with dry-tank configurations, see amongst others Ref. Hall et al. (2015).
However the standard engineering practices for wing (and aircraft) design do not consider the effect of the fuel movement
for the estimation of the aircraft design loads, such as those arising from discrete gust and continuous turbulence. In fact
the type of excitation in those scenarios is quite peculiar to air transport, with the acceleration direction perpendicular
to the free surface and magnitudes in excess of 10 g. Also the tank motion is the result of the dynamic response of its
flexible container, making the phenomenon inherently hydroelastic in nature.
In Civil Engineering, liquid tuned mass dampers are commonly employed to reduce the response of tall towers to the
wind or earthquakes. These are tanks, partially full of liquid, whose purpose is to increase the damping of a structure
by using the free surface motion of the fluid to counter any excitation (Kareem et al., 1999). Similarly, anti-roll tanks
can be employed on ships in Refs. Gawad et al. (2001) and Bouscasse et al. (2014a,b) to damp the vessel motion in
rough sea conditions. Inspired by these applications, which take advantage of the liquid motion to benefit the overall
design, we investigate the effects of fuel sloshing on the dynamic response of flexible wing-like structure. While traditional
approaches for load alleviation aim at active or passive control of the external aerodynamic load, via rapid deflection of
control surfaces (Fonte et al., 2017) or passive structural methods such as aeroelastic tailoring (Stodieck et al., 2017), in the
present work we focus on the dissipative effects induced by the liquid sloshing. It is worthwhile noting that the sloshing
phenomenon investigated in this paper received little attention in the past. In fact in classic liquid damper applications,
the tank moves parallel to the liquid free surface thus generating waves inside the cavity. Consequently, the sloshing
fluid behaves as a dynamic absorber that dissipates energy through breaking waves and impacts with the side walls of
the tank. This behavior, together with a proper tuning of the sloshing natural frequencies, allows for the passive control of
structural vibrations. Instead the subject of this paper is sloshing induced by high vertical acceleration of the tank, hence
perpendicular to the free surface. This triggers Rayleigh–Taylor instabilities (Lewis, 1950) during the first oscillation cycle,
determines a chaotic flow regime with air/water mixing in subsequent cycles until a well-defined free surface reforms and
standing waves are clearly observable. Turbulence, impacts and the continuous generation of the free surface (bubbles)
cause additional dissipation of energy compared to the case of a solid mass. The total balancing of elastic potential energy
and fluid energy results in a noticeable increase in the effective damping of the structural motion. We focus our work on
the analysis of a set of physical experiments carried out at the Airbus Protospace facilities in Filton (UK) in August 2018.
The campaign demonstrated that liquid sloshing affects the dynamics of a free-vibrating cantilever beam, leading to a
substantive increase in its damping characteristics (Gambioli et al., 2019; Titurus et al., 2019). These preliminary findings
formed the basis for the SLOshing Wing Dynamics (SLOWD) H2020 project, which aims to exploit the potential of fuel
sloshing as a mean for aircraft loads alleviation (Gambioli et al., 2020).
In terms of experimental characterization of sloshing flows, activities during 1960’s NASA space programs focused
on the determination of equivalent mechanical model parameters via system identification techniques, complemented
with photographic images of the generally linear sloshing flow regimes (Abramson, 1966). More recently non-intrusive
measurements such as PIV (Particle Image Velocimetry) have been employed successfully to measure the velocity field
and track the free surface (Simonini et al., 2014, 2015, 2016). Owing to its importance for Liquid Natural Gas (LNG)
tanker design, the measurement of unsteady pressures due to wave impacts have been the focus of intense research, see
amongst others (Delorme et al., 2009; Brosset et al., 2009; Ibrahim, 2020). In general sloshing experiments are filmed
with high-speed cameras, but are resulting images are used for the identification of qualitative features of the flow, with
limited exceptions (Saatci and Yazıcı, 2019; Tosun et al., 2017).
In this paper we apply the Proper Orthogonal Decomposition technique (POD) in combination with the wavelet
transform to snapshots of the sloshing fluid recorded via high-speed camera. To our best knowledge, the combined use
of these techniques is novel and not found in existing literature. Similarly to the POD, the Dynamic Mode Decomposition
(DMD) is used to extract relevant flow features in terms of modal frequency and damping. Both techniques decompose
ψ
the hydrodynamic flow field in coherent modes in time and space (aij and φi ). There are limited examples in literature
of modal decompositions of free surface flow based on data, and, to the knowledge of the authors, none based on
2
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 1. Sketch of the experimental set-up. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)

experimental images. In fact, Refs. Williams et al. (2012) and Moya et al. (2019) applied POD and DMD respectively to
free-surface waves and sloshing in a cylindrical tank based on numerical simulations. We find that the spacial coherent
structures identified via POD and DMD are similar and that the assessment of the time coefficients via the POD-Wavelet
transform compares well with the modal damping and frequency identified using by the DMD.
Focusing on the structural dynamics, we use the Wavelet transform to analyze the accelerometer signals and to
characterize the dynamic behavior of the system in time and frequency. In particular we are able to estimate the time-
varying damping coefficients at given frequency bands, and compare the corresponding dissipation for wet and dry
configurations.
Further we show that the cross-wavelet spectra of acceleration signals and POD time coefficients can be used to detect
strong interactions between the fluid and structural dynamics. We conclude that the proposed techniques, when applied
in combination to accelerometer and video signals, can provide useful insights into the dissipative effects caused by the
coupling of violent sloshing motion with structural flexibility.
This paper is organized as follows: first we present the experimental setup and the free oscillation tests in Section 3,
then we provide a basic description of the methods used analysis of the experimental data in Section 2. Finally we analyze
and correlate the POD/DMD results with the accelerometer signals via cross wavelet in Section 4.2.

2. Experimental setup

The main structural member is a cantilevered steel beam of 2.35 m span (scale 1:5 with respect to a single aisle airliner
wing). A water tank is mounted onto the beam at the free-end side, spanning 0.7 m, as shown in Fig. 1. The quick release
mechanism (shown in the bottom right insert of Fig. 2) is employed to force the free-end down, to a condition close to
maximum elastic stress in the beam but still within the geometrically linear regime (the tip deflection is less than 5% of the
span), and suddenly released. The free oscillation of the system are then monitored by the various triaxial accelerometers
(Isotron 65L-100) at a sampling frequency of 1 kHz and the fluid motion is imaged by a high-speed camera (Photron
Fastcam SA1.1) at 3000 fps with a resolution of 1024 by 1024 pixel for a duration of 5 s circa. The data acquisition is
performed via a Vishay System 6200 data logger and acceleration time histories are measured at the geometrical center
of the cantilever tip section (shown by the arrow in Fig. 2). The tank is mounted on the cantilever at an angle so that in the
initial deflected position the tank bottom (and roof) are horizontal. It is attached to the beam at four points (Fig. 2 bottom
left) by pin-joints. This design allows for minimal transfer of longitudinal forces during the free oscillations, so that the
tank can be assumed rigid in first approximation if the deflection of the beam is limited to the linear deformation regime.
By design the overall setup is configured to limit three-dimensional effects in both the motion of vibrating structure and
the fluid flow in the tank, to ease the characterization of the coupled system. It is also important to notice that the total
mass of the system can be controlled by the ballast placed beneath the tank on the underside of the cantilever (Fig. 2
bottom center). The analysis of such types of couple fluid–structure system is not available in literature. In fact, the nature
of the excitation applied to the fluid by the structure is peculiar to aircraft wing (or lifting surface) applications, in that it
is predominantly vertical (in direction perpendicular to the free surface); is of periodic nature, and not limited to a single
impact event; can reach high acceleration levels (well above gravity), leading to the liquid departing from the bottom of
the tank. In space applications the sudden deceleration at the separation of a launcher stage can lead to severe impact
forces, but these are generally resolved to a single event as opposed to the cyclic wing oscillations which are repetitive
in nature. Furthermore the shape of the tanks in launchers is significantly different from those in aircraft wings.
The experimental configuration is designed to be similar to a real wing geometrically (spanwise position, width and
height of the tank); to achieve vertical accelerations of the same order of magnitude of limit design cases of an airliner
wing (10 g) and such that the ratio of liquid and solid masses (10%) is representative of typical design conditions. The
tank is divided in seven sealed compartments, to mimic the typical internal geometry of a wing tank which includes
ribs. There are no openings on interior walls, to reduce the complexity of the fluid flow, as these are assumed to have
secondary effects for primarily vertical excitations.
3
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 2. Photos of the experimental set-up.

Fig. 3. Time response of the acceleration time history: time response is marked with blue solid line for the baseline, red dashed line for the test
case 1, and black dash-dotted line for the test case 2. (For interpretation of the references to colour in this figure legend, the reader is referred to
the web version of this article.)

Table 1
Considered test cases.
Case Ballast mass [kg] α [–] Fluids Water mass [kg] Ballast-water mass [kg]
Baseline 0.528 0.0 Air 0.000 0.528
Test case 1 0.330 0.3 Water & air 0.198 0.528
Test case 2 0.198 0.5 Water & air 0.330 0.528

Among the tests carried out during the experimental campaign (Ref. Gambioli et al. (2019)), the present activity focuses
on three different sloshing configuration expressed in Table 1 by defining different combination of ballast placed beneath
the tank and filling level ratio α . Specifically, the amount of sloshing fluid mass and ballast is set constant in order to
keep the vibration frequencies as similar as possible. Fig. 3 shows the time response of the accelerometer placed at the
beam tip (red arrow in Fig. 1). Blue solid line response is associated with the beam response in the baseline configuration,
whereas red dashed line and black dashed–dotted line provide, respectively, the measured beam tip acceleration for the
test case 1 (α = 0.3) and test case 2 (α = 0.5). It is already noticeable that when sloshing water is present inside the
tank, a complex fluid–structure interaction provides an increased value of overall damping.

3. Materials and methods

In the present work, we apply different data analysis techniques and their combinations to investigate the fluid–
structure interaction of a sloshing fluid and a flexible slender structure. Fig. 4 depicts schematically the process followed
for the analysis. The fluid-dynamic input data is a set of time-resolved flow images, which are decomposed in spacial
and temporal coefficients using proper orthogonal decomposition (POD), see Fig. 4a. The temporal coefficients are further
analyzed using the wavelet transform: the wavelet coefficients have been integrated in time to obtain a representation
of the signal energy expressed in terms of ‘‘stabilized’’ spectra (a spectrum quite similar to the one obtainable by Fourier
4
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 4. Sketch of the processing strategies applied to the experimental data: flow chart of the data analysis strategy applied to the fluid dynamics
(a) and structural dynamics database (b). Data post-processing adopted for highlighting the Fluid–structure interaction (FSI) (c).

transform, but much less noisy Pagliaroli et al., 2018a, 2020). This procedure allowed us to identify global temporal
coherence represented in terms of energetic peaks in the frequency domain.
The structural dynamics of the system is measured using accelerometers, whose signals are analyzed using wavelets
(Fig. 4b). The resulting wavelet coefficients have been used for multiple purposes:

1. to calculate the damping in each of the beam modes as a function of time;

2. to represent the dynamics of the beam in terms of stabilized spectra;
3. to characterize the dynamic behavior of the beam vibration mode frequencies that are not constant over time.
Finally cross-wavelets coherency and phase delays between the acceleration time history and POD time coefficients,
to characterize the coupling within the systems, were computed (Fig. 4c).
We have found that the wavelet transform is particularly effective in the characterization of the transition between
different flow regimes and facilitated their interpretation, which is typically one of the limitations of the decomposition.

3.1. Moving volume tracking algorithm, POD and DMD

Before applying the POD and the DMD the image sequence was subjected to a pre-processing related to the fact that
the camera is not integral with the oscillating beam. For this reason the region of interest (ROI) of the image, contained
in the field of view (FOV), has a time-dependent boundaries. During testing, the beam is released from an initial deflected
shape and oscillates within the FOV. The sloshing flow is recorded within a volume in relative motion to the camera, and
the video includes a large background not required for the flow analysis. Typically both POD and DMD occur applied to
ROIs that have constant boundaries in time. To overcome this aspect pre-processing is applied to the flow images:
5
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

1. tracking of the tank position in space, using an edge detection algorithm, assuming planar motion;
2. applying a series of geometric transformations (rotation/translation and scaling) so that the tank is moved back to
its initial position;
3. cropping the area of interest for the flow analysis.
The resulting video sequence is fed to the POD and DMD algorithms as an input for the flow decomposition. The video
sequence is arranged in a matrix of snapshots, whose columns are vectors of gray-scale intensity.
As already mentioned the fluid dynamics will be analyzed through two techniques, one combining POD (Berkooz et al.,
1993; Maurel et al., 2001; Meyer et al., 2007; Sieber et al., 2016) and wavelet transform, herein proposed for the first time
by the authors, and the conventional DMD (Tu et al., 2014). The first technique aims at analyzing an image time-resolved
sequence I(x, t) such that
N N M
ψ
∑ ∑ ∑
I(x, t) ≈ ai (t)φiPOD (x) ≈ aij (t)φiPOD (x) (1)
i=1 i=1 j=1

ψ
where the modes φiPOD (x) given by Proper Orthogonal Decomposition and the coefficients aij (t) given by Wavelet
Transform are, respectively, coherent in time and space. On the other hand, Dynamic Mode Decomposition aims at
estimating natural modes, frequencies and damping ratios of the system. The homogeneous solution can thus be expressed
as
N √
−ζi ωi ± iωi 1−ζi2
∑
I(x, t) ≈ φiDMD (x) c0i e (2)
i=1

where φiDMD (x) are the DMD modes expressed on the basis of POD modes, c0i is a constant related to the initial condition
and ωi and ζi are the natural frequencies and damping ratios, respectively.

3.2. Wavelet transform

Wavelets are a relatively recent instrument developed in applied mathematics around thirty years ago (Farge,
1992) and have become a common and attractive tool for time series analysis. In particular, in the last two decades,
wavelet transform (WT) has been used for a number of studies, e.g. capturing sea surface thermodynamic variable
oscillations in ocean science (Torrence and Compo, 1998), studying coherent structures in turbulent flows (Farge,
1992), signal denoising (Stefanutti and Bruni, 2017), studying intermittent phenomena in aeroacoustics and solid rocket
motors (Pagliaroli et al., 2015, 2018a), and, very recently, investigating the acoustic behavior of meta-surfaces for boundary
layer transition control (Pagliaroli et al., 2020, 2018b). WT permits to decompose a space/time history expanding the
choice for a proper basis among a great variety of suitable functions, allowing to tailor the selected wavelet to each
specific problem (Ashmead, 2012). One of the main properties of wavelet basis functions is their localization in both
time and frequency domains. As a result, wavelet series better fit wave forms usually found in nature and often converge
faster than corresponding Fourier series. The starting point for continuous wavelet transform (CWT) consists in choosing
a proper wavelet function ψ (t), called the mother wavelet, which can be real or complex (Mallat, 2009). In particular,
the present analysis was performed by selecting a complex mother wavelet named Morlet, which is a solution of wave
equation (Visser, 2003) and fits the nature of our problem. The Morlet mother function is a plane wave modulated by
Gaussian envelope of unit width (see Fig. 5(a)):
1 t2
ψ (t) = π − 4 eiω0 t e− 2 (3)
where t and ω0 are the dimensionless time and wavelet center frequency, respectively (Farge, 1992; Torrence and Compo,
1998). Incidentally, this mother wavelet is only marginally admissible: it is of zero average only if some very small
correction terms are added (Farge, 1992; Ashmead, 2012). In practice, if we take ω0 = 6, the correction terms can be
neglected. Another way to ensure the admissibility is to impose equal to zero the Fourier transform of the Morlet wavelet
for ω = 0, i.e. ψ̃ (0) = 0.
In Fourier space, the Morlet wavelet is given by:
{
(ω−ω0 )2
−1 −
ψ̃ (ω) = π 4 e 2 for ω > 0 (4)
0 for ω ≤ 0
Starting from the mother wavelet function ψ , a family of continuously translated and dilated wavelets (the orthogonal
basis function) can be generated and normalized in energy norm:
t −τ
( )
ψs,τ (t) = sn ψ with s, τ ∈ R+ (5)
s
where s is the scaling parameter, τ the time shifting,
∫ ∞ and n the normalization exponent. In the present analysis,
L2 -normalization has been applied, thus: n = −1/2 , −∞ |ψs,τ |2 dt = 1.
6
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 5. Real (ℜ(ψ )) and imaginary (ℑ(ψ )) part of the Morlet mother wavelet (a); dilation effect on wavelet function represented in time (b) and
Fourier domain (c).

The role of s in time and Fourier domain is clarified observing Fig. 5(b), (c): increasing s corresponds to dilation in time
and contraction in frequency domain.
Finally, the continuous WT is defined as:
+∞
t −τ
∫ ( )
−1/2
w(s, τ ) = s x(t)ψ ∗
dt (6)
−∞ s
where w (s, τ ) are the wavelet coefficients (Pagliaroli et al., 2015, 2018a; Mancinelli et al., 2017, 2016) and ψ ∗ is the
complex conjugate of the dilated and translated mother wavelet.
Importantly, for non-sinusoidal periodic waveforms, the Fourier transform spreads the energy across several sub-
harmonics. By contrast, wavelets select only few components with a significant energy content.
It is worth noting that the scale s can be easily converted into pseudo-frequency by using the equation (Torrence and
Compo, 1998):
⎡ ⎤−1
4π s
f =⎣ √ ⎦ (7)
ω0 + 2+ω 2
0

This conversion allows to maintain a link with Fourier domain, even though the signal processing is performed by applying
WT.

3.2.1. Wavelet damping ratio

The Morlet wavelet coefficients can be used to compute the natural frequency and the damping ratios after isolating
individual modes of vibration systems (Staszewski and Cooper, 2002). The damping ratio ζ and natural frequency ωn can
be estimated, respectively, from the slopes of the logarithm of modulus and phase of the wavelet coefficients with respect
to parameter τ . According to the mathematical formulation proposed by Chen et al. (2009), we define a wavelet-based
and time-dependent damping ratio:
1 ∂ ln |w (f , τ )|
ζ ψ (τ ) = − (8)
ωn ∂τ
where f is pseudo-frequency previously described. This indicator can be useful to monitor the effects of sloshing in terms
of damping over time, particularly as damping is affected by the hydrodynamic motion regime.

3.2.2. Wavelet coherence and phase

In analogy with Fourier energy density spectrum, we can readily define a wavelet auto-spectrum for a data series x(t)
as (Liu, 1994):

Wxx (s, τ ) = wx (τ , s)wx∗ (τ , s) = [ℜ(wx )]2 + [ℑ(wx )]2 = |wx (τ , s)|2 (9)

Results from the application of short-time Fourier transforms have been called spectrograms, whereas results from the
application of wavelet transforms have been called scalograms. Since in practice the scale, s, and translation, τ , can be
associated with a corresponding frequency, f , and time, t, can be considered as a representation of the time-varying,
localized energy auto-spectrum for a given time series.
We can similarly define a cross wavelet spectrum for the study of two simultaneously time histories x(t) and y(t) as:

Wxy (s, τ ) = wx (τ , s)wy∗ (τ , s) (10)

7
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Here, the wavelet coherency is defined using smoothing in both time and scale.
Traditionally, Fourier coherency has been used to identify frequency bands where two time series are related. On
the other hand wavelet coherency identifies both frequency bands and time intervals when the time series were
related (Torrence and Webster, 1999). This latter feature is the main motivation to apply wavelet coherency to strongly
unsteady phenomenon instead of taking the Fourier route.
The wavelet coherence of two time series x(t) and y(t) is:
Wxy (s, τ )
Γ (s, τ ) = [ ] 12 (11)
Wxx (s, τ )Wyy (s, τ )
For real-valued time series, the wavelet coherence is real-valued if you use a real-valued analyzing wavelet, and complex-
valued if you use a complex-valued analyzing wavelet. The wavelet-coherence based phase difference is given by:
[ ]
−1 ℑ(Wxy )
Wxy = tan (12)
ℜ(Wxy )
Another interesting wavelet-based variable is the wavelet coherency, defined as follows:
|Wxy (s, τ )|2
Γ 2 (s, τ ) = (13)
Wxx (s, τ )Wyy (s, τ )
It is worth noting that Γ 2 (s, τ ) is always a real value. Finally, it is noted that because the wavelet transform conserves
variance, the wavelet coherency is an accurate representation of the (normalized) covariance between the two time
series (Torrence and Webster, 1999). S is a smoothing operator in time and scale.

4. Results

The study on the fluid–structure interaction of the present experimental model is carried out on several levels. First,
we proceed to identify the nominal free response of the structure without sloshing. A flow field pre-qualification is then
done using certain snapshots of the free response. Finally, these analyses allow to study the structural and fluid dynamics
response through wavelet analysis and the fluid–structure interaction through the recognition of specific patterns and
cross-wavelet analysis.

4.1. Free vibration response

The characterization of the structural behavior of the cantilever beam being object of this analysis plays a key role to
understand the type of unsteady boundary conditions that the structural dynamics provide to the fluid dynamics domain.
We focus on the first two out-of-plane bending modes since they are the ones clearly involved in the free response of
the beam. Table 2 lists the natural frequencies and damping ratios associated with the case in which the liquid inside
the tanks is replaced with equivalent masses. Further details about their experimental evaluation can be found in Titurus
et al. (2019). By making use of a finite element model properly tuned to have a natural frequency scenario similar to the
experimental one, the modal shapes associated with the first two out-of-plane bending modes are shown in Fig. 6. Thus,
the motion of the tank due the first vibration mode consists of a marked vertical deflection able to trigger the Rayleigh–
Taylor instabilities. The high amplitude values of this mode can also cause a shortening effect of the beam spanwise
projection that provides acceleration to the liquid inside the tank in a centripetal direction. Concerning the second mode,
the tank center position is close to the node of the second mode and, therefore, the consequent motion mainly consists
of rotation.
Considering the free response of the structure, it is certainly of interest to understand the participation of each vibration
mode in the initial conditions imposed by the quick release mechanism (see Fig. 2). In particular, it is worth providing the
modal participation levels to the acceleration that the cantilever tip would experience in the ideal case of instantaneous
release. By performing a numerical linear static analysis to replicate the initial conditions set in the experiment (i.e. 7 cm
vertical displacement at the tip), it has been found out that the modal participation factor (obtained by projecting the
overall displacements on the actual mode shapes) of the tip vertical displacement is 6.83 cm and 0.32 cm downward for
first and second modes, respectively (Titurus et al., 2019). The total initial tip displacement is y0 = y0,1 + y0,2 + · · · where
the terms y0,i may assume negative or positive values. As expected, the imposed initial conditions excite mainly the first
vibration mode (i.e. the initial condition ratio is about 21), and, in case of free harmonic oscillation, one has
y(t) = y0,1 cos(ω1 t) + y0,2 cos(ω2 t) + · · · (14)
However, truncating the above decomposition to the first two modes, one has
ÿ(t) = −y0,1 ω12 cos(ω1 t) − y0,2 ω22 cos(ω2 t) = ÿ0,1 cos(ω1 t) + ÿ0,2 cos(ω2 t) (15)
where ÿ0,1 = 147.65 m s and ÿ0,2 = 352.32 m s for a total initial acceleration tip value of around 500 m s . A non-
−2 −2 −2

instantaneous release of the structure results in an initial acceleration lower than the ideal case as shown in Fig. 3, that
8
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Table 2
Experimental modal analysis results.
Mode fn [Hz] ζn [–]
1 7.12 0.52%
2 49.26 0.76%

Fig. 6. Out-of-plane bending modes of cantilever beam.

depicts the acceleration tip free response for the three considered cases, that is, the baseline (frozen mass configuration),
the case 1 (α = 0.3) and case 2 (half filled tanks, α = 0.5). It can be noticed immediately that in cases 1 and 2 there is a
much higher damping than the baseline. This finding will be considered more in detail in the following sections.

4.2. Flow field pre-qualification

The tank motion is driven by the initial deflection of the cantilever. The displacement is mainly vertical, with additional
effects due to the rotation in x–y plane as well as displacements along y due to beam shortening. These result in
violent vertical free-surface motion, with secondary effects due to centrifugal forces experienced by the fluid. The typical
time history measured by the tip accelerometer is shown in Fig. 7a, where the vertical dashed lines identify the time
windows corresponding to four different flow regimes. For each hydrodynamic regime sample time points are indicated
by red circles on the accelerometer signal, and Fig. 7b–m show the corresponding flow configurations. We note that the
acceleration ÿt (t) plotted in Fig. 7a is derived from the accelerometer signal as:

ÿt (t) = g − ÿ(t) (16)

2
where g is the constant gravitational acceleration (9.81 m/s ). This is the acceleration experienced by the liquid in the
non-inertial reference frame moving with the tank, and therefore is representative of the time-varying body forces seen
by the fluid during the oscillations.
Immediately after the quick-release, in the non-inertial reference frame attached to the tank, the fluid experiences
’’augmented gravity’’ due to the tank being accelerated upwards by the oscillation of the beam. During the first quarter-
cycle there is minimal fluid motion relative to the tank, but two distinct effects are observed. First, and most apparent,
the free-surface starts drifting towards the free-end of the cantilever (Fig. 7b), as a result of the centripetal acceleration
experienced by the tank. Second, due to adhesion effects, the free surface of the liquid is elevated in the shape of a
meniscus in small regions close to the tank walls (not shown in Fig. 7); these evolve into gravity waves which perturb
the free surface.
After the first quarter-cycle the acceleration of the tank changes direction, and the fluid experiences a short transition
phase of reduced gravity (in the frame moving with the tank) eventually reaching zero-g until the acceleration reverts
direction and starts to push the liquid upwards (inverted gravity). At this time the free surface perturbations, initiated
by the elevation at the side walls, evolve into Rayleigh–Taylor (RT) instabilities which eventually assume the finger-like
structure shown in Fig. 7c. These jetting structures are typical of conditions in which a heavy fluid is accelerated towards
a lighter one (liquid and air in the case under consideration), and are due the opposing effects of buoyancy and inertia
forces.
Fig. 7d shows the time of first liquid impact with the tank roof. Subsequent to this impact, the flow at the top of the
tank starts to become incoherent, with free surface breakup and air bubble entrapment, while the RT fingers on the left
side of the tank roll-up and eventually form the lambda-like structure indicated in Fig. 7e. Fig. 7f shows the propagation
of lambda structure from left to right at the bottom of the enclosure and expansion of the free surface breakup area in
the upper part of the tank, following from the impact with the roof.
The expansion of the breakup area continues, leading to a completely incoherent flow field anywhere in the tank
(defined washing machine hereafter), as it can be seen in Fig. 7g. In the washing machine regime, the fluid is excited by
the vertical oscillatory motion with high acceleration levels (or the order of 10 g) and continuously impacts the tank walls
9
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 7. Vertical acceleration time history, reported in semi-log scale (a) and pictures of the hydrodynamic flow field corresponding to the different
motion regime: augmented gravity (b); Rayleigh–Taylor instability (c–e); first transition (f); washing machine (g); second transition (h); and standing
Faraday wave (i–m). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

preventing the formation of a stable free surface, while promoting the air–water mixing and air bubbles entrapment. This
phase lasts for six oscillation cycles and, as it will be illustrated later, is associated with the greatest amount of energy
dissipation with significant influence on the structural response of the cantilever.
As soon as the acceleration levels decay below zero, and consequently the direction of the inertia force stays constant
and downwards, the fluid transitions back to a condition of a well-defined free surface. This effect is illustrated in Fig. 7h,
where the free surface begins to re-establish with the liquid collecting at the bottom of the tank. In the final regime a
standing wave forms inside the tank. As shown in Fig. 7i–m and the corresponding time points on the graph in Fig. 7a,
the oscillation period of the wave is twice that of the cantilever, as is typical for Faraday waves. It is of interest to notice
that the shape of the free surface is indicative of a non-linear, high amplitude behavior. In fact at its apex, see Fig. 7m,
the two side slopes of the wave meet almost at a right angle, departing from the linear sinusoidal shape typical of linear
phenomena.

4.3. Wavelet analysis

The present analysis aims at deepening the content of the accelerometer measurement of the tip of the H-beam
and the role of the sloshing fluid contained in the tanks. In particular, the harmonic content and the damping of the
first two vibration modes (see Ref. Titurus et al. (2019) and Constantin et al. (2021)) are featured by complex fluid–
structure interactions that evolve over time with the transition between the different fluid dynamic regimes. This type
of characterization requires an analysis in the time–frequency domain, such as continuous wavelet transform, that, in
several applications proved particularly suitable for studying unsteady signals that change their properties over time. The
proposed application, featured by three different phases, makes the wavelet transform appropriate to study the transition
from one regime to another. In this regard, the wavelet transform performs the projection of the measured time histories
on compact support properly defined in the same domain. This feature results useful to identify in time–frequency domain
the behaviors related to transients that could correspond to changes of regimes.
The logarithmic values of the squared wavelet coefficients of the acceleration, computed for the baseline and test case
2, are reported in Fig. 8.
In both cases three dominant narrow-band components can be observed: (i) the first harmonic of the first longitudinal
mode of the H-beam (f1 ); (ii) the second harmonic of the first mode (2f1 ); (iii) the second structural mode (f2 ). It is
noticeable that the first and second structural modes are time-dependent for both cases. Specifically, three different
outlines can be extracted from the wavelet analysis, i.e., (i) the damping associated to the first mode, (ii) the frequency
drift from a lower frequency to a weakly higher frequency, (iii) the frequency modulation of the second structural mode
that for the first 2 s oscillates in time. This feature is confirmed by the serrated pattern pointed in Fig. 8a by the black
arrow.
Consider now physically characterizing and interpreting the highlighted narrow-band components and their temporal
dependence. Specific attention is given to the dissipative behavior introduced by the fluid by focusing on the incremental
10
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 8. Wavelet transform of the vertical acceleration time series represented upon time and pseudo-frequency for two different cases: baseline (a)
and case 2 (α = 0.5) (b).

Fig. 9. Wavelet energy distribution in time for three different modes f1 (a); 2f1 (b) and f2 (c). For all plots solid line is referred to the baseline,
dashed line represents the case 1 and dashed–dotted line is used for case 2.

modal damping due to sloshing. For the sake of clarity, it is worth noting that the baseline behavior is also not strictly
linear due to possible unidentified frictions within the structure. In particular, Fig. 8a shows how the drift of the first mode
frequency is already present in the baseline dry case, as well as the modulation of the second vibration mode frequency.
Therefore, these features are caused only in part by water sloshing.
Fig. 9 shows the wavelet energy distribution in time associated to the harmonic contents f1 , 2f1 and f2 for the three
considered cases.
11
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 10. Damping ratios based on wavelet transform in time dependent form, calculated at the fundamental frequency (f1 ), for baseline (solid line);
case 1 (dashed line); and case 2 (dashed–dotted).

From the comparison of the instantaneous energy associated to such harmonics, it is clear that the water contained
in the tank significantly dampens the f1 , 2f1 and f2 modes that decay to zero earlier than the dry case. Indeed, the dry
case maintains the fluctuating energy level higher for about three seconds: in the first two seconds (time window A) the
amplitude decay is linear in log scale, whereas in the third second (time window B) the mode decay is abrupt. Finally
within the time window C the decay starts to be linear again, but with a less steep slope. With reference to the first mode,
this aspect suggests a significant variation in damping over time. Since even the baseline instantaneous energy does not
exhibit constant damping, the resulting damping associated to the first mode slightly depends also on the unidentified
friction within the structure. Nevertheless, the amplitude of the wavelet coefficients for the baseline is always higher than
in cases 1 and 2 even after the first damping phase (for t > 3 s).
In order to have a representation of the damping in time, a method of calculation of this fundamental parameter
based on the wavelet transform has been applied. This methodology allows us to represent the damping, ζ ψ (fp , t), on a
time–frequency domain, as any other wavelet coefficient. Setting fp = f1 (t) it is possible to show the time history of the
damping relative to the first mode of the beam, the most energetic one. In Fig. 10a, we can see that ζ ψ is a significantly
time dependent, in fact it exhibits for all cases a noticeable hump. The left tail of the bump is much less steep than the
right, this aspect indicates that the damping increases gradually over time until it reaches a critical value, its maximum,
downstream of which there is an abrupt reduction of the oscillations and an entry into a dynamic regime different from
the previous one. This bump occurs earlier in case 1 and 2 than in the dry case and assumes an higher maximum value.
In particular, for case 2 the global maximum of the damping is the highest, so according with Refs. Titurus et al. (2019)
and Constantin et al. (2021) this value is a growing function of the water mass. It is also interesting to note the behavior
of the damping in the first 3 s and afterwards. In all cases the damping value is ζ ψ ∼ O(10−2 ) for the first three seconds
while in the remaining seconds ζ ψ ∼ O(10−3 ) (see Fig. 10a). This aspect, too, indicates the presence of two macro regimes
in the structural dynamics. Although time dependence is useful to describe the phenomenon qualitatively, for modeling
purposes dependency from amplitude is easier to handle, Observing Fig. 10b we can conclude that ζ is a function of
wavelet coefficients and time, i.e., of acceleration and time:

ζ = f [w(f1 , t)] = ζ (ÿ) (17)

This latter analysis allows us to propose a preliminary mathematical model to describe the first mode. Neglecting the
small frequency drift observed, assuming in first approximation the frequency of the first mode constant over time, based
on the experimental evidence we can write the following equation that describes the dynamics of the first mode:

ÿ + 2ζ (ÿ)ω1 ẏ + ω12 y = 0 (18)

with ω1 = 2π f1 . The Eq. (18) is descriptive of a nonlinear dynamical model: the second term at the first member is
dependent by both ÿ and ẏ. Nonlinear dynamical models often exhibit a strong sensitivity to initial conditions. If this is
the case for the model described we should expect a chaotic transition between two or more states of the system.
At this point we would like to shift our attention to the second mode of vibration of the beam, referring to Fig. 9
c, we observe that the wavelet coefficient relative to the mode (f2 ) presents significant oscillation in the first damping
phase (t < 3 s), these oscillations are due to the oscillation at frequency f2 , as confirmed by the scalogram in Fig. 8a. This
aspect is commonly addressed in classical and quantum physics as an oscillator with time-dependent frequency (Dittrich
12
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 11. Absolute value of the Fast Fourier Transform of the wavelet coefficient time series 2 log[w (f2 , t)] for the baseline.

and Reuter, 2017) whose differential equation is given by:

ÿ + ω22 (t)y = 0 (19)

for a known functional dependency of the stiffness, ω2 (t), the Lagrangian can be exploited for modeling purposes.
The oscillations of the frequency f2 has been calculated by means of Fast Fourier Transform (FFT) of the wavelet
coefficient represented in Fig. 9c (solid line), the resulting FFT is shown in Fig. 11. As result the oscillation frequency
of f2 is equal to f1 . Hence the time-dependent natural frequency can be modeled as:

ω2 (t , ω1 ) = ⟨ω2 ⟩T (1 + c sin (ω1 t)) (20)

where ⟨·⟩T stands for the Reynolds average over a period. In other words there is a coupling between the first and second
mode of the beam.
As for the response of the second mode, rather than introducing a time-dependent damping, it can be assumed that
this is characterized by a forcing term due to beatings present when the first mode has large oscillations. Ultimately, the
experimental results led to the writing of a simplified mathematical model of the first and second modes of oscillations
such that:
ÿ1 + 2ζ (ÿ1 )ω1 ẏ1 + ω12 y1 = 0
{
(21)
ÿ2 + 2ζ ω2 (t , ω)1 ẏ2 + ω2 (t , ω1 )2 y2 = F (ÿ1 )
The main objective of the mathematical model is to synthesize the results obtained so far and highlight the need to use
advanced and ad hoc analysis techniques for this type of phenomenon.

4.4. Comparison between POD and DMD results

In Fig. 12 are shown the normalized eigenvalues of the first ten POD modes of the sloshing flow for test case 2,
calculated from the video image sequence. It is worth nothing that the 90% of the image energy is associated to the
first dominant mode. This is because, the average image has not been subtracted from each instantaneous realization.
In this way, the temporal coefficient of the first mode embeds useful information to monitor the transient behavior
of the phenomenon and the higher modes are representative of the fluctuation around the transient. This aspect is
confirmed observing Fig. 13a, where the first POD time coefficients are plotted in normalized form. For this coefficient
different characteristics can be observed into two different time windows labeled I and II. From 0 to 1 s, around a
mean value equal to −0.0064 uncorrelated oscillations can be observed. Whereas, from 1.5 to about 3 s there are a
low-frequency periodic oscillations that alternately presents a very energetic pulse and one of lower amplitude. These
fluctuating components oscillate around an average value of −0.0132 much lower than the previous value in the first
time window. Furthermore, looking at the inset in Fig. 13b, still referring to time window II, another interesting feature
of the coefficient in its oscillatory regime can be seen: there is an additional periodic but not sinusoidal component at
higher frequency, characterized by numerous spikes. The difference between the average values and the nature of the
fluctuating component of the first POD coefficient in the I and II time windows helps us to interpret the role of this first
mode in the dynamics of the system: the first POD shows a distinct switching (on/off switch) character that corresponds
to the transition between the washing machine and standing wave regimes depicted in Fig. 7. This first result highlights
the information content of sloshing fluid images. In other words, the light intensity passing through the gas and the liquid
phases results in a passive scalar, which is convected by the flow field and resolved in time by using the camera.
The analysis of this two-dimensional scalar field, although it refers to a three-dimensional dynamical system, is
nevertheless promising, and as we will see hereafter, the quantitative information extracted is in perfect agreement
with that obtained through the use of accelerometers, but more useful for interpreting the physics associated with the
phenomenon. We emphasize this aspect because POD of the sloshing flow images are not very common in the literature
despite appearing to be a very effective and promising approach.
The POD time coefficients of modes 2–4 are depicted in Fig. 13c. As for the first mode, they present a clear transition
between 1 to 1.5 s, however they differ from the first mode, as they transition from an uncorrelated signal to a clearly
13
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 12. Singular value spectrum of POD modes identified from the image sequence of the complete experiment time-history.

Fig. 13. First four POD time coefficients reported in normalized form: first POD coefficient (a) and its enlargement in the time window [2 2.5] s (b);
second (red dashed line), third (black dot-dashed line) and fourth (green dotted line) POD coefficient.

oscillatory behavior. With respect to the information content of the time coefficients, this topic will be taken up in the
in last section of the manuscript. In particular, we will try to understand what kind of correlation there is between the
oscillations of these coefficients and the vibration modes of the beam.
However, since the oscillatory content of the POD coefficients is very evident and since we are also interested in their
damping, a DMD of the image sequence has been done. For the sake of clarity, since the DMD is not well suited to analyze
the initial washing machine regime, the DMD are applied only for the second part of the time series (between 1.7 and 3 s),
where the POD coefficients indicate an oscillatory behavior of the flow field.
Hereafter, the results of POD and DMD are compared and analyzed in details in Figs. 14–17. Similarities between the
two POD and DMD modes are identified using the square of their normalized scalar product nij defined as:
( T
)2
φ POD
i φ DMD
j
nij = ( )( ) (22)
T T
φ POD
i φ POD
i φ DMD
j φ DMD
j

where the indexes i and j indicate mode numbers. The calculated nij values are shown in Fig. 14, where values close to
unit (and therefore similar shapes) are indicated by the dark squares. The POD and DMD modes that result to be nearly
parallel have been represented in Fig. 15 on the same row. It is worth to noting a very good agreement between the mode
shapes identified using the two techniques. We identified two modes whose spacial coefficients are almost identical to
14
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 14. Square of the normalized scalar product nij of DMD and POD mode shapes for the complete time series.

Table 3
Frequency and damping coefficients calculated for DMD modes.
Mode fn [Hz] ζn [–]
0.5f1 3.92 1.24%
f1 7.89 1.40%

those of modes two and four shown in Fig. 15, and therefore are not reproduced for brevity. The corresponding time
coefficients are presented Figs. 16 and 17, together with their respective Fourier spectra calculated via FFT. Similarly to
what was done for the full time series, the POD histories are paired with the DMD coefficients using the square of the
normalized scalar product. We note that both time and frequency characteristics of the modes are similar, with the DMD
presenting a single frequency component by construction. The mode frequencies and damping coefficients (calculated as
the ratio of the real over imaginary part of the eigenvalue) are listed in Table 3.
We propose an interpretation of the POD/DMD results based on the linear parametric sloshing theory developed
in Benjamin and Ursell (1954), where the movement of the fluid free surface in a tank subject vertical period motion
is represented as linear superposition of an infinite series of spacial mode shapes Sl,m , given by:
lπ x ( mπ z )
( )
Sl,m = cos cos (23)
L M
where L and M are the width and depth of the tank; and l, m are positive integers. The time coefficients of these are the
solution of an infinite series of independent ordinary differential equations of the Mathieu’s type:
∞ [ 2 ]
∑ d al , m ( )
+ pl,m − 2ql,m cos 2T al,m = 0 (24)
dT 2
1

where the dimensionless parameters for time T , frequency pl,m and excitation amplitude ql,m are given by:
1 (ω )2 |ÿ|i ( ωl,m )2
l,m
T = ωt , p(l, m) = 4 , q(l, m) = 2 (25)
2 ω g ω
with ω and ωl,m the excitation and natural sloshing frequencies respectively, expressed in radiants per second. The latter
is calculated as:
)] 1
ωl,m = gkl,m tanh hkl,m 2
[ (
(26)

where h is liquid height and kl,m is given by:

) 12
l2 m2
(
kl,m = π + (27)
a2 b2
In the calculation of the sloshing frequency we neglect the effect of surface tension, estimated to be less than 1%. The
stability of the solution of the Mathieu’s equations can be analyzed in the q - p plane (see Eq. (25)), presented in Fig. 18a.
15
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 15. First four modes extracted from the image dynamics of the cavity number four for a cavity filling factor equal to 0.5 by applying POD
(a,c,e,g) and DMD (b,d,f,h). Red and blue colors are used to represent positive and negative value of the mode respectively. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 16. Normalized time coefficients (a) and Fourier spectra (b) of the second POD mode (red dashed line) and the corresponding DMD mode (red
solid line) for the second half of the time series - 1.7 to 3 s circa.

16
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 17. Normalized time coefficients (a) and Fourier spectra (b) of the fourth POD mode (green dotted line) and the corresponding DMD mode
(green solid line) for the second half of the time series - 1.7 to 3 s circa.

When the parameter combination lays with the Half Harmonic or Harmonic instability regions, the amplitude of the
time coefficients grows in time and parametric oscillation of the free surface can be observed. Experimentally the wave
amplitude is limited by viscous dissipation and non-linear effects which are not modeled by Eq. (24) (Abramson, 1966).
The position on the q - p plane of mode shapes S2,0 and S2,1 are also shown in Fig. 18, assuming an excitation frequency
f1 of 7.4 [Hz] as previously derived from the wavelet analysis of the accelerometer signal (Fig. 9 and Table 2). Both modes
lay within the Half Harmonic region, and hence the solution of Mathieu’s equation predicts a wave oscillation at 3.7 [Hz]
(or half of the excitation frequency).
For mode S2,0 , this frequency value compares well with the POD and DMD results for mode two, with a relative
difference of ≈6% when compared with f1 in Table 3. We attribute this difference to non-linear stiffening of the wave
response (Abramson, 1966) and reserve a more in-depth analysis for future work. Also the spacial mode shape, shown
in the inset in Fig. 18b compares well with POD and DMD spacial coefficients in Fig. 15c and d, seen as the projection of
the wave height onto the x-y plane. Similarly the shape of mode S2,1 as per inset in Fig. 18c is similar to Fig. 15g and h,
however for this mode the frequency identified via the DMD is not consistent with the half harmonic response expected
for a Faraday wave, but rather close to the excitation frequency due to the oscillation of the cantilever. We highlight
that the spectral content of the POD coefficient signal does include a component at ≈ 4 [Hz] as seen in Fig. 17b, which
seems to indicate a frequency modulation of the main component. The final remarks in this section relate to third mode
in Fig. 15e and f. The spectral components of this mode are shown Figs. 19 and 20.
In Fig. 19b and c, it is noticeable the presence of energy at all frequencies in the time interval 0 < t < 1.5 s, i.e., during
the regime called washing machine. This broadband energy distribution is typical in spatially and temporally incoherent
phenomena. The result indicates a loss of temporal coherence related to the loss of spatial coherence of the fluid dynamic
field during this regime. From the spectrograms we notice two frequencies, which corresponds to those of mode two and
four, however it was not possible to identify this mode from the POD and DMD analysis of the second half of the time
series.

4.5. Wavelet and cross-wavelet comparison between POD coefficients and acceleration time series

In Fig. 19 are represented the wavelet transform of the time coefficients referring to the first four POD modes. The POD
coefficient of the first mode, represented in Fig. 19a, exhibits an energetic band, slightly discontinuous, at frequencies
named 1/2f1 . Below this band, at a frequency in the range of 1 Hz, in a time window ranging from 0 to 1.5 s, an energy
band related to the nature of the first POD mode can be observed which represents a switching between ‘‘washing machine
regime’’ and ‘‘standing wave regime’’. At frequency equal to twice the first sloshing mode we have a synchronous energetic
band (f 1) with the first mode of beam. It is worth to noting that this band is common to both physics: structural and
hydrodynamics. For this reason, the component can be interpreted as the linear response of the fluid to the periodic
forcing due to the oscillations of the beam in its first longitudinal mode. A last band, with lower energy than the others,
appears at high frequency (see 2f2 ) and only in the low vibrational energy time window. The latter frequency is exactly
twice the frequency of the beam’s second structural mode. Wavelet analysis and its representation is very useful for
understanding the temporal behavior of spectral components. Also, the wavelet spectra obtained by temporal integration
of the coefficients are a good visualization method to compare which tonal components exist in the POD coefficients
and in the accelerometer time series as well. The wavelet spectra referring to the accelerometer signals and the POD
coefficients are shown in Fig. 20a and b, respectively. In Fig. 20a we can see, as already mentioned, that there are three
components corresponding to the three peaks labeled f1 , 2f1 and f2 . The component f1 is dominant while its harmonic
17
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 18. Stability regions for the Mathieu’s equation solution (black solid lines). The values assumed by p,q for mode S2,0 at f = 3.9 Hz and S2,1 at
f = 4.1 Hz are indicated with the red circle and blue triangle respectively (a); representation of the mode S2,0 at f = 3.9 Hz (b) and the mode S2,1
at f = 4.1 Hz.

has a much smaller amplitude as expected. The peak f2 on the other hand is related to the second mode of the beam.
Observing Fig. 20b, related to the sloshing modes, we have a clear information about the sloshing dynamics:

1. the first sloshing mode 1/2f1 oscillates at around half that of the first longitudinal mode of the beam, this is the
linear part of the fluid response to parametric excitation (see Eq. (24)), induced by the first structural mode of the
cantilever;
2. the other sloshing modes exhibit harmonic content at super-harmonics of 1/2f1 , namely f 1 and 3/2f1 possibly
indicating non-linear response of the fluid to the large amplitude motion of the cantilever;
3. The frequency content of the first sloshing mode at f2 indicates a linear sloshing response to the second structural
mode of the beam. Note that the response is not sub-harmonic in this case, because the shape of the structural
mode (as shown in Fig. 6(b)) is such that the tank rotates more than translates vertically;
4. Similarly to the first sloshing mode, the presence of higher order harmonics at 2f2 for the other sloshing modes
seem to indicate a non-linear sloshing response to the structural excitation induced by the second beam mode.
We summarize the findings of the frequency analysis of the accelerometer data and POD time coefficients in Table 4 where
we also indicate the patterns identified for high and low amplitude oscillation via wavelet analysis. The corresponding
figures are reported in the table together with the coupling pattern specified in the FSI column.
18
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 19. Wavelet transform of the first four time coefficients extracted from the POD decomposition of the image sequence: first mode (a); second
mode (b); third mode (c) and fourth mode (d).

Table 4
Summary of the findings of the frequency analysis of the accelerometer data and POD time coefficients
and the coupling patterns.
High amplitude Low amplitude FSI
3.7 Hz Fig. 20a → 4.0 Hz Fig. 20a–b–c 0.5f1
7.9 Hz Fig. 20a → 7.9 Hz Fig. 20a–c–d 1.0f1 *
POD
– – 12.0 Hz Fig. 20b 1.5f1
– – 102.0 Hz Fig. 20a–b 2.0f2 **

7.4 Hz Fig. 9b → 1.0f1

7.9 Hz Fig. 9b *
Acc. 13.9 Hz Fig. 9b ↗ 2.0f1
50.2 Hz Fig. 9b → 50.2 Hz Fig. 9b 1.0f2 **

Indeed, in order to better understand which sloshing modes are coupled with structural dynamics and for which time
windows, the wavelet coherency can clarify the nature of the coupling and addresses this issue. In Fig. 21a it can be
seen that beam vibrations and first POD mode are coupled all the time in a frequency range between f1 and 2f1 . Instead,
in Fig. 21b we see that there is almost no coupling between the second POD mode and the beam vibrations, i.e. the
second mode is a hydrodynamic only mode. In Fig. 21c and d we see that there is a significant coupling between the
19
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 20. Semi-log scale spectra of the acceleration time history: spectra is marked with blue solid line for the baseline, red dashed line for the test
case 1, and black dash-dotted line for the test case 2 (a). Log–log scale spectra of the first four POD time coefficients. Blue solid line is used to
represents the first coefficient, red dashed line is for the second one, black dash-dotted line for the third coefficient and in green dotted line is
plotted the spectra of the fourth coefficient. All main frequency components have been marked with an indicator. Furthermore, the spectra have
been computed by a temporal integration of the wavelet coefficients (b).

third and fourth POD modes and the beam vibrations only after the first second, i.e. during the phase of formation of the
standing waves. Finally, in order to further investigate the coupling between structural dynamics and sloshing POD modes,
the phase of the wavelet cross-spectrum was represented for modes 1, 3 and 4 (the POD modes for which the wavelet
coherency has pointed out a coupling with structural dynamics). Observing the phase of the first mode with respect
to the acceleration it is noticeable that the first mode is in advance with respect to the acceleration: more specifically
in quadrature. Furthermore, after 1.75 s the phase changes of sign and the first mode is in delay with respect to the
acceleration and then asymptotically reaches the phase opposition. It is very interesting to note that for modes 3 and 4
after about 1 and 1.5 s respectively there is a phase opposition between the structural dynamics and the sloshing POD
modes, while the first mode is affected by this phase opposition only in the last second. It can be concluded that the most
energetic sloshing modes, which are linearly coupled with the beam modes, are in phase opposition during the so-called
standing wave regime. Since we can define the modes less than a constant, the information obtained from this analysis
is that the phase during this regime is constant. Finally, the switching between π and −π , on the other hand, should not
be ascribed to any physical phenomenon but only to a mathematical issue known as phase wrap (see Fig. 22).

5. Conclusions

In this paper, we present POD, DMD and wavelet analyses of sloshing flows and structural dynamics based on image
sequences and acceleration time series. Specifically, we look at the free transient response and fluid–structure interaction
experienced by a cantilever beam with partially filled tank, representative of an aircraft wing. From the experimental
results we showed that internal hydrodynamics in aircraft fuel tanks has a significant effect on the dynamics of the
20
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Fig. 21. Wavelet coherency map computed for acceleration time history and first (a), second (b), third (c) and fourth (d) POD coefficient. A dominant
band appear bounded f1′ and f1′′ is highlighted via two white dotted line.

Fig. 22. Phase of the cross-wavelet spectrum of the acceleration and the first (blue solid line), third (red dashed line) and fourth (black dash-dotted
line) POD coefficient calculated at central frequency, (f1 + 2f1 )/2, of the dominant band in the coherency map.

wing structure, and has the potential to be exploited for optimal design. We demonstrated the use of POD, DMD and
wavelet transform on video images and sensor data to discover several characteristics of such a coupled fluid/structure
interaction problem. From the POD and DMD we are able to decompose the sloshing flow field in spacial and temporal
coherent structures, and compare the wavelet transforms of these temporal coefficients and accelerometer signal. The
analysis highlighted a clear transition between an initial incoherent ‘‘washing machine’’ regime to a more orderly one,
21
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

once the incremental acceleration level decays below 1g. Also, we identified several mechanisms of interaction between
the structural forcing and the fluid flow, in particular a half harmonic response typical of Faraday’s waves, accompanied
by harmonic and 3/2 harmonic components.
An interesting outcome of the analysis performed with both POD and DMD decomposition, is their ability to capture
the frequencies and damping of free surface oscillations directly from flow images, with no additional processing required.
Conversely, we recognized that the ‘‘washing machine’’ regime has a distinctively incoherent nature and it is therefore
less suitable for periodic analysis.

CRediT authorship contribution statement

Tiziano Pagliaroli: Conceptualization, Methodology, Software, Formal analysis, Investigation, Writing – original draft,
Writing – review & editing. Francesco Gambioli: Conceptualization, Software, Formal analysis, Investigation, Writing –
original draft, Project administration, Writing – review & editing, Funding acquisition. Francesco Saltari: Software, Formal
analysis, Investigation, Writing – original draft, Writing - review & editing. Jonathan Cooper: Writing – original draft,
Writing – review & editing.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

Acknowledgments

The authors acknowledge the support of Lewis Lea and Ben Rivington of Mathworks, for their support in the image
®
processing with the Matlab . We thank Leo Gonzalez and Franco Mastroddi for the lively discussions about the POD
and DMD algorithms. The SLOWD project has received funding from the European Union’s Horizon 2020 research and
innovation programme under grant agreement No 815044.

References

Abramson, H.N., 1966. The dynamic behavior of liquids in moving containers, with applications to space vehicle technology. Natl. Aeronaut. Sp. Adm.
464.
Ashmead, J., 2012. Morlet wavelets in quantum mechanics. Quanta 1 (1), 58–70.
Behruzi, P., Michaelis, M., Khimeche, G., 2006. Behavior of the cryogenic propellent tanks during the first flight of the ariane 5 ESC-A upper stage.
In: Collect. Tech. Pap. - AIAA/ASME/SAE/ASEE 42nd Jt. Propuls. Conf., Vol. 9. pp. 7077–7086, no. July.
Benjamin, T.B., Ursell, F.J., 1954. The stability of the plane free surface of a liquid in vertical periodic motion. Proc. R. Soc. A 225 (1163), 505–515.
Berkooz, G., Holmes, P., Lumley, J., 1993. The proper orthogonal decomposition in the analysis of turbulent flows. Annu. Rev. Fluid Mech. 250,
539–575.
Bouscasse, B., Colagrossi, A., Souto-Iglesias, A., Cercos-Pita, J.L., 2014a. Mechanical energy dissipation induced by sloshing and wave breaking in a
fully coupled angular motion system. I. Theoretical formulation and numerical investigation. Phys. Fluids 26 (3), 033103.
Bouscasse, B., Colagrossi, A., Souto-Iglesias, A., Cercos-Pita, J.L., 2014b. Mechanical energy dissipation induced by sloshing and wave breaking in a
fully coupled angular motion system. II. Experimental investigation. Phys. Fluids 26 (3), 033104.
Brosset, L., Mravak, Z., Kaminski, M., Collins, S., Finnigan, T., 2009. Overview of Sloshel project. In: Proc. Int. Offshore Polar Eng. Conf., no. 1. pp.
115–124.
Chen, S.-L., Liu, J.-J., Lai, H.-C., 2009. Wavelet analysis for identification of damping ratios and natural frequencies. J. Sound Vib. 323 (1–2), 130–147.
Constantin, L., De Courcy, J., Titurus, B., Rendall, T., Cooper, J., 2021. Analysis of damping from vertical sloshing in a SDOF system. Mech. Syst. Signal
Process. 152, 107452.
Delorme, L., Colagrossi, A., Souto-Iglesias, A., Zamora-Rodríguez, R., Botía-Vera, E., 2009. A set of canonical problems in sloshing, Part I: Pressure field
in forced roll-comparison between experimental results and SPH. Ocean Eng. 36 (2), 168–178.
Dittrich, W., Reuter, M., 2017. Linear oscillator with time-dependent frequency. In: Classical and Quantum Dynamics. Springer, pp. 259–274.
Dodge, F.T., 2009. Series page. p. iii.
Farge, M., 1992. Wavelet transforms and their applications to turbulence. Annu. Rev. Fluid Mech. 24 (1), 395–458.
Fonte, F., Ricci, S., Mantegazza, P., 2017. Wind tunnel evaluation of a static output feedback controller for gust load alleviation on a regional aircraft.
In: 17th Int. Forum Aeroelasticity Struct. Dyn. IFASD 2017, Vol. 2017-June. pp. 1–14, no. June.
Gambioli, F., Chamos, A., Jones, S., Guthrie, P., 2020. Sloshing wing dynamics -project overview sloshing wing dynamics – project overview. no. April.
Gambioli, F., Malan, A.G., 2017. Fuel loads in large civil airplanes. In: 17th Int. Forum Aeroelasticity Struct. Dyn. IFASD 2017, Vol. 2017-June. pp.
1–20, no. June.
Gambioli, F., Usach, R.A., Wilson, T., Behruzi, P., 2019. Experimental evaluation of fuel sloshing effects on wing experimental evaluation of fuel
sloshing effects on. pp. 1–14, no. June.
Gawad, A.F., Ragab, S.A., Nayfeh, A.H., Mook, D.T., 2001. Roll stabilization by anti-roll passive tanks. Ocean Eng. 28 (5), 457–469.
Hall, J., Rendall, T., Allen, C., Peel, H., 2015. A multi-physics computational model of fuel sloshing effects on aeroelastic behaviour. J. Fluids Struct.
56, 11–32.
Ibrahim, R.A., 2020. Assessment of breaking waves and liquid sloshing impact. Nonlinear Dynam. 100 (3), 1837–1925.
Kareem, A., Kijewski, T., Tamura, Y., 1999. Mitigation of motions of tall buildings with specific examples of recent applications. Wind Struct. 2 (3),
201–251.
Konopka, M., De Rose, F., Strauch, H., Jetzschmann, C., Darkow, N., Gerstmann, J., 2019. Active slosh control and damping-Simulation and experiment.
Acta Astronaut. 158, 89–102.

22
T. Pagliaroli, F. Gambioli, F. Saltari et al. Journal of Fluids and Structures 112 (2022) 103603

Lewis, D.J., Proc R. Soc. Lond. A, 1950. The instability of liquid surfaces when accelerated in a direction perpendicular to their planes. II. Proc. R. Soc.
A 202 (1068), 81–96.
Liu, P.C., 1994. Wavelet spectrum analysis and ocean wind waves. In: Foufoula-Georgiou, E., Kumar, P. (Eds.), Wavelets in Geophysics. In: Wavelet
Analysis and its Applications, vol. 4, Academic Press, pp. 151–166.
Mallat, S., 2009. A Wavelet Tour of Signal Processing, third ed. Academic press.
Mancinelli, M., Pagliaroli, T., Di Marco, A., Camussi, R., Castelain, T., 2017. Wavelet decomposition of hydrodynamic and acoustic pressures in the
near field of the jet. J. Fluid Mech. 813, 716–749.
Mancinelli, M., Pagliaroli, T., Di Marco, A., Camussi, R., Castelain, T., Leon, O., 2016. Hydrodynamic and acoustic wavelet-based separation of the
near-field pressure of a compressible jet. In: 22nd AIAA/CEAS Aeroacoustics Conference. p. 2864.
Maurel, S., Borée, J., Lumley, J., 2001. Extended proper orthogonal decomposition: application to jet/vortex interaction. Flow Turbul. Combust. 67,
125–136.
Meyer, K.E., Pedersen, J.M., Oktayøzcan, 2007. A turbulent jet in crossflow analysed with proper orthogonal decomposition. J. Fluid Mech. 583,
199–227.
Moya, B., González, D., Alfaro, I., Chinesta, F., CUETO, E.G., 2019. Learning slosh dynamics by means of data. Comput. Mech. 64 (2), 511–523.
Pagliaroli, T., Camussi, R., Giacomazzi, E., Giulietti, E., 2015. Velocity measurement of particles ejected from a small-size solid rocket motor. J. Propuls.
Power 31 (6), 1777–1784.
Pagliaroli, T., Mancinelli, M., Troiani, G., Iemma, U., Camussi, R., 2018a. Fourier and wavelet analyses of intermittent and resonant pressure components
in a slot burner. J. Sound Vib. 413, 205–224.
Pagliaroli, T., Pagliaro, A., Patanè, F., Tatì, A., Lv, P., 2020. Wavelet analysis ultra-thin metasurface for hypersonic flow control. Appl. Acoust. 157,
107032.
Pagliaroli, T., Patanè, F., Pagliaro, A., Lv, P., Tatí, A., 2018b. Metamaterials for hypersonic flow control: Experimental tests on novel ultrasonically
absorptive coatings. In: 2018 5th IEEE International Workshop on Metrology for AeroSpace (MetroAeroSpace). pp. 284–289.
Saatci, E., Yazıcı, G., 2019. Sloshing displacement measurements based on morphological image analysis. Acta Phys. Pol. A 135 (5), 949–954.
Sieber, M., Paschereit, C., Oberleithner, K., 2016. Spectral proper orthogonal decomposition. J. Fluid Mech. 792, 798–828.
Simonini, A., Colinet, P., Vetrano, M.R., 2015. Reference Image Topography technique applied to harmonic sloshing. 1 (1).
Simonini, A., Peveroni, L., Vetrano, M.R., 2016. Experimenatl characterisationof LN2 sloshing by means of non-intrusive opticaol techniques. pp. 2–3,
no. August.
Simonini, A., Vetrano, M.R., Colinet, P., Rambaud, P., 2014. Particle image velocimetry applied to water sloshing due to a harmonic external excitation.
pp. 7–10.
Staszewski, W.J., Cooper, J.E., 2002. Wavelet approach to flutter data analysis. J. Aircr. 39 (1), 125–132.
Stefanutti, E., Bruni, F., 2017. Signal denoising using the stationary wavelet decomposition. In: IMEKO TC19 Workshop on Metrology for the Sea,
MetroSea 2017: Learning to Measure Sea Health Parameters, Vol. 2017-October. pp. 104–110.
Stodieck, O., Cooper, J.E., Weaver, P.M., Kealy, P., 2017. Aeroelastic tailoring of a representative wing box using tow-steered composites. AIAA J. 55
(4), 1425–1439.
Titurus, B., Cooper, J.E., Saltari, F., Mastroddi, F., Gambioli, F., 2019. Analysis of a sloshing beam experiment. In: International Forum on Aeroelasticity
and Structural Dynamics. Savannah, Georgia, USA, Paper, Vol. 139.
Torrence, C., Compo, G.P., 1998. A practical guide to wavelet analysis. Bull. Am. Meteorol. Soc. 79 (1), 61–78.
Torrence, C., Webster, P.J., 1999. Interdecadal changes in the ENSO–monsoon system. J. Clim. 12 (8), 2679–2690.
Tosun, U., Aghazadeh, R., Sert, C., Özer, M.B., 2017. Tracking free surface and estimating sloshing force using image processing. Exp. Therm. Fluid
Sci. 88, 423–433.
Tu, J.H., Rowley, C.W., Luchtenburg, D.M., Brunton, S.L., Kutz, J.N., 2014. On dynamic mode decomposition: Theory and applications. J. Comput. Dyn.
1 (2), 391–421.
Visser, M., 2003. Physical wavelets: Lorentz covariant, singularity-free, finite energy, zero action, localized solutions to the wave equation. Phys. Lett.
A 315 (3–4), 219–224.
Williams, M.O., Shlizerman, E., Wilkening, J., Kutz, J.N., 2012. The low dimensionality of time-periodic standing waves in water of finite and infinite
depth. SIAM J. Appl. Dyn. Syst. 11 (3), 1033–1061.

23