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

net/publication/371170465

A Data-Driven Approach for Generating Vortex-Shedding Regime Maps for an

Oscillating Cylinder

Article in Energies · May 2023

DOI: 10.3390/en16114440

CITATIONS READS

0 64

8 authors, including:

Fue-Sang Lien Matthew Cann

University of Waterloo University of Waterloo
62 PUBLICATIONS 575 CITATIONS 5 PUBLICATIONS 3 CITATIONS

SEE PROFILE SEE PROFILE

Ryley McConkey W.W. Melek

University of Waterloo University of Waterloo
16 PUBLICATIONS 61 CITATIONS 165 PUBLICATIONS 3,594 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Matthew Cann on 01 June 2023.

The user has requested enhancement of the downloaded file.

 energies
Article
A Data-Driven Approach for Generating Vortex-Shedding
Regime Maps for an Oscillating Cylinder
Matthew Cann, Ryley McConkey, Fue-Sang Lien * , William Melek and Eugene Yee
Citation: Cann, M.; McConkey, R.;
Lien, F.-S.; Melek, W.; Yee, E. A
Data-Driven Approach for
Generating Vortex-Shedding Regime
Maps for an Oscillating Cylinder.
Energies 2023, 16, 4440. https://
doi.org/10.3390/en16114440
Academic Editor: Eugen Rusu
Received: 30 April 2023
Revised: 27 May 2023
Accepted: 28 May 2023
Published: 31 May 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Energies 2023, 16, 4440. https://doi.org/10.3390/en16114440
Department of Mechanical and Mechatronics Engineering, University of Waterloo,
Waterloo, ON N2L 3G1, Canada; matthew.cann@uwaterloo.ca (M.C.); ryley.mcconkey@uwaterloo.ca (R.M.);
william.melek@uwaterloo.ca (W.M.); eugene.yee@uwaterloo.ca (E.Y.)
* Correspondence: fslien@uwaterloo.ca
Abstract: This study presents a data-driven approach for generating vortex-shedding maps, which
are vital for predicting flow structures in vortex-induced vibration (VIV) wind energy extraction
devices, while addressing the computational and complexity limitations of traditional methods. The
approach employs unsupervised clustering techniques on subsequences extracted using the matrix
profile method from local flow measurements in the wake of an oscillating circular cylinder generated
from 2-dimensional computational fluid dynamics simulations of VIV. The proposed clustering
methods were validated by reproducing a benchmark map produced at a low Reynolds number
(Re = 4000) and then extended to a higher Reynolds number (Re = 10,000) to gain insights into the
complex flow regimes. The multi-step clustering methods used density-based and k-Means clustering
for the pre-clustering stage and agglomerative clustering using dynamic time warping (DTW) as the
similarity measure for final clustering. The clustering methods achieved exceptional performance at
high-Reynolds-number flow, with scores in the silhouette index (0.4822 and 0.4694) and Dunn index
(0.3156 and 0.2858) demonstrating the accuracy and versatility of the hybrid clustering methods.
This data-driven approach enables the generation of more accurate and feasible maps for vortex-
shedding applications, which could improve the design and optimization of VIV wind energy
harvesting systems.
Keywords: vortex shedding; machine learning; unsupervised clustering; time series clustering;
vortex-induced vibrations
1. Introduction
Vortex shedding is an aerodynamic phenomenon when fluid flows around a bluff body,
such as a cylinder. Vortex shedding produces unbalanced forces acting on the structure,
which cause vortex-induced vibrations (VIV). Most studies on the consequences of vortex
shedding have focused on lessening the imbalanced forces in applications that are affected
by the phenomenon. VIV arises in many domains, such as the design of skyscrapers,
pipelines [1], offshore structures [2], and bridges [3]. VIV is a subset of flow-induced
vibration (FIV) which is relevant for oscillatory motions in structures, primarily in the
case of flow past a circular cylinder. However, for more complex structures, such as a
square cylinder or a cylinder–plate assembly, the flow-induced motions can be much more
complex, involving not only VIV but also galloping (self-limited or unlimited as a function
of the reduced velocity) and an overlapping form of VIV and galloping [4]. While these
are all important phenomena to consider in the study of fluid–structure interaction, this
paper will focus specifically on VIV as it pertains to flow past a circular cylinder for the
application of VIV energy harvesters. Bladeless wind turbines are a new concept of wind
harvesting machines that utilize VIV to oscillate a vertical cylinder. Instead of reducing
the oscillations, the main principle of bladeless wind turbines is to take advantage of bluff
bodies’ natural phenomenon to extract renewable energy from the motion.
https://www.mdpi.com/journal/energies
Energies 2023, 16, x FOR PEER REVIEW 2 of 32

Energies 2023, 16, 4440 2 of 30

There are two primary methodologies to study vortex-induced vibration for a
mounted circular cylinder in a uniform crossflow, namely, free and forced vibration. A
free-vibration
There areexperiment
two primary ofmethodologies
VIV allows thetocylinder to oscillate due
study vortex-induced to the external
vibration and
for a mounted
unbalanced forcesinproduced
circular cylinder a uniformby the fluid,
crossflow, while free
namely, forced-vibration methodsAprescribe
and forced vibration. the
free-vibration
oscillatory
experiment motion
of VIVof the cylinder
allows the cylinder to approximate
to oscillate due thetofree-vibration
the external and behaviour.
unbalanced Due to
forces
varying
produced mechanisms
by the fluid, of VIV
while between free- and forced-vibration
forced-vibration methods prescribe experiments,
the oscillatoryspecial
motioncon-of
sideration
the cylinder is required to ensure
to approximate theconsistent
free-vibrationfluid behaviour.
forces to represent positivemechanisms
Due to varying energy wakeof
VIV between free- and forced-vibration experiments, special consideration is required to
excitation.
ensure
Morse consistent fluid forces
and Williamson [5]toaddressed
represent the positive
validityenergy wake
of the excitation.
predicting power of forced
vibrationMorse and Williamson
by finding that thorough[5] addressed
matching the validity offrequency,
of amplitude, the predicting power of num-
and Reynolds forced
vibration
ber resulted byinfinding that thorough
consistent matching
fluid forces. of amplitude,
The parameter frequency,
space and Reynoldsexperi-
of forced-vibration number
resulted
ments in consistent
identifies fluid forces. regimes
vortex-shedding The parameter on thespace of forced-vibration
normalized experiments
amplitude–wavelength
plane for the sinusoidal transverse motion of the cylinder. The plane is definedplane
identifies vortex-shedding regimes on the normalized amplitude–wavelength by the for
∗motion of the cylinder. The plane is defined by ∗ the normalized
normalized wavelength, 𝜆 , and the normalized oscillation amplitude, 𝐴 , which are cal-
the sinusoidal transverse
wavelength,
culated λ∗ , and of
for a cylinder theconstant
normalized 𝐷, in amplitude,
oscillation
diameter, crossflow, 𝑈; A∗oscillation
, which arewavelength,
calculated for 𝜆; a
cylinder
and of constant
oscillation diameter,
frequency, D, inEquations
𝑓, using crossflow,(1) U;and
oscillation wavelength, λ; and oscillation
(2) as follows:
frequency, f , using Equations (1) and (2) as follows:
𝐴
𝐴∗ = , (1)
∗ 𝐷A
A = , (1)
D
𝜆 𝑈
𝜆∗ = = . (2)
∗ λ𝐷 𝑓𝐷 U
λ = = . (2)
Morse and Williamson [5] produced a vortex-shedding D fD map for the normalized am-
plitude–wavelength
Morse and Williamson plane by conducting
[5] produced numerous forced-vibration
a vortex-shedding mapexperiments involv-
for the normalized
ing the flow past a circular
amplitude–wavelength planecylinder at a Reynolds
by conducting numerousnumber of 4000. The generation
forced-vibration experiments andin-
shedding
volving the of large coherent
flow past vortex
a circular structures
cylinder occur in distinct
at a Reynolds numbermodes.
of 4000.TheThemost prevalent
generation and
vortex modes are 2S and 2P, which correspond, respectively,
shedding of large coherent vortex structures occur in distinct modes. The most prevalent to two single opposingly
spinning
vortex modesvortices and
are 2Stwo
andpairs of opposing
2P, which correspond,spinning vortices shed
respectively, to per
twooscillating period.
single opposingly
Morse
spinning vortices and two pairs of opposing spinning vortices shed per oscillatingthe
and Williamson [5] discovered a transition mode between the boundary of 2S
period.
and
Morse2P modes named ‘2P [5]
and Williamson Overlap’ or reduced
discovered ‘2PO’. mode
a transition The 2PO mode is
between thedescribed
boundary byof twothe
pairs
2S and of vortices
2P modes being
namedshed‘2P per cycle, with
Overlap’ one vortex
or reduced ‘2PO’. in each
The 2POoscillation
mode is much weaker
described by
and
twointermittently
pairs of vortices switching
being shed between the 2S
per cycle, withandone2P vortex
modes.inFinally, the authors
each oscillation muchoutlined
weaker
a and
P + Sintermittently
mode as an asymmetric
switching betweenform ofthe the2S 2Pandmode 2P described by a pair
modes. Finally, of vortices
the authors and a a
outlined
single
P + S mode as an asymmetric form of the 2P mode described by a pair of vortices runs
vortex shed per oscillating period. The authors conducted 5860 experimental and a
insingle
a water channel
vortex shedtoperobtain high-resolution
oscillating period. The vortex-shedding
authors conducted regimes, as shown in Figure
5860 experimental runs in
1.a water channel to obtain high-resolution vortex-shedding regimes, as shown in Figure 1.

Figure 1. Forced-vibration vortex-shedding map in normalized amplitude–wavelength space [5].

Energies 2023, 16, 4440 3 of 30

The vortex-shedding map is crucial for understanding cylinder kinematics and energy
extraction for VIV generators. It provides information on energy excitation, ensuring the
validity of forced-vibration experiments for free-vibration cases. Morse and Williamson [5]
validated the predicting power of forced vibration by matching amplitude, frequency, and
Reynolds number, which resulted in consistent fluid forces.

1.1. Related Work

Previous approaches to generating vortex-shedding maps have been reliant on gath-
ering substantial amounts of experimental data. The vortex-shedding map shown in
Figure 1 was produced through an expanded investigation that built upon the first am-
plitude and frequency map created by Williamson and Roshko [6]. They conducted
5680 experimental runs in a water channel, producing 500 h of data, using digital par-
ticle image velocimetry (DPIV) to visualize wake regimes with high-resolution images
taken at 10- and 20-millisecond intervals. The determination of various regimes required a
coupled analysis of the flow images and the fluid force measurements by the experimental
apparatus. The authors used abrupt jumps in the fluid force as transitions into different
regimes to define the boundaries of the shedding map. However, more complex transition
states between vortex-shedding modes required investigation of the vortex phase. Generat-
ing the vortex-shedding map necessitated large amounts of data, including fluid forces and
their decomposition into vortex force and potential force, alongside reference to fluid flow
images of the wake. Visual inspection by an expert was necessary to manually identify
spatiotemporal patterns, but this has limitations for more complex flow regimes associated
with higher Reynolds numbers.
The previous studies are the extent of research in generating vortex-shedding maps
and are limited to a low Reynolds number of less than 4000. The lack of investigations in
generalizing the vortex-shedding behaviour in the parameter space at a higher Reynolds
number may be attributed to the more complex dynamics. Wu et al. [7] studied vortex-
shedding patterns from Reynolds number 35,000 to 130,000 and identified several relatively
stable modes, including 2S, 2P, 2PO, P + S, and 2P + 4S. Due to instability and variations
between cycles, the authors noted the inability to generalize some vortex-shedding patterns
from Reynolds number 60,000 to 75,000. Zhang et al. [8] echoed this limitation at the highest
free-stream turbulence intensity (5%); the vortex structure was indistinguishable in the
flow field due to the large amounts of mixing from the increased dissipation energy. The
authors found that the turbulence intensity of the incoming stream has a dissipation effect
that causes the vortices to become weaker and increasingly difficult to distinguish.
VIV energy harvesters are expected to perform optimally in ambient airflow condi-
tions characterized by Reynolds numbers that are greater than 4000. For instance, a mean
wind velocity of 5 m/s corresponds to a Reynolds number of 10,000 [9]. The advantage of
operating VIV machines in higher-Reynolds-number flows lies in the increased efficiency
and energy harvesting potential resulting from increases in lift force, VIV oscillation ampli-
tude, and synchronization range [9]. Although a vortex-shedding map at higher Reynolds
numbers is beneficial, generating such a map is challenging due to high computational
costs and complex dynamics, with few demonstrated methods available.
The traditional method of generating vortex-shedding maps requires large amounts
of data in the form of fluid forces, resolved flow images, and visual inspection from an ex-
pert [10–14]. Data-driven methods for identifying and clustering distinct flow regimes from
high-dimensional, time-resolved flow fields provide a versatile and automated approach to
improve confidence intervals while reducing input information. Recently, in the literature,
the benefits of the data-driven method of clustering on high-dimensional time-resolved
flow fields have been gaining attention, specifically for identifying and grouping distinct
flow dynamics of vortex shedding from VIV.
In the progression of research within this stream, our earlier publication Cann et al. [15]
served as a pivotal proof of concept for validating the efficacy of the proposed clustering
analysis. The paper presents an effective method of classifying 2S, 2P, and 2PO vortex
Energies 2023, 16, 4440 4 of 30

structures in the wake of an oscillating circular cylinder, employing various machine-

learning algorithms trained on simulated fluid flow data. An exceptional classification
accuracy of 99.8% was achieved using a random forest model trained on the y-component
of velocity. The impressive classification results highlight the notable feature separation
attainable, which is crucial for the successful unsupervised clustering of time series data.
The demonstrated ability to accurately identify global vortex structures in the wake of an
oscillating circular cylinder using local flow signatures provides a solid foundation for the
application of the proposed clustering methods in this paper.
Clustering is a subset of unsupervised machine learning when the input data is
assumed to have discrete structures that can be extracted. Unsupervised learning uses
the input data structure to provide insights since there are no expert labels associated
with the data. Huera-Huarte and Vernet [16] applied fuzzy clustering on digital particle
image velocimetry (DPIV) data for the vortex shedding of a flexible cylinder. Proper
orthogonal decomposition (POD) was used to reduce the dimensionality of the image
data to identify patterns. The method adequately identified the 2S and 2P modes at
two elevations along the long flexible cylinder exposed to crossflow of Reynolds numbers
from 1200 to 12,000. The work by Huera-Huarte and Vernet [16] provides a pertinent
application of clustering on vortex-shedding structures for an oscillating cylinder. However,
the clustering analysis was only utilized to identify the 2S and 2P modes for the free-
vibration experiment. The parameter space of amplitude and frequency was not explored
since the experiment considered the free vibration of the cylinder. Furthermore, the entire
vorticity flow field was required for clustering, despite the use of POD to reduce the
data dimensions.
Menon and Mittal [17] studied pitching airfoil wake dynamics using clustering meth-
ods to identify and track vortices. The main objective of this study was to analyze the
force production on the airfoil from local vortical regions using the force and moment
partitioning method (FMPM). The proper application of this method requires the vortex
regions near the airfoil to be accurately isolated. The authors used the clustering analysis
to isolate and track the vortices primarily in the leading and trailing edge due to their
overall effects on pitching airfoils. The authors used the DBSCAN clustering algorithm for
the analysis based on its advantages for flow field clustering. The dataset comprises the
vorticity fields from forced-vibration CFD simulations of an oscillating airfoil. The pitching
frequency and amplitude parameter space were sampled extensively for the varying vortex
dynamics. The authors demonstrated the utility of this methodology for the investigation
of the dynamic influence of vortex-induced forces.
Another effort directed at clustering vortex-shedding modes was conducted by Calvet
et al. [18] to cluster the vortex wakes of bio-inspired propulsors. The authors generated
the flow fields behind an oscillating airfoil using computational fluid dynamics (CFD) at
a Reynolds number of 106 . The vorticity flow field was reduced with a deep convolu-
tional autoencoder and clustered using k-Means++. The authors selected the k-Means++
algorithm after comparing the results of k-Medoids, hierarchical clustering, and DBSCAN,
which all showed no improvement over k-Means. The combination of feature extraction
and clustering based on the latent space produced an exceptional method for identifying
wake kinematics. The clustering method achieved 100% accuracy in extracting airfoil
wake patterns when trained on a simple one-degree-of-freedom labelled dataset, which
included 2P + 4S, 2P + 2S, and 2S modes. The trained autoencoder was then applied to
an unlabelled dataset of a two-degree-of-freedom oscillating airfoil with more complex
modes. The optimal number of clusters for this case was determined to be six through the
elbow method, silhouette index, and visual classification of the modes. The success of the
authors provides a basis for the methodology used in this study, which shares similarities
of developing a clustering technique for a straightforward problem of vortex shedding and
extrapolating to more complex problems. However, the authors highlight the limitations
of the investigation, which are linked to the input data used to train the autoencoder
and clustering algorithm. Higher-amplitude vortex modes may be under-represented in
Energies 2023, 16, 4440 5 of 30

the training set, potentially corrupting the results. The wake width significantly affected
the clustering algorithm, indicating the need to train it to cluster based on overall vortex
pattern parameters.

1.2. Objectives
In reviewing the literature, it was observed that the traditional method used for
generating vortex-shedding mode maps requires large amounts of data and intensive
supervision. An opportunity was identified to implement a versatile and automated data-
driven approach that addresses these limitations. Furthermore, there is a lack of proposed
methods for generating the vortex-shedding mode map at higher Reynolds numbers due
to the increased computational cost and complex dynamics.
The main objective of this study was to develop a data-driven approach for generating
vortex-shedding maps for a cylinder under forced vibration. This primary objective was
discretized into two manageable goals, namely,
1. Develop data-driven methods and validate their performance in the vortex-shedding
map using a reference case at a low Reynolds number;
2. Extend the method from Objective 1 to generate a vortex-shedding map at a higher-
Reynolds-number case.
In this paper, Objective 1 was achieved by developing an unsupervised clustering
method to extract a small number of well-defined clusters that are rooted in the flow
physics of the low-Reynolds-number case to reproduce the benchmark vortex-shedding
map. The final objective was met by generating vortex-shedding maps using the methods
from Objective 1 for higher Reynolds numbers and discussing the insights gained on the
underlying dynamical regimes of the physical system.

1.3. Contributions and Perspectives of This Work

The contribution of our study is twofold. First, we present a novel approach for gen-
erating vortex-shedding maps that is data-driven, automated, and accurate compared to
traditional methods. Second, we demonstrate the effectiveness of our approach in handling
high-Reynolds-number flow, which is a significant contribution to the field. The effort of
this study deviates from previous work concerning vortex-shedding map generation by
demonstrating the viability of data-driven unsupervised subsequence clustering techniques
to identify vortex-shedding modes from local flow signatures sampled in the wake. Previ-
ous unsupervised clustering approaches have shown promising results for identification
and dissection of vortex-shedding modes. However, these methods require extensive input
data to compute the entire flow field. Furthermore, the primary contribution of this method
is its use in generating vortex-shedding maps for high Reynolds numbers, which have been
limited in the literature due to the increased computational cost and complex dynamics.
The data-driven approach for generating vortex-shedding maps presented in this study
has the potential for several future engineering applications. One potential application is
in the design and optimization of VIV wind energy harvesting systems, where accurate
predictions of flow structures are essential for maximizing energy output. The current low-
Reynolds-number vortex-shedding maps have limited utility in designing VIV harvesting
systems that operate at higher flow speeds. By using this method, we are able to generate
vortex-shedding maps for a much wider range of Reynolds numbers, which provides
critical insights into the flow dynamics and aids in the design and optimization of VIV
systems. The approach can also be applied to other engineering applications, such as heat
exchangers, aerospace vehicles, and marine engineering, where vortex-induced vibrations
can cause structural damage and impact the overall performance of the system. Moreover,
the unsupervised clustering techniques employed in our approach can be extended to other
fluid mechanics problems where flow structures play a significant role, such as turbulence
and fluid–structure interaction.
The framework developed in this paper presents a refreshing alternative to conven-
tional approaches for elucidating wake dynamics in the flow past a structure. This approach
 extended to other fluid mechanics problems where flow structures play a significant role,
such as turbulence and fluid–structure interaction.
Energies 2023, 16, 4440
The framework developed in this paper presents a refreshing alternative to conven- 6 of 30
tional approaches for elucidating wake dynamics in the flow past a structure. This ap-
proach could lead to new and innovative designs of a new generation of energy harvesters
that can exploit all forms of flow-induced vibration for energy extraction. While the meth-
could
odologyleadistotested
new and innovative
in the paper fordesigns of arestricted
the more new generation
case ofof
VIVenergy harvesters
of a circular that canit
cylinder,
exploit all forms of flow-induced vibration for energy extraction. While the methodology
can be applied to a more general case of FIV in more general structures. For example, even
is tested in the paper for the more restricted case of VIV of a circular cylinder, it can be
for the cylinder–plate assembly in laminar flow (at low Reynolds numbers), the vortex-
applied to a more general case of FIV in more general structures. For example, even for the
shedding map is much more complex than that produced by Morse and Williamson [5]
cylinder–plate assembly in laminar flow (at low Reynolds numbers), the vortex-shedding
for a circular cylinder at Re = 4000. Therefore, the framework developed in this study
map is much more complex than that produced by Morse and Williamson [5] for a circular
could provide a powerful tool for exploring and understanding the complex flow dynam-
cylinder at Re = 4000. Therefore, the framework developed in this study could provide
ics in various engineering systems, paving the way for more efficient and effective designs.
a powerful tool for exploring and understanding the complex flow dynamics in various
engineering systems, paving the way for more efficient and effective designs.
1.4. Paper Structure
The structure
1.4. Paper Structure of this paper is as follows: Section 2 presents the vortex-shedding da-
tasetThe
used for clustering,
structure Section
of this paper is as3 follows:
explainsSection
the novel unsupervised
2 presents clustering methods,
the vortex-shedding dataset
Section
used for 4clustering,
presents the method’s
Section performance
3 explains the novel using cluster evaluation
unsupervised metrics and
clustering methods, repro-
Section 4
duction of the benchmark vortex-shedding map, and Section 5 reports
presents the method’s performance using cluster evaluation metrics and reproduction the vortex-shed-
ding
of themap generation
benchmark performance for
vortex-shedding map,theand
clustering
Sectionmethods
5 reportsextended to a higher-Reyn-
the vortex-shedding map
olds-number case. Finally, key takeaways from this investigation are provided
generation performance for the clustering methods extended to a higher-Reynolds-number in Section
6. Finally, key takeaways from this investigation are provided in Section 6.
case.

2.
2. Vortex-Shedding
Vortex-SheddingDataset
Dataset
The
The analysis in this study
analysis in this study requires
requires the
the univariate
univariate time
time series
series signatures
signatures of
of local
local flow
flow
measurements
measurementsininthe thewake
wakeofofananoscillating circular
oscillating cylinder
circular experiencing
cylinder forced
experiencing vibration.
forced vibra-
The
tion.vortex shedding
The vortex in a turbulent
shedding wake dataset
in a turbulent was obtained
wake dataset from Kaggle
was obtained from [19]. The[19].
Kaggle dataset
The
was generated
dataset through 2D
was generated CFD simulations
through in OpenFOAMv2006
2D CFD simulations [20]. Turbulent
in OpenFOAMv2006 VIV was
[20]. Turbulent
modelled using the unsteady
VIV was modelled using theReynolds-averaged Navier–Stokes
unsteady Reynolds-averaged (RANS) methodology,
Navier–Stokes (RANS) meth- and
aodology, and a wall-resolved 𝑘-𝜔 shear stress transport (SST) turbulence closure model
wall-resolved k-ω shear stress transport (SST) turbulence closure model was employed.
Furthermore,
was employed. a deformable
Furthermore,mesh was usedmesh
a deformable for fluid–structure interaction to accurately
was used for fluid–structure interaction
capture
to accurately capture the complex flow dynamics. The 2-dimensionalboundary
the complex flow dynamics. The 2-dimensional domain and domain andconditions
bound-
used for the numerical
ary conditions used forstudy are shownstudy
the numerical in Figure 2.
are shown in Figure 2.

Figure2.2. Domain
Figure Domain and
andboundary
boundaryconditions
conditionsused
usedfor
forthe
the2D
2Dsimulation
simulation(not
(notto
toscale).
scale).

The
The accuracy
accuracy ofof 2D
2D RANS
RANS simulations
simulations for
for turbulent
turbulent VIV
VIV has
hasbeen
beendemonstrated
demonstrated in in
past numerical studies [21–23], with the k-ω SST turbulence model performing
past numerical studies [21–23], with the 𝑘-𝜔 SST turbulence model performing well for well for
separated
separated flows
flows with high pressure
with high pressure gradients
gradientscommon
commoninincylinder
cylinderwakes
wakes[24].
[24].
AA wall-
wall-re-
resolved approach was targeted to improve accuracy because of the effect small instabilities
solved approach was targeted to improve accuracy because of the effect small instabilities
near
near the
the surface
surface of
of the
the cylinder
cylinder have
have on
on VIV
VIV response,
response, with
with aa structured,
structured, hexahedral,
hexahedral,
body-fitted mesh generated around the cylinder. The simulations
body-fitted mesh generated around the cylinder. The simulations were were completed using
completed the
using
open-source finite-volume library OpenFOAMv2006 with second-order accuracy achieved
in both spatial and temporal schemes. The established numerical models validated against
similar experimental and computational data to replicate the vortex-shedding patterns
identified by Morse and Williamson [5] provide strong evidence that the data presented in
the study are of high quality and can be used with confidence in future research.
Energies 2023, 16, 4440 7 of 30

The sensor measurements are taken every 0.25 s, for 100 s, along six sampling lines
located downstream in the wake of the oscillating cylinder, as shown in Figure 2. The
sampling lines were orthogonal to the x-axis at streamwise distances of 2D, 4D, 6D, 8D,
10D, and 12D, where D is the cylinder diameter, and each sampling line recorded flow field
measurements at 1000 locations. The data from four types of sensors were recorded from
the simulations, namely, flow speed, |u|; the x-component of velocity, u x ; the y-component
of velocity, uy ; and the vorticity, ω.
The standard fluid measurement used for wake classification is the magnitude of
vorticity [25,26]. The conventional wisdom is that vorticity, a measurement of the local
fluid rotation, will provide a clear signature of the rotating wakes. However, recent studies
have shown that the y-component of flow velocity yielded higher classification accuracy
than vorticity. Alsalman et al. [27] trained several neural networks on time series data
using vorticity, x- and y-components of flow velocity, and flow speed sensors, and found
that training on the y-component of velocity resulted in exceptional classification results,
outperforming vorticity-trained networks even for shallow networks. The performance
gain was also supported by Cann et al. [15], who demonstrated the improved accuracy
of identifying vortex structures from a reduced input feature space by training machine-
learning algorithms on features extracted from the frequency domain of the y-component
of velocity. Based on the evidence from the literature, the y-component of flow velocity
was used in this study to capture flow characteristics more accurately and improve the
performance of the machine-learning models.
Numerous CFD simulations were conducted to provide a higher-density sampling of
the parameter space investigated previously by Morse and Williamson [5]. Furthermore, the
simulations were conducted at Reynolds numbers 4000 and 10,000. The Reynolds number
data at 4000 were used to match the experimental setup of Morse and Williamson [5] and
validate the vortex-shedding map generation procedure, referencing the benchmark map.
The Reynolds number data at 10,000 were used to extend the map generation method to
unknown vortex-shedding regimes. A Reynolds number of 4000 holds specific significance
as it falls within the VIV regime range where turbulent vortex shedding occurs and has
been identified by Wu et al. [7] as the approximate threshold at which the 2S and 2P
modes emerge. Additionally, the relatively lower flow complexity at this Reynolds number
allows for a comprehensive understanding of the underlying flow mechanisms. As a result,
Reynolds numbers around 4000 are often selected as a benchmark value for VIV studies.
The higher Reynolds number of 10,000 was chosen to maintain turbulent vortex shedding
within the flow, while introducing dynamic changes to the vortex-shedding behaviour from
the increased turbulence, which is not present in cases at the lower Reynolds numbers.
The study examined the impact of Reynolds number, streamwise location, and vortex-
shedding modes on signal variance. Data from sampling lines 6D and 4D were used for
clustering analysis in the low- and high-Reynolds-number cases, respectively, based on
their ability to capture wake signals without significant downstream dissipation effects.

3. Unsupervised Clustering Methods

Clustering analysis is a subset of unsupervised machine learning in which a dataset
is decomposed into groups, or “clusters,” based on detected shared structures in the data.
The unsupervised nature of clustering means there is limited external information on the
data structure other than the intrinsic properties.
A time series is a sequence of data points indexed at specific points in time, denoted
Xt = { X (t)}, t ∈ T, where the observation at the time, t, is a subset of allowed timesteps,
T. The set of T represents the time domain over which the time series is defined. In this
application, T is a set of consecutive timesteps in seconds during which the CFD simulations
were conducted. There are three different types of time series clustering strategies, namely,
whole time series [28], subsequence [28], and time point clustering [29]. Subsequence
time series clustering was chosen for this application as it can identify specific patterns in
 A time series is a sequence of data points indexed at specific points in time, denoted
𝑋 = 𝑋(𝑡) , 𝑡 ∈ 𝑇, where the observation at the time, 𝑡, is a subset of allowed timesteps,
𝑇. The set of 𝑇 represents the time domain over which the time series is defined. In this
application, 𝑇 is a set of consecutive timesteps in seconds during which the CFD simula-
Energies 2023, 16, 4440tions were conducted. There are three different types of time series clustering strategies, 8 of 30
namely, whole time series [28], subsequence [28], and time point clustering [29]. Subse-
quence time series clustering was chosen for this application as it can identify specific
patterns in full-time
full-timeseries with
series withnon-steady-state
non-steady-statesignal behaviour
signal and
behaviour isolate
and thethe
isolate desired
desired behaviour
behaviour even when multiple shedding modes are present.
even when multiple shedding modes are present.

3.1. Subsequence
3.1.Data Mining Data Mining
Subsequence
SubsequenceSubsequence
time series clustering
time series involves
clusteringclustering frequent frequent
involves clustering patterns patterns
in an ex- in an extended
tended time series. The frequent
time series. The frequentpatterns within
patterns a time
within series
a time areare
series identified
identifiedusing
usingshape-
shapelet/motif dis-
let/motif discovery
covery techniques.
techniques. The data data mining
miningcommunity
communityhas hasextensively
extensively researched
researched and developed
and developed various
various algorithms
algorithms forseries
for time time motif
seriesdiscovery.
motif discovery.
Torkamani Torkamani
and Lohweg and[29] provide a
Lohweg [29] comprehensive
provide a comprehensive review
review of these of these algorithms.
algorithms.
The time series Themotif
time extraction
series motifmethod for method
extraction this study forwas
thisconducted using the ma-
study was conducted using the matrix
trix profile. The matrix
profile. Theprofile,
matrix first presented
profile, by Yeh et al.
first presented by[30],
Yeh isetaal.
novel
[30],algorithm
is a novelforalgorithm for
conducting aconducting
time series subsequence all-pairs similarity
a time series subsequence all-pairssearch. The algorithm
similarity search. The uses a fast uses a fast
algorithm
similarity search algorithm under 𝑧-normalized Euclidean distance. The method is method
similarity search algorithm under z-normalized Euclidean distance. The sim- is simple,
parameter-free,
ple, parameter-free, and exact,and exact, implying
implying that no
that no false false positives
positives or falseor false negatives
negatives are in- are incurred
curred in thein the process.
process. The profile
The matrix matrix records
profile the
records the distances
distances of a subsequence
of a subsequence with slid- with sliding
length, 𝑚,
ing window window to allm,other
length, to allsubsequences
other subsequencesof the of the same
same length. length. The matrix
The matrix profile contains
profile
contains twotwo components,
components, namely,
namely, a distance
a distance profileand
profile andaaprofile
profile index.
index. The
The distance profile is the
profile is the vector of minimum
minimum Euclidean
Euclideandistance,
distance,and andthetheprofile
profileindex
indexisisthe
thefirst
firstnearest
near- neighbour’s
index. Figure 3 illustrates an example of a time series signal
est neighbour’s index. Figure 3 illustrates an example of a time series signal and its corre- and its corresponding distance
profile obtained from the matrix profile analysis. The distance
sponding distance profile obtained from the matrix profile analysis. The distance profile profile in Figure 3b shows
in Figure 3b the
showssimilarity betweenbetween
the similarity each subsequence and all other
each subsequence and all subsequences of the same length in
other subsequences
the time
of the same length inseries. Theseries.
the time blue trace
The in thetrace
blue time inseries
the highlights
time seriesa highlights
motif and its nearest neighbour,
a motif
which corresponds to the minimum distance value in
and its nearest neighbour, which corresponds to the minimum distance value in the dis- the distance profile (marked with a
red line) and profile index.
tance profile (marked with a red line) and profile index.

(a)

(b)
Figure 3. Matrix profile
Figure of (a) example
3. Matrix time
profile of (a) series
exampleandtime
(b) distance
series andprofile with a highlighted
(b) distance profile with motif
a highlighted motif
and neighbourand
(blue trace) identified by minimum distance profile value (red line).
neighbour (blue trace) identified by minimum distance profile value (red line).

Together, theTogether,
time series
theand distance
time profile
series and in Figure
distance 3 provide
profile a visual
in Figure represen-
3 provide a visual represen-
tation of the tation
structure andstructure
of the dynamics of dynamics
and the time series
of thedata,
timeallowing us to
series data, identifyustime
allowing to identify time
series patterns and anomalies. Low distance profile values indicate similarity between
subsequences within the series, corresponding to motifs, while high values indicate unique
subsequences, corresponding to anomalies. The lowest values in the profile represent the
prominent subsequence, which can be extracted as the motif. Careful selection of the sliding
window length is crucial for domain-specific motif extraction.
Energies 2023, 16, 4440 9 of 30

3.2. Clustering Evaluation Metric

Cluster evaluation is essential for obtaining meaningful and quantifiable results. Scalar
accuracy indices assess cluster quality through numerical values, consisting of internal and
external indices. External indices cannot be utilized for unsupervised clustering due to
the absence of external data, while internal indices use intrinsic information to determine
clustering performance. The clustering quality is based on two parameters: compactness
and separation. Compactness measures how close the objects within the same cluster
relate to each other, known as intra-cluster similarity [31]. Separation measures how well
separated a cluster is from other clusters, known as an inter-cluster similarity [31]. Many
internal indices are proposed in the literature, but the silhouette and Dunn indices were
selected for the evaluation metrics based on their ability to provide a measure of both
intra-cluster similarity and inter-cluster dissimilarity [32].

3.2.1. Silhouette Index

The silhouette coefficient, introduced by Rousseeuw [33], measures the similarity of an
element to its cluster (compactness) compared to the other sets (separation). The silhouette
coefficient is determined as follows:
b−a
silhouette index = , (3)
maximum(b − a)

where the parameter a is the mean distance between a sample and all the points within the
same set, and the parameter b is the mean distance from the sample to all the other points
in the next nearest cluster. The value of the silhouette index is bounded between −1 and 1.
As the silhouette approaches 1, the cluster containing the sample element will be far from
the following cluster but compact within the elements of the same cluster. Negative values
indicate that a sample element is closer to another cluster’s elements than to its assigned
elements. The clustering analysis’s quality can be measured by calculating the average
silhouette coefficient for all objects in the dataset.

3.2.2. Dunn Index

The Dunn index was first introduced by Dunn [34]. The Dunn index is the ratio
of the minimum distance between clusters (inter) and the maximum distance within a
cluster (intra). To quantify the Dunn index, consider any two clusters Ci and Cj with the
corresponding members Xi and X j . The Dunn index is defined as



minimum intercluster distance min(Ci , Cj ) dist Xi , X j
Dunn index = = . (4)
maximum intracluster distance max(Ck ) (dist( Xk , Xk ))

The Dunn index should be maximized such that the minimum intercluster distance is
large (separation) and the maximum intracluster distance is small (compactness).

3.3. Proposed Clustering Algorithms

The selection of clustering algorithms is crucial for clustering analysis, particularly for
time series data. There are two main approaches for clustering time series data: conven-
tional and hybrid [35]. In this study, the best-performing algorithms for clustering time
series subsequences of vortex shedding were chosen, including two traditional single-step
and two proposed hybrid methods.
The traditional clustering methods of k-Means partitioning and agglomerative hierar-
chical clustering were chosen for their demonstrated performance. The k-Means method
was selected based on its demonstrated ability to extract highly separated clusters. The
k-Means algorithm was implemented using the k-Means++ initialization method to select
initial cluster centres that are distant, resulting in better results than random initialization
by avoiding local minima [36]. The agglomerative hierarchical method was selected based
on the improvements over the partitioning methods, which do not require the number of
 The traditional clustering methods of 𝑘-Means partitioning and agglomerative hier-
archical clustering were chosen for their demonstrated performance. The 𝑘 -Means
method was selected based on its demonstrated ability to extract highly separated clus-
ters. The 𝑘 -Means algorithm was implemented using the 𝑘 -Means++ initialization
method to select initial cluster centres that are distant, resulting in better results than ran-
Energies 2023, 16, 4440 10 of 30
dom initialization by avoiding local minima [36]. The agglomerative hierarchical method
was selected based on the improvements over the partitioning methods, which do not
require the number of partitions to be prespecified, as the hierarchy structure can be
partitions
spliced at to
thebeappropriate
prespecified, as the
level hierarchy
to extract the structure can hyperparameters
clusters. The be spliced at the appropriate
selected for
level to extract the clusters. The hyperparameters selected
the agglomerative algorithm were complete linkage and cosine for the agglomerative algorithm
affinity distance, which
were complete linkage and cosine affinity
achieved the best balance of evaluation metrics. distance, which achieved the best balance of
evaluation metrics.
Multi-step clustering methods, known as hybrid methods, can improve the clustering
Multi-stepofclustering
performance methods,
traditional methods known as hybrid
[37–40]. Thesemethods,
methodscan improve
conduct the clustering
a pre-clustering
performance of traditional methods [37–40]. These methods conduct a pre-clustering
phase that produces a large number of separate clusters that are then merged using a final
phase that produces a large number of separate clusters that are then merged using a
clustering method to obtain the desired number of clusters. A reduced dataset of proto-
final clustering method to obtain the desired number of clusters. A reduced dataset of
types is used for the final clustering stage, which comprises samples extracted from the
prototypes is used for the final clustering stage, which comprises samples extracted from the
subclusters that minimize the pairwise distances between themselves and the other cluster
subclusters that minimize the pairwise distances between themselves and the other cluster
members. The various steps involved in the proposed hybrid algorithms for clustering the
members. The various steps involved in the proposed hybrid algorithms for clustering the
motif dataset are illustrated in the flow diagram in Figure 4.
motif dataset are illustrated in the flow diagram in Figure 4.

Pre- Reduced
processed Similarity in Time Prototype
Pre-Clustering
Motif Distance Motif
Dataset Dataset

Similairty in Shape
Final Clustering Clusters
Distance

Figure 4. Block diagram summarizing the key steps in the proposed hybrid algorithm.
Figure 4. Block diagram summarizing the key steps in the proposed hybrid algorithm.
Hybrid methods yield the best clustering results when the pre-clustering phase max-
imizes Hybrid methodsindex,
the silhouette yield the best clustering
indicating that theresults when the pre-clustering
high-resolution phase
clusters capture max-
distinct
imizes the silhouette index, indicating that the high-resolution
groups. The silhouette index is expected to decrease in the final clustering stage as the clusters capture distinct
groups. merge
clusters The silhouette
to produce index
more is general
expected to decrease
clusters in theoverarching
that reveal final clustering stageleading
patterns, as the
clusters merge to produce more
to an increase in the Dunn index value. general clusters that reveal overarching patterns, leading
to anMost
increase
hybridin the Dunn use
methods index value.distance metrics for each step, usually a similarity
varying
in timeMost hybridfor
distance methods use varyingand
the pre-clustering distance metrics for
a shape-based each step,
similarity usuallyfor
distance a similarity
the final
in time distance for the pre-clustering and a shape-based similarity distance
clustering. The advantage of multi-step clustering is the ability to calculate the computation- for the final
clustering. The advantage of multi-step clustering is the ability
ally expensive dynamic time warping (DTW) similarity matrix on the reduced dataset of to calculate the computa-
tionally expensive
prototypes. dynamic
The clustering time warping
results obtained(DTW) using thesimilarity matrix on
DTW distance the reduced
provide dataset
an advantage
of prototypes.
for shape-basedThe clustering
clustering results obtained using the DTW distance provide an ad-
analysis.
vantage
The for shape-based
advantage of theclustering
multi-stepanalysis.
clustering procedure is depicted in Figure 5 using an
exampleTheof advantage
clusteringofsubsequences
the multi-stepusingclustering procedure
a two-step hybridismethod.
depictedThe
in Figure
figure 5illustrates
using an
example
the initialof clustering subsequences
subclusters’ inter-similarity using
based a two-step hybrid
on similarity in method.
time andThethefigure illustrates
intra-similarity
the initialthe
between subclusters’
prototypesinter-similarity
achieved in thebased on similarity
final clustering in time and the intra-similarity
stage.
between
Two the prototypes
hybrid methods achieved in the final
are proposed clustering
in this study. The stage.
first hybrid method, denoted
Hybrid A, uses a combination of density-based spatial clustering of applications with noise
(DBSCAN) [41] and agglomerative clustering algorithms. The DBSCAN method provides
an advantageous initial clustering step since the clusters are automatically determined
based on the input data structures. The DBSCAN algorithm only requires two parameters,
namely, the number of minimum samples and radii. These parameters represent the core
samples, which are a subset of the data which include a minimum number of samples,
min_samples, that are within a distance radius, ε. The samples that are not core and are
further than ε distance from any core sample are considered an outlier. The agglomerative
algorithm was then used to group the subclusters from the DBSCAN algorithm into the
final number of clusters.
The final hybrid method proposed in this study, denoted Hybrid B, was conceptual-
ized by combining the best-performing single-step clustering analysis into a multi-step
method. The hybrid method uses the k-means algorithm for the pre-clustering phase and
Energies 2023, 16, 4440 11 of 30

Energies 2023, 16, x FOR PEER REVIEW

theagglomerative method for the final clustering with the same hyperparameters 11
as of
the32
single-step methods.

Figure5.5.Demonstration
Figure Demonstrationof
ofthe
theclustering
clusteringprocedure
procedureof
ofthe
thefull
fullsubsequence
subsequenceset
setand
andprototype
prototypeset
set
using
usingvarying
varyingdistance
distancemetrics
metricsininaatwo-stage
two-stagehybrid
hybridclustering
clusteringmethod.
method.
3.4. Vortex-Shedding Map Generation
Two hybrid methods are proposed in this study. The first hybrid method, denoted
The A,
Hybrid vortex-shedding
uses a combination maps of were generated using
density-based spatialanclustering
ensemble-type method. This
of applications with
method decomposed each node in the normalized amplitude–wavelength
noise (DBSCAN) [41] and agglomerative clustering algorithms. The DBSCAN map into primary
method
and secondary
provides time series signatures
an advantageous based on
initial clustering thesince
step clustering results.
the clusters areSpecifically,
automaticallynodes
de-
with a strong vortex-shedding mode will have a larger proportion
termined based on the input data structures. The DBSCAN algorithm only requires two of time signatures
clustered
parameters, together
namely, in the
thenumber
same cluster. On thesamples
of minimum other hand, unstable
and radii. Thesenodes such as
parameters the
repre-
2PO mode, which can exhibit both 2S and 2P modes, will share multiple clusters within
sent the core samples, which are a subset of the data which include a minimum number
the node. By decomposing each node into primary and secondary time series signatures,
of samples, min_samples, that are within a distance radius, 𝜀. The samples that are not
the ensemble approach used to generate the vortex-shedding maps may allow for a more
core and are further than 𝜀 distance from any core sample are considered an outlier. The
nuanced understanding of the underlying flow dynamics. The primary and secondary
agglomerative algorithm was then used to group the subclusters from the DBSCAN algo-
signatures represent the dominant and sub-dominant modes of variability in the data,
rithm into the final number of clusters.
respectively. The clustering methods used to identify these signatures may help to identify
The final hybrid method proposed in this study, denoted Hybrid B, was conceptual-
patterns or groups of nodes with similar flow behaviour, which could provide insights
ized by combining the best-performing single-step clustering analysis into a multi-step
into the mechanisms that drive vortex shedding and the formation of flow patterns. The
method. The hybrid method uses the 𝑘-means algorithm for the pre-clustering phase and
ensemble-type method is a promising approach for generating vortex-shedding maps and
the agglomerative method for the final clustering with the same hyperparameters as the
studying the complex flow dynamics that occur in a variety of fluid systems.
single-step methods.
4. Vortex-Shedding Map Generation at Low Reynolds Number
3.4. Vortex-Shedding Map Generation
This section applies the unsupervised clustering methods to a low-Reynolds-number
scenarioTheofvortex-shedding
vortex shedding maps from an were generated
oscillating usingtoan
cylinder ensemble-type
replicate the benchmarkmethod. This
regime
method
map decomposed
[7]. The performance eachofnode
each in the normalized
clustering method amplitude–wavelength
is evaluated based on both map into pri-
clustering
mary and
accuracy secondary
and the qualitytime series
of the signatures
resulting based on themaps.
vortex-shedding clustering results. Specifically,
nodes with a strong vortex-shedding mode will have a larger proportion of time signa-
4.1. Introduction
tures clustered together in the same cluster. On the other hand, unstable nodes such as the
2POThe mode, which can
following exhibitonboth
locations the 2S and 2Pvortex-shedding
reference modes, will share multiple
map producedclusters within
by Morse
the Williamson
and node. By decomposing each to
[5] were selected node intothe
create primary anddataset
clustering secondary time
based on series signatures,
observed vortex-
the ensemble
shedding approach
behaviour used
in the to generate the vortex-shedding maps may allow for a more
signals.
nuanced understanding of the underlying flow dynamics. The primary and secondary
signatures represent the dominant and sub-dominant modes of variability in the data, re-
spectively. The clustering methods used to identify these signatures may help to identify
patterns or groups of nodes with similar flow behaviour, which could provide insights
into the mechanisms that drive vortex shedding and the formation of flow patterns. The
Energies 2023, 16, 4440 12 of 30

The sampling nodes included three vortex-shedding modes (2S, 2P, and 2PO) that
exhibited stable behaviour. Nodes on the frequency sampling line of λ∗ =4 were excluded
due to the absence of coherent patterns in the data near the C (2S) to 2S transition. The
absence of a distinct transition between the C (2S) and pure 2S mode was reported as the
least distinct boundary compared to any modes by Morse and Williamson [5]. The data
points in the C (2S) region were carefully selected due to the small vortices that coalesce in
the near wake detected by the sampling line at 6D.
The subsequences in the data were mined using the matrix profile motif extrac-
tion method. The specified window size for the algorithm was set to the equivalent
of 64 timesteps to represent a total of at least two cycles of oscillation. Selecting a sliding
window to capture multiple oscillations improves clustering results by producing more con-
sistent and representative patterns. The subsequence extraction process was performed on
all of the raw time series to generate a single representative pattern for clustering analysis.
The number of clusters specified in the cluster algorithms is five, considering the do-
main knowledge that three distinct vortex-shedding modes should be present in the data, as
Energies 2023, 16, x FOR PEER REVIEW 13 of 32
shown in Figure 6. The two additional clusters were designated for separating transitional
modes such as 2PO and any noise or outlier points identified in the clustering procedure.

∗ ∗
Figure6.6.Dataset
Figure Datasetsampled
samplednodes
nodesoverlaid
overlaidreference
referencenormalized
normalizedamplitude
amplitude(A (𝐴∗))–wavelength (𝜆∗))
–wavelength(λ
plane
plane[5],
[5],with
withthree
threesubpanels
subpanelsshowing example
showing examplesignals extracted
signals from
extracted each
from corresponding
each node
corresponding on
node
on map,
map, denoted
denoted 1, 2,1,and
2, and
3. 3.

4.2.
4.2. k-Means
𝑘-Means
The
Thefirst
firstclustering
clusteringmethod
methodimplemented
implementedto tocluster
clustertime
timeseries
seriessubsequences
subsequenceswas wasthe
the
k-Means algorithm. The internal metrics of the silhouette and Dunn indices were
𝑘 -Means algorithm. The internal metrics of the silhouette and Dunn indices were em- employed
to evaluate
ployed the clustering
to evaluate performance,
the clustering and the and
performance, results
theare summarized
results in Table in
are summarized 1. Table 1.

Table1.1.Clustering
Table Clusteringperformance
performancemetrics ofk 𝑘-Means
metricsof -Means method
methodatatRe
Re==4000.
4000.
Evaluation Metric
Evaluation Metric
Clustering
ClusteringAlgorithm
Algorithm Initialization Method
Initialization Method Sil Dunn
Sil Dunn
k-Means
𝑘-Means k-Means++
𝑘-Means++ 0.6559
0.6559 0.15295
0.15295

The relatively high evaluation metrics indicate that the clusters generated are suffi-
ciently separate and compact, which can visually be confirmed by the time series subse-
quence clusters shown in Figure 7.
 Energies 2023, 16, 4440 13 of 30

The relatively high evaluation metrics indicate that the clusters generated are suffi-
Energies 2023, 16, x FOR PEER REVIEW 14 of 32
Energies 2023, 16, x FOR PEER REVIEW ciently
separate and compact, which can visually be confirmed by the time series subse-
14 of 32
quence clusters shown in Figure 7.

Figure Subsequence dataset

7. Subsequence datasetclustered
clusteredbyby 𝑘-Means
k-Means method at Re 4000, with 𝑁
Re ==4000, N representing the
Figure 7. Subsequence dataset clustered by 𝑘-Means method at Re = 4000, with with
𝑁 representing the
number of samples in each cluster relative to the total dataset.
number of samples in each cluster relative to the total dataset.

The similarities
The similarities of
of cluster
cluster 11 and
and cluster
cluster 55 are
are illustrated
illustrated in in Figure
Figure 7,
7, both clusters
clusters ex-
The similarities of cluster 1 and cluster 5 are illustrated in Figure 7, both both
clusters ex- ex-
hibiting
hibiting a double-peak
a double-peak behaviour
behaviour synonymous
synonymous with the 2P group. However, the algorithm
hibiting a double-peak behaviour synonymous withwith
the 2P thegroup.
2P group. However,
However, the algo-
the algo-
identified
rithm a subtle
identified a ramping
subtle of the
ramping peak
of in
the cluster
peak 1,
in which
cluster is
1, characteristic
which is of the 2PO
characteristic mode.
of the
rithm identified a subtle ramping of the peak in cluster 1, which is characteristic of the
The
2PO vortex-shedding
mode. The map was
vortex-shedding thenmapplotted
was with
then the primary
plotted with cluster
the candidates
primary identified,
cluster candi-
2PO mode. The vortex-shedding map was then plotted with the primary cluster candi-
as shown in Figure 8.
datesdates identified,
identified, as shown
as shown in Figure
in Figure 8. 8.

Figure
Figure 8. Vortex-shedding
8. Vortex-shedding map using
map using 𝑘-Means
𝑘-Means method
method at Reatat
=ReRe = 4000.
4000.
Figure 8. Vortex-shedding map using k-Means method = 4000.

The generated
The generated vortex-shedding
vortex-shedding map map identifies
identifies several
several regions
regions of interest,
of interest, withwith cluster
cluster
(𝜆 ∗ ∗)
1 designated to the node at
∗ ∗ , 𝐴
1 designated to the node at (𝜆 , 𝐴 ) = (6, 0.7) resembling a 2PO mode with ∗a reduced first first
= (6, 0.7) resembling a 2PO mode with a reduced
peakpeak
in theinpattern.
the pattern.
∗
Cluster
Cluster 2 is primarily
2 is primarily forlow-frequency
for the the low-frequency
∗
𝜆∗ = 𝜆2, which
case case = 2, which
occurs
∗
occurs
mainly at 𝐴 = 0.3 and as the secondary mode at 𝐴 = 0.5. The low-amplitude,
mainly at 𝐴∗ = 0.3 and as the secondary mode at 𝐴∗ = 0.5. The low-amplitude, 𝐴∗ = 0.1, 𝐴 = 0.1,
spacespace exhibits
exhibits the regular
the regular sinusoidal
sinusoidal pattern
pattern indicative
indicative of 2Sof behaviour
2S behaviourunder under
the the
Energies 2023, 16, 4440 14 of 30

The generated vortex-shedding map identifies several regions of interest, with cluster
Energies 2023, 16, x FOR PEER REVIEW1 designated to the node at (λ∗ , A∗ ) = (6, 0.7) resembling a 2PO mode with a reduced 15 of 32
first peak in the pattern. Cluster 2 is primarily for the low-frequency case λ∗ = 2, which
occurs mainly at A∗ = 0.3 and as the secondary mode at A∗ = 0.5. The low-amplitude,
A∗ = 0.1, space exhibits the regular sinusoidal pattern indicative of 2S behaviour under
identification of cluster
the identification 3. A similar
of cluster sinusoidal
3. A similar pattern is
sinusoidal observed
pattern with cluster
is observed 4, differing
with cluster 4,
∗ ∗
differing by a lower signal amplitude. Finally, cluster 5 is primarily located at λat
by a lower signal amplitude. Finally, cluster 5 is primarily located at 𝜆 = 6 ∗ =𝐴6 =
at
(𝜆∗ ∗ )
(0.3,
∗ 0.5) with an additional split located at , 𝐴 =∗ (2, 0.5).
∗ The additional
A = (0.3, 0.5) with an additional split located at (λ , A ) = (2, 0.5). The additional node node shows
ashows
ratheraweak
ratheridentification due to the
weak identification dueswitching between
to the switching cluster cluster
between 5 and cluster 2.
5 and cluster 2.

4.3.
4.3.Agglomerative
Agglomerative
The
The following
followingclustering
clusteringalgorithm
algorithm implemented
implemented waswasthe
theagglomerative
agglomerativehierarchical
hierarchical
algorithm.
algorithm. The internal indices used to quantify the clustering performance are summarized
internal indices used to quantify the clustering performance are summa-
in Table
rized 2.
in Table 2.

Table 2.
Table Clustering performance
2. Clustering performance metrics
metrics of
of agglomerative
agglomerative method
method at
at Re
Re ==4000.
4000.

Evaluation Metric
Evaluation Metric
Clustering
ClusteringAlgorithm
Algorithm Linkage
Linkage Affinity
AffinityDistance
Distance Sil
Sil Dunn
Dunn
Agglomerative
Agglomerative Complete
Complete Cosine
Cosine 0.6794
0.6794 0.61721
0.61721

The
The agglomerative
agglomerativemethod
methodproduces
producesclusters
clusterswith
withaamuch
muchlarger
largerDunn
Dunnindex
indexcom-
com-
pared
pared to the partitioning method, indicating the quality of the groups, which can be
to the partitioning method, indicating the quality of the groups, which can be seen
seen
by
by the
the distinct
distinct clusters
clusters shown
shown in
in Figure
Figure 9.
9.

Figure 9.
Figure Subsequencedataset
9. Subsequence datasetclustered
clusteredby
byagglomerative
agglomerative(complete,
(complete,cosine)
cosine)method
methodat atRe
Re==4000,
4000,
with 𝑁
with representing the
N representing the number
number ofof samples
samples in
in each
each cluster
clusterrelative
relativeto
tothe
thetotal
totaldataset.
dataset.

Thecluster
The clusterdiagram
diagramin inFigure
Figure99displays
displaysthe
thedistinctiveness
distinctivenessof ofthe
thegenerated
generatedclusters,
clusters,
particularly clusters
particularly clusters 2,
2, 3,
3, and
and 4.4. However,
However, cluster
cluster 55 appears
appears toto inhabit
inhabit the
the same
same subspace
subspace
as cluster
as cluster 1,
1, potentially
potentially due
due toto its
its higher
higher amplitude
amplitude and and longer
longer wavelength.
wavelength.Despite
Despitethis,
this,
the vortex-shedding map clearly illustrates the primary cluster candidates,
the vortex-shedding map clearly illustrates the primary cluster candidates, as shown in as shown in
Figure 10.
Figure 10.
Energies 2023, 16,
Energies 2023, 16,4440
x FOR PEER REVIEW 15 of
16 of 30
32

Figure 10. Vortex-shedding

Figure 10. Vortex-shedding map
map using
using the
the agglomerative
agglomerative method
methodat
atRe
Re== 4000.
4000.

The
The vortex-shedding
vortex-sheddingmap mapisolates several
isolates similar
several regions
similar of subsequence
regions patterns,
of subsequence con-
patterns,
sistent with those observed in the k-Means method. Additionally, clusters 1 and
consistent with those observed in the 𝑘-Means method. Additionally, clusters 1 and 3 ex- 3 exhibit
patterns similarsimilar
hibit patterns to the signature expected
to the signature for the 2S
expected formode, while
the 2S a group
mode, whileofa nodes
groupidentified
of nodes
as cluster 2 on the sampling line ∗ =6 demonstrates
∗ the double peak of the pair of vortices
identified as cluster 2 on the sampling line 𝜆 = 6 demonstrates the double peak
λ of the
being shed from the 2P mode.
pair of vortices being shed from the 2P mode.
4.4. Hybrid A
4.4. Hybrid A
The clustering analysis results are presented for the corresponding phases using
The clustering analysis results are presented for the corresponding phases using
DBSCAN as the pre-clustering phase and agglomerative for the final merging of clusters.
DBSCAN as the pre-clustering phase and agglomerative for the final merging of clusters.
The pre-clustering phase identified six clusters, with three outliers, and a prototype was
The pre-clustering phase identified six clusters, with three outliers, and a prototype was
created for each cluster, which was used for the final clustering stage. The clustering
created for each cluster, which was used for the final clustering stage. The clustering per-
performance results of both phases are summarized in Table 3.
formance results of both phases are summarized in Table 3.
Table 3. Clustering performance metrics of Hybrid A method at Re = 4000.
Table 3. Clustering performance metrics of Hybrid A method at Re = 4000.
Evaluation Metric
Phase Clustering Algorithm Number of Clusters Evaluation Metric
Phase Clustering Algorithm Number of Clusters Sil Dunn
Sil Dunn
1: Pre-Clustering DBSCAN 6 (3 Noise Points) 0.7185 0.36049
1:
2: Pre-Clustering
Final Clustering DBSCAN
Agglomerative 6 (3 Noise
5 Points) 0.7185
0.7031 0.36049
0.39836
2: Final Clustering Agglomerative 5 0.7031 0.39836

The DBSCAN algorithm in the pre-clustering phase produced well-separated clusters

The DBSCAN algorithm in the pre-clustering phase produced well-separated clus-
resulting in a high silhouette index. The final merging step improved the Dunn index
ters resulting in a high silhouette index. The final merging step improved the Dunn index
marginally at the cost of a slight reduction of the silhouette index. The generated clusters
marginally at the cost of a slight reduction of the silhouette index. The generated clusters
merged for the entire dataset are shown in Figure 11.
merged for the entire dataset are shown in Figure 11.
The associated vortex-shedding map was generated based on the primary cluster
candidates and the proportions of cluster samples at each node shown in Figure 12.
Using the Hybrid A method, the generated regime map shares its overall structure
with the other methods. Specifically, the nodes along the line λ∗ =6 all belong to the
primary mode identified in cluster 4 that resembles the 2P mode. The low amplitude nodes
along the horizontal line A∗ = 0.1 share cluster 1. The sinusoidal mode with a lower
amplitude is located at (λ∗ , A∗ ) = (4, 0.7) and (λ∗ , A∗ ) = (4, 0.9). Finally, the node located
at (λ∗ , A∗ ) = (2, 0.5) comprises two modes labelled by cluster 4 and cluster 3.
Energies 2023, 16, x FOR PEER REVIEW 17 of 32

Energies
Energies 2023,
2023, 16,
16, x4440
FOR PEER REVIEW 1716ofof32
30

Figure 11. Subsequence dataset clustered by Hybrid A method at Re = 4000, with 𝑁 representing
the number of samples in each cluster relative to the total dataset.

Figure
FigureThe
11. Subsequencevortex-shedding
11. associated
Subsequence dataset
dataset clustered
clusteredby Hybrid
map
by wasAA
Hybrid method
methodatatRe
generated = 4000,
based
Re onwith
= 4000, the Nprimary
with representing the
𝑁 representing
cluster
number of
candidates
the samples
number of and in each cluster
the proportions
samples relative to
of relative
in each cluster the
cluster to total
samples dataset.
the totalatdataset.
each node shown in Figure 12.

The associated vortex-shedding map was generated based on the primary cluster
candidates and the proportions of cluster samples at each node shown in Figure 12.

.
Figure 12. Vortex-shedding map using Hybrid A method at Re = 4000.
Figure 12. Vortex-shedding map using Hybrid A method at Re = 4000.
4.5. Hybrid B .
Using the Hybrid
This section A method,
presents results the
for generated regime map
the pre-clustering and shares its overallphases
final clustering structure
of
with
Figure
the lastthe other
12.proposed methods.
Vortex-shedding
hybrid Specifically,
map thethe
using Hybrid
method. In nodes alongat the
Apre-clustering
method Re =line
4000.𝜆∗ =the
phase, 6 all belong algorithm
k-Means to the pri-
mary
requires mode the identified in clusterof4clusters
optimum number that resembles
to maximizethe 2P themode. The low
silhouette index.amplitude nodes
The optimum
along
number Using the Hybrid
theofhorizontal
clusters Athe
line
for 𝐴∗ pre-clustering
= 0.1 the
method, sharegenerated
phase1.regime
cluster The 20
was map
as itshares
sinusoidal modeits with
maximized overall
the structure
a lower am-
silhouette
with
index,
plitude the isother
representing
locatedmethods.(𝜆∗Specifically,
, 𝐴∗ ) = (4, 0.7)
atwell-separated the
clusters.
and (𝜆
nodes ∗ along
The ∗)
, 𝐴clustering
the0.9).
= (4, line 𝜆∗ = 6 all
performance
Finally, thebelong
of bothto
node the pri-
phases
located is
at
(𝜆 ∗ ∗mode)
mary , 𝐴 = (2, 0.5) comprises two modes labelled by cluster 4 and cluster 3.
summarized identified
in Table in
4. cluster 4 that resembles the 2P mode. The low amplitude nodes
alongThe the evaluation
horizontal line
metrics 𝐴∗ = 0.1 share
indicate cluster 1. separated
significantly The sinusoidal mode
clusters with
in the a lower
first stage am-
and
(𝜆∗ ∗) (𝜆∗ ∗)
plitude is located at , 𝐴 = (4, 0.7) and , 𝐴 = (4, 0.9). Finally,
more general clusters merged in the final stage observed by the decrease in the silhouette the node located at
(𝜆∗ ∗
, 𝐴 )but
index = (2,an 0.5) comprises two
improvement modes
to the Dunn labelled
index.byThe cluster
final4 clusters
and cluster 3. for the entire
merged
dataset are shown in Figure 13.
 dex, representing well-separated clusters. The clustering performance of both phases is
summarized in Table 4.

Table 4. Clustering performance metrics of Hybrid B method at Re = 4000.

Energies 2023, 16, 4440 Evaluation Metric

17 of 30
Phase Clustering Algorithm Clusters
Sil Dunn
Pre-Clustering 𝑘-Means 20 0.7464 0.09175
Final
Table Clustering
4. Clustering performanceAgglomerative
metrics of Hybrid B method at5Re = 4000.0.5336 0.11211

The evaluation metricsClustering

indicate significantly separated Evaluation Metric
Phase Algorithm Clusters clusters in the first stage and
more general clusters merged in the final stage observed by the decrease in the Dunn
Sil
silhouette
index Pre-Clustering
but an improvement to the k-Means 20 clusters0.7464
Dunn index. The final merged for 0.09175
the entire
Final
dataset areClustering
shown in Figure 13. Agglomerative 5 0.5336 0.11211

Figure 13. Subsequence dataset

dataset clustered
clusteredbybyHybrid
HybridB Bmethod
method atat
ReRe
= 4000, with
= 4000, with 𝑁 representing
N representing the
number
the of samples
number in each
of samples cluster
in each relative
cluster to the
relative to total dataset.
the total dataset.

Overall, the
Overall, the clusters
clusters capture
capturedistinct
distinctsubsequence
subsequencepatterns.
patterns.However,
However,the
thelow
lownumber
num-
of samples in cluster 2 and signal variation in cluster 1 could be responsible for the relatively
ber of samples in cluster 2 and signal variation in cluster 1 could be responsible for the
Energies 2023, 16, x FOR PEER REVIEW 19 of 32
poor performance
relatively of the Dunn
poor performance index.
of the DunnTheindex.
generated vortex-shedding
The generated map is shown
vortex-shedding map inis
Figure 14 for the respective cluster candidate proportions.
shown in Figure 14 for the respective cluster candidate proportions.

Vortex-shedding map using Hybrid B method at Re = 4000.

Figure 14. Vortex-shedding

The vortex-shedding map exhibits a similar structure to the regime map generated
using the Hybrid A method. For example, the nodes along the line 𝜆∗ = 6 are indicative
to the 2P and 2P0 modes. The sinusoidal signals of clusters 1 and 5 are located in similar
locations to the prior hybrid method results. These results provide further validation of
the effectiveness of the proposed hybrid method.
Energies 2023, 16, 4440 18 of 30

The vortex-shedding map exhibits a similar structure to the regime map generated
using the Hybrid A method. For example, the nodes along the line λ∗ = 6 are indicative
to the 2P and 2P0 modes. The sinusoidal signals of clusters 1 and 5 are located in similar
locations to the prior hybrid method results. These results provide further validation of the
effectiveness of the proposed hybrid method.

4.6. Discussion
Each clustering method offers varying benefits to the unsupervised clustering task,
and their respective performance and qualitative attributes are discussed in this section,
culminating in an assessment of their suitability for producing vortex-shedding maps. The
clustering method’s ability to produce separate and compact clusters was accessed by the
internal evaluation metrics summarized in Table 5.

Table 5. Final clustering performance metrics of proposed methods at Re = 4000.

Representation Evaluation Metric

Type Clustering Algorithm
Method Sil Dunn
Partitioning k-Means Raw Time Series 0.6559 0.15295
Hierarchical Agglomerative Raw Time Series 0.6794 0.61721
Hybrid DBSCAN/Agglomerative Raw Time Series 0.7031 0.39836
Hybrid k-Means/Agglomerative Raw Time Series 0.5336 0.11211

The k-Means method is beneficial for implementation due to its simple algorithm and
linear time complexity. However, the requirement to specify the number of clusters and
the assumption of spherical cluster shape can impact performance for the application of
generating vortex-shedding maps. Hierarchical methods offer benefits such as not requiring
the number of clusters to be specified and great visualization power but have drawbacks
such as increased computational complexity and the inability to reassign samples once
merged. The hierarchical agglomerative algorithm outperforms the partitioning method in
both the silhouette and Dunn indices. However, maximizing the Dunn index can result
in few samples being clustered together, which affects generalizability. Nevertheless, the
agglomerative algorithm’s clustering performance is still competitive and is considered for
Energies 2023, 16, x FOR PEER REVIEW
high-Reynolds-number cases. 20 of 32

The single-step methods using traditional clustering are validated in the generation of
vortex-shedding maps by comparing them to the reference map [5]. The vortex-shedding
shedding map produced
map produced using theusing the 𝑘-Means
k-Means method is method is overlaid
overlaid with thewith the regimes
regimes in the map
in the reference
reference map as shown in Figure 15. The coloured ellipses in the figure are used
as shown in Figure 15. The coloured ellipses in the figure are used purely for the purpose purely
for the purposeregions
of visualizing of visualizing regions
of similar of similar vortex-shedding
vortex-shedding behaviour; thebehaviour;
size andthe size and of the
orientation
orientation of the ellipses were arbitrarily determined to convey a
ellipses were arbitrarily determined to convey a sense of these regions.sense of these regions.

Figure
Figure15.
15.Overlaid
Overlaidbenchmark regimes
benchmark on vortex-shedding
regimes map produced
on vortex-shedding with 𝑘-Means.
map produced with k-Means.

The 𝑘-Means method produced groups that align with the expected regimes estab-
lished by Morse and Williamson [5]. Specifically, the algorithm identified a cluster solely
associated with the 2PO mode. Clusters 4 and 5 exclusively inhabit the 2S and 2P modes
on the benchmark map, with the assigned cluster numbers showing more variation in the
Energies 2023, 16, 4440 Figure 15. Overlaid benchmark regimes on vortex-shedding map produced with 𝑘-Means.
19 of 30

The 𝑘-Means method produced groups that align with the expected regimes estab-
lishedTheby k-Means
Morse and Williamson
method produced [5]. Specifically,
groups that alignthewith
algorithm identified
the expected a cluster
regimes estab- solely
lished by Morse
associated and 2PO
with the Williamson
mode. [5]. Specifically,
Clusters 4 andthe algorithm identified
5 exclusively a cluster
inhabit the 2S andsolely
2P modes
associated with the 2PO mode. Clusters 4 and 5 exclusively inhabit the 2S and
on the benchmark map, with the assigned cluster numbers showing more variation in the 2P modes
Con(2S)
the region.
benchmark
Themap, with
nodes the assigned
located in thiscluster
region numbers
at low showing
values ofmore variation in the ampli-
non-dimensional
C (2S) region. The nodes located in this region at low values of non-dimensional amplitude
tude 𝐴∗ < 0.2 are strongly controlled by the clear sinusoidal cluster of number 3. The vor-
A∗ < 0.2 are strongly controlled by the clear sinusoidal cluster of number 3. The vortex-
tex-shedding map produced using the agglomerative algorithm is overlaid with the re-
shedding map produced using the agglomerative algorithm is overlaid with the regimes in
gimes in the reference
the reference mapinasFigure
map as shown shown 16.in Figure 16.

Figure
Figure 16.
16. Overlaid benchmark
Overlaid benchmark regimes
regimes on vortex-shedding
on vortex-shedding map with
map produced produced with agglomerative
agglomerative algorithm. al-
gorithm.
Clusters 2 and 4 identify general regions fitting the expected modes of 2P and 2S,
respectively.
Clusters It is not4 surprising
2 and that cluster
identify general 4 is designated
regions in the 2POmodes
fitting the expected region,ofas2Ptheand 2S,
vortex-shedding behaviour of 2PO is similar to 2P and can switch intermittently.
respectively. It is not surprising that cluster 4 is designated in the 2PO region, as the vor-
This paper proposes two hybrid clustering methods with varying attributes and
tex-shedding behaviour of 2PO is similar to 2P and can switch intermittently.
inherent limitations. Hybrid A method combines DBSCAN and agglomerative clustering,
This paper
leveraging DBSCAN’s proposes two
ability hybrid
to define clustering
highly methods
separated clusterswith varyingoutliers
and identify attributes
withand in-
herent limitations.
unbounded Hybrid
parameters. A method
Another combines
advantage DBSCAN and
of incorporating agglomerative
DBSCAN clustering,
is the reduced
complexity in the parameter initialization, as it eliminates the need to specify the number
of partitions.
The Hybrid B method was specifically developed to optimize the clustering perfor-
mance based on the results of single-stage clustering analysis, combining the clustering
algorithms k-Means and agglomerative. The k-Means algorithm offers several advantages,
including easy implementation due to its relatively simple algorithm and linear time com-
plexity, denoted as O(n), where n represents the number of data objects [42]. However,
the k-Means algorithm requires the initialization of the number of clusters, which poses a
limitation in this data-driven study due to the absence of domain knowledge. Additionally,
the k-Means algorithm is sensitive to input data, initial seeds, and outliers, primarily due
to the crucial step of initializing the cluster centres.
Both methods share the advantages and drawbacks associated with hierarchical clus-
tering techniques employed for final clustering. One significant advantage is the flexibility
in determining the number of the partitions, as the hierarchy structure can be spliced at
the appropriate level to extract the clusters. However, hierarchical methods also have
limitations that can impact clustering performance. For instance, the
computational com-
plexity of agglomerative algorithms is considered quadratic, O n2 [43], which hampers
scalability for larger datasets. Additionally, the sequential approach of merging clusters in
the algorithm restricts the reassignment of samples once they are merged.
 in determining the number of the partitions, as the hierarchy structure can be spliced at
the appropriate level to extract the clusters. However, hierarchical methods also have lim-
itations that can impact clustering performance. For instance, the computational complex-
ity of agglomerative algorithms is considered quadratic, 𝑂(𝑛 ) [43], which hampers
Energies 2023, 16, 4440 scalability for larger datasets. Additionally, the sequential approach of merging 20 of 30 clusters
in the algorithm restricts the reassignment of samples once they are merged.
Hybrid A outperforms Hybrid B in terms of silhouette and Dunn indices, showing a
31.8% and 255%
Hybrid increase, respectively.
A outperforms The of
Hybrid B in terms high silhouette
silhouette andindex
Dunn demonstrates
indices, showing DBSCAN’s
a
proficiency in creating
31.8% and 255% well-definedThe
increase, respectively. clusters and removing
high silhouette noise points
index demonstrates in the subspace
DBSCAN’s
proficiency in creating
Incorporating well-defined
agglomeration clusters
in the finaland removing
clustering noiseresults
stage points in the subspace.
a slightly enhanced
Incorporating agglomeration in the final clustering stage results in a slightly
Dunn index, generating more comprehensive clusters. To validate the accuracy of the hy-enhanced
Dunnmethod’s
brid index, generating
ability tomore comprehensive
create clusters.
vortex-shedding To validate
maps, the accuracy
a comparison of thewith the
was made
hybrid method’s ability to create vortex-shedding maps, a comparison was made with the
reference map generated by Morse and Williamson [5]. Figure 17 illustrates the vortex-
reference map generated by Morse and Williamson [5]. Figure 17 illustrates the vortex-
shedding map produced using Hybrid A overlaid with the reference map’s regimes.
shedding map produced using Hybrid A overlaid with the reference map’s regimes.

17.Overlaid
Figure 17.
Figure Overlaidbenchmark regimes
benchmark on vortex-shedding
regimes map produced
on vortex-shedding with Hybrid
map produced with A.
Hybrid A.
The identified clusters are in line with the expected regions from Morse and Williamson [5],
The identified
with cluster clusters
2 and cluster arecorresponding
4 points in line with the expected
to the pure 2S regions from Morse
and 2P regions, and William-
respectively.
son [5], with cluster 2 and cluster 4 points corresponding to the pure
While the 2PO mode does not have a distinct cluster identification, vortex-shedding patterns 2S and 2P regions
respectively.
at this node canWhile
exhibitthe 2PO
both puremode does and
2P signals not intermittent
have a distinct
2PO cluster identification,
modes. Similar to all of vortex-
shedding patterns
the clustering methods at presented,
this nodevariation
can exhibit
in theboth pure
groups 2P signals
of clusters in theand intermittent
C (2S) regime 2PO
is observed. Cluster 1 strictly shows a regular sinusoidal pattern at low values
modes. Similar to all of the clustering methods presented, variation in the groups of clus- of non-
dimensional
ters in the Camplitude
(2S) regimeA∗ <is 0.2. The clusters
observed. identified
Cluster 1 at locations
strictly shows (λ
a
∗ , A∗ ) = (2, 0.3) and
regular sinusoidal pattern
(λ∗ , A∗ ) = (2, 0.5) show an unexpected 2P mode.
This common theme of unexpected 2P mode signals identified by the clustering
techniques in the C (2S) region may be attributed to the flow physics at this low-wavelength
case. The C (2S) mode is described by its synchronized 2S mode comprising small vortices
that coalesce in the near wake, forming large-scale structures downstream. This coalescence
and transformation of small vortical structures in the streamwise direction corrupt the
patterns in the data. In conclusion, the results of the data-driven methods for generating
vortex-shedding maps were determined to agree with the benchmark regime map produced
by Morse and Williamson [5]. The regime maps and corresponding clusters reveal the
underlying signatures of each mode based only on the local flow measurements of the
y-component of velocity. The cluster candidates used to plot the regime map provided more
precise distinctions between the modes with a satisfactory agreement with the reference
map, validating the use of clustering methods for generating vortex-shedding maps for
an oscillating cylinder at low Reynolds numbers. This section achieved its objective of
implementing a data-driven approach requiring less input data and supervision, relying
solely on local flow measurements and unsupervised clustering. The clustering methods
presented in this section are extended to produce vortex-shedding maps for high Reynolds
numbers and more complex vortex structures, which can be difficult to discern using
traditional methods.
Energies 2023, 16, 4440 21 of 30

5. Vortex-Shedding Map Generation at High Reynolds Number

This section applies unsupervised clustering methods validated in Section 4 for
the low-Reynolds-number case to a higher-Reynolds-number case to produce a vortex-
shedding map for more complex flow regimes. The performance of each clustering method
is evaluated based on both clustering accuracy and the quality of the resulting vortex-
shedding maps.

5.1. Introduction
This section presents the application of the validated unsupervised clustering strate-
gies to a case of high Reynolds number where the mapped domain is unknown to gain
insights into the complex flow regimes. The quality of the clustering analysis was evaluated
using internal clustering metrics, visual analysis of the clusters, and finally exploring the
patterns of the generated vortex-shedding maps. The main contributions of this section
include gaining insights into the underlying dynamical regimes of the vortex shedding
through map generation. Moreover, we also quantify the clustering performance to extract
meaningful patterns from the local flow field experiencing increased instability and mixing
due to the increased dissipation in the energy of the flow.
The high-Reynolds-number dataset [19] sampled the normalized amplitude–wavelength
plane of forced oscillations along five sampling lines. A subset of nodes was identified
as suitable for the clustering analysis based on the exhibited stable and synchronized
vortex-shedding behaviour. Nodes that did not exhibit stable vortex-shedding behaviour
or showed no synchronized pattern were excluded from the analysis to avoid adding noise
to the dataset. The nodes selected for the cluster analysis based on these23 features
Energies 2023, 16, x FOR PEER REVIEW of 32 are shown
in Figure 18.

Dataset sampled ∗ ∗
Figure18.18.
Figure Dataset sampled nodesnodes
in the in the normalized
normalized amplitudeamplitude (A )–wavelength
(𝐴∗ )–wavelength (𝜆∗ ) plane, with(λ ) plane, with
subpanels
subpanels showing
showing example
example signals
signals extracted
extracted from
from each each corresponding
corresponding node on the node on the map, denoted 1,
map, denoted
1,2,2,and
and 3.
3.

The repeated
The patterns
repeated in the
patterns inlocal flow measurements
the local were then
flow measurements isolated
were thenusing the using the sub-
isolated
subsequence extraction method. The quality of extracted subsequences for the high-Reyn-
sequence extraction method. The quality of extracted subsequences for the high-Reynolds-
olds-number cases is pivotal in the clustering results since more fluctuating signals are
number The
observed. cases is pivotal
window size in thefor
used clustering
the matrixresults
profile since more
algorithm fluctuating
was confirmed signals
for ap- are observed.
The window size used for the matrix profile algorithm was confirmed for applying the
plying the high-Reynolds-number case by visual inspection of a relatively consistent pat-
high-Reynolds-number
tern shown in Figure 19. case by visual inspection of a relatively consistent pattern shown in
Figure 19.
 The repeated patterns in the local flow measurements were then isolated using the
subsequence extraction method. The quality of extracted subsequences for the high-Reyn-
olds-number cases is pivotal in the clustering results since more fluctuating signals are
observed. The window size used for the matrix profile algorithm was confirmed for ap-
Energies 2023, 16, 4440 22 of 30
plying the high-Reynolds-number case by visual inspection of a relatively consistent pat-
tern shown in Figure 19.

Figure 19.
Figure 19. Motif
Motif extraction
extractionfor
forhigh-Reynolds-number
high-Reynolds-numbersignals of consistent
signals pattern
of consistent observed
pattern at at
observed
(𝜆∗∗, 𝐴∗ )∗= (0.5, 6).
(λ , A ) = (0.5, 6).

The
The window
windowsizesizeofof6464extracts
extractsmeaningful
meaningful patterns from
patterns fromthethe
sample signal
sample eveneven
signal withwith
unstable signals. Smaller window sizes would not capture the repeated global
unstable signals. Smaller window sizes would not capture the repeated global patterns, and patterns,
and larger
larger window
window sizessizes
wouldwouldnotnot extract
extract valuablepatterns
valuable patternstotoaid
aid in
in the
the clustering
clustering anal-
analysis.
ysis. Applying the clustering methods selected from the low-Reynolds-number
Applying the clustering methods selected from the low-Reynolds-number cases casesrequires
re-
quires the number of clusters to be specified. Since this study limits
the number of clusters to be specified. Since this study limits the domain knowledge the domain
knowledge required for map generation, the optimum number of clusters must be
required for map generation, the optimum number of clusters must be determined. The
number of clusters should optimize the clustering performance and the insights that can
be extracted through clustering. To ensure that the generated clusters can provide value
in the map generation, a limit of 10 clusters is specified. The upper bound adheres to the
rule of thumb
√ in cluster analysis, which approximates the number of clusters using the
formula n/2 for a training dataset of n instances [44], which yields an estimated limit of
11 clusters.

5.2. k-Means
The first step in implementing the k-Means method was determining the number of
clusters to consider for this case. Based on the balance of the silhouette and Dunn indices,
nine clusters were ultimately chosen. By balancing these metrics, we were able to select
a suitable number of both internally cohesive and well-separated clusters that allowed
for the clear identification of stable vortex-shedding signals while avoiding overfitting or
underfitting the data. The k-Means method validated in the previous section was imple-
mented using the same initialization method of k-Means++. The clustering performance
for the high-Reynolds-number case was quantified using the internal metrics summarized
in Table 6.

Table 6. Clustering performance metrics of k -Means method at Re = 10,000.

Evaluation Metric
Clustering Algorithm
Sil Dunn
k-Means 0.5750 0.49261

The evaluation metrics of the clusters indicate a good balance of high silhouette and
Dunn indices, implying that the clusters are both distinct from one another and internally
homogeneous. The corresponding clusters generated are shown in Figure 20.
From the generated clusters, some patterns begin to appear. Clusters 7 and 9 both
exhibit a regular sinusoidal variation with prominent peaks and troughs in the signal. A
similar pattern with low amplitude is observed in cluster 6. The expected pattern of the
2PO mode is observed in cluster 2 and cluster 4, highlighted by a smaller peak in between
the peaks. More irregular signals are highlighted in clusters 1, 3, 5, and 8. Despite the
variations in the signals of clusters 1 and 8, an underlying pattern persists of smaller dual
peaks. Clusters 3 and 5 exhibit a regular, albeit smaller-amplitude, sinusoidal variation of
the clusters.
 Evaluation Metric
Clustering Algorithm
Sil Dunn
𝑘-Means 0.5750 0.49261

The evaluation metrics of the clusters indicate a good balance of high silhouette and
Energies 2023, 16, 4440 23 of 30
Dunn indices, implying that the clusters are both distinct from one another and internally
homogeneous. The corresponding clusters generated are shown in Figure 20.

Energies 2023, 16, x FOR PEER REVIEW 25 of 32

the peaks. More irregular signals are highlighted in clusters 1, 3, 5, and 8. Despite the
variations in the signals of clusters 1 and 8, an underlying pattern persists of smaller dual
Subsequence
20. Subsequence
Figure Clusters
peaks. dataset
3 and 5dataset clustered
exhibitclustered
a regular, by
by 𝑘-Means
k-Means
albeit method
method at at
ReRe
smaller-amplitude, = sinusoidal
10,000,
= 10,000, with
with 𝑁 representing
Nvariation
representing
of the
the number
number of of samples
samples
the clusters. in in
eacheach cluster
cluster relative
relative to to
the the total
total dataset.
dataset.

5.3. Agglomerative
From the generated clusters, some patterns begin to appear. Clusters 7 and 9 both
5.3. Agglomerative
Thea regular
exhibit validated agglomerative method uses the samepeakscomplete linkage and cosine
The validatedsinusoidal variation
agglomerative method with prominent
uses the same complete and troughs
linkage andincosine
the signal.
af- A
affinity distance as that used for the low-Reynolds-number case. The internalpattern
indicesofused
finity distance as that used for the low-Reynolds-number case. The internal indices used the
similar pattern with low amplitude is observed in cluster 6. The expected
to
2POquantify
mode thethe
is clustering cluster
observed performance are summarized in Tablea7.smaller peak in between
to quantify clusteringinperformance
2 andare
cluster 4, highlighted
summarized in Tableby
7.
Table 7. Clustering performance metrics of agglomerative method at Re = 10,000.
Table 7. Clustering performance metrics of agglomerative method at Re = 10,000.
Evaluation
EvaluationMetric
Metric
Clustering Algorithm
Clustering Algorithm Sil Dunn
Sil Dunn
Agglomerative
Agglomerative 0.5760
0.5760 0.51160
0.51160

Theclusters
The clustersassociated
associatedwith
with the
the evaluation
evaluation metrics
metrics areare shown
shown in Figure
in Figure 21. 21.

Figure
Figure 21. Subsequence dataset clustered
21. Subsequence clustered by
by agglomerative
agglomerative (complete,
(complete,cosine)
cosine)method
methodatatReRe =
= 10,000,
10,000,
with Nwith 𝑁 representing
representing the number
the number of samples
of samples in cluster
in each each cluster relative
relative tototal
to the the total dataset.
dataset.

The clusters generated using the agglomerative procedure share many similarities
with the clusters of the 𝑘-Means method. A sub-peak in between cycles is synonymous
with the 2PO mode and is identified in clusters 1 and 3. Pure modes were identified in
clusters 1, 3, 5, 7, and 9, with slight variation in the samples. Inconsistent patterns are
Energies 2023, 16, 4440 24 of 30

The clusters generated using the agglomerative procedure share many similarities
with the clusters of the k-Means method. A sub-peak in between cycles is synonymous with
the 2PO mode and is identified in clusters 1 and 3. Pure modes were identified in clusters 1,
3, 5, 7, and 9, with slight variation in the samples. Inconsistent patterns are observed in
clusters 4 and 5, where the clusters would benefit from additional merging.

5.4. Hybrid A
The hybrid methods validated in the low-Reynolds-number case are compared based
on internal evaluation metrics, cluster plots, and the generated vortex-shedding map. The
number of clusters in the pre-clustering phase is not required to be determined since the
DBSCAN algorithm automatically finds the optimum number and corresponding noise
Energies 2023, 16, x FOR PEER REVIEW 26 of 32
points. The algorithm found 14 separate clusters and 30 noise points. The clustering
performance results of both phases are summarized in Table 8.

Table8.8. Clustering
Table Clusteringperformance
performancemetrics
metricsof
ofHybrid
HybridAAmethod
methodat
atRe
Re== 10,000.
10,000.

Clustering Algorithm
Evaluation
Evaluation Metric
Metric
Phase
Phase Clustering Algorithm Clusters
Clusters Sil
Sil Dunn
Dunn
Pre-Clustering
Pre-Clustering DBSCAN
DBSCAN 14 14
(30(30
Noise)
Noise) 0.7250
0.7250 0.20279
0.20279
Final
FinalClustering
Clustering Agglomerative
Agglomerative 6 6 0.4822
0.4822 0.31561
0.31561

The
The generated
generated clusters
clusters merged
mergedfor
forthe
theentire
entiredataset
datasetare
areshown
shownin
inFigure
Figure22.
22.

Figure 22. Subsequence

Figure 22. Subsequence dataset
dataset clustered
clustered by
by Hybrid 𝑁 representing
Hybrid A method at Re = 10,000, with N representing
the number
the numberof ofsamples
samplesin
ineach
eachcluster
clusterrelative
relativeto
tothe
thetotal
totaldataset.
dataset.

The
Thecluster
clustersamples
samplesinineach
eachlabel demonstrated
label demonstratedthethe
hybrid method
hybrid methodB’s B’s
ability to extract
ability to ex-
similar shapeshape
tract similar patterns. Cluster
patterns. 3 contains
Cluster out-of-phase
3 contains samples,
out-of-phase but thebut
samples, similarity in shape
the similarity in
is observed
shape between
is observed the patterns.
between the patterns.
5.5. Hybrid B
5.5. Hybrid B
The pre-clustering phase involved determining the optimum number of clusters,
The pre-clustering phase involved determining the optimum number of clusters,
which was achieved by selecting the number that maximized the silhouette index. Through
which was achieved by selecting the number that maximized the silhouette index.
this approach, the optimal number of clusters was found to be 30, as it maximized the
Through this approach, the optimal number of clusters was found to be 30, as it maxim-
separation of the clusters in the first stage. The clustering performance results of both
ized the separation of the clusters in the first stage. The clustering performance results of
phases are summarized in Table 9.
both phases are summarized in Table 9.

Table 9. Clustering performance metrics of Hybrid B method at Re = 10,000.

Evaluation Metric
Phase Clustering Algorithm Clusters
 Energies 2023, 16, 4440 25 of 30

Table 9. Clustering performance metrics of Hybrid B method at Re = 10,000.

Evaluation Metric
Phase Clustering Algorithm Clusters
Sil Dunn
Pre-Clustering k-Means 30 0.8693 0.05395
Final Clustering Agglomerative 9 0.4694 0.28585

Energies 2023, 16, x FOR PEER REVIEW 27 of 3

In the final cluster phase, the merged cluster labels produced the subsequence clusters
shown in Figure 23.

Figure 23.Subsequence
Figure23. Subsequence dataset clustered
dataset by Hybrid
clustered by HybridB method at Re =at10,000,
B method with N with
Re = 10,000, 𝑁 representing
representing
the
thenumber
numberofofsamples in each
samples cluster
in each relative
cluster to thetototal
relative thedataset.
total dataset.

The patterns of interest in the generated clusters include the resemblance of a 2S mode
The patterns of interest in the generated clusters include the resemblance of a 2S
for clusters 5 and 8. Clusters 1, 3, and 6 include more irregular signals, but an underlying
patternfor
mode clustersdual
of smaller 5 and 8. Clusters
peaks 1, 3, and 6 include
can be distinguished. more
Finally, irregular
cluster numberssignals, but an un
4, 7, and
9derlying pattern
exhibit strong 2POofbehaviour.
smaller dual peaks can be distinguished. Finally, cluster numbers 4, 7
and 9 exhibit strong 2PO behaviour.
5.6. Discussion
5.6. Discussion
The validated clustering methods were implemented for the case of a high Reynolds
number. The clustering method’s ability to produce separate and compact clusters was
The validated clustering methods were implemented for the case of a high Reynold
accessed by the internal evaluation metrics summarized in Table 10.
number. The clustering method’s ability to produce separate and compact clusters was
accessed
Table by clustering
10. Final the internal evaluation
performance metrics
metrics summarized
of proposed methods atinRe
Table 10.
= 10,000.

Table 10.Type
Final clustering performance Evaluation
metrics of proposed methods Metric
at Re = 10,000.
Clustering Algorithm
Sil Dunn
Partitioning k-Means Algorithm 0.5750 Evaluation
0.49261Metric
Type Clustering
Hierarchical Agglomerative 0.5760 Sil 0.51160Dunn
Partitioning
Hybrid A 𝑘-Means
DBSCAN/Agglomerative 0.48220.5750 0.49261
0.31561
Hybrid B k-Means/Agglomerative 0.4694 0.28585
Hierarchical Agglomerative 0.5760 0.51160
Hybrid A DBSCAN/Agglomerative 0.4822 0.31561
InHybrid
terms ofB the evaluation metrics of the silhouette
𝑘-Means/Agglomerative and Dunn indices,
0.4694 the single-stage
0.28585
clustering methods performed better than the hybrid methods. The reduced clustering
performance of the hybrid methods is attributed to the use of dynamic time warping (DTW)
In terms of the evaluation metrics of the silhouette and Dunn indices, the single-stage
clustering methods performed better than the hybrid methods. The reduced clustering
performance of the hybrid methods is attributed to the use of dynamic time warping
(DTW) in the final clustering phase, which groups signals based on shape rather than time
Energies 2023, 16, 4440 26 of 30

in the final clustering phase, which groups signals based on shape rather than time. The
pairwise distance calculation in the silhouette and Dunn indices will score time series
poorly if the patterns are out of phase, even if the shape is similar in the cluster.
However, the clusters produced using the hybrid methods are more similar based
on shape than the single-stage methods. The similarity in shape of the hybrid methods
will yield better vortex-shedding maps based on signature shapes. The improved vortex-
Energies 2023, 16, x FOR PEER REVIEW shedding maps are observed in the plotted domain of non-dimensional amplitude 28 of 32 and
wavelength. The maps generated using the traditional methods display more scattered and
irregular patterns, which results in the need for a larger number of clusters to capture the full
range of vortex-shedding behaviour. In contrast, the proposed ensemble method produces
produces more coherent and consistent maps, allowing for a more comprehensive under-
more coherent and consistent maps, allowing for a more comprehensive understanding of
standing of the flow dynamics with a smaller number of clusters.
theDespite
flow dynamics
the lowerwith a smaller
overall number
evaluation of clusters.
metrics, it was determined that hybrid methods
Despite the lower overall evaluation metrics,
outperformed single-stage methods in the quality of the it was determined
clusters based onthat
the hybrid methods
similarity
of outperformed
shape and the single-stage
lower number methods in therequired
of clusters quality toof represent
the clusters
thebased
data.on
Thethevarious
similarity of
regions of vortex-shedding behaviour at high Reynolds numbers can be obtained by ana-regions
shape and the lower number of clusters required to represent the data. The various
of vortex-shedding
lyzing behaviour
the produced cluster maps.at high
The Reynolds numbers
non-dimensional can be
amplitude obtained
and by analyzing
wavelength plane the
produced cluster maps. The non-dimensional amplitude and wavelength
populated with the corresponding clusters derived using Hybrid A method is shown in plane populated
with 24.
Figure the corresponding clusters derived using Hybrid A method is shown in Figure 24.

Figure 24. Vortex-shedding map regions using the Hybrid A method at Re = 10,000.
Figure 24. Vortex-shedding map regions using the Hybrid A method at Re = 10,000.
The coloured regions only visualize similar vortex-shedding behaviour determined by
The coloured
cluster number. Tworegions only visualize
independent similar
regions vortex-shedding
are located behaviour
at the top of the map,determined
(λ∗ , ∗A∗∗) = (4, 0.9)
byand (λ , A )∗ = ∗(6, 0.9), with consistent vortex-shedding signals. The former(𝜆node
cluster∗ number.
∗ Two independent regions are located at the top of the map, , 𝐴 )signal
= is
(4,periodic (𝜆 , 𝐴distinct
0.9) and with ) = (6, peaks
0.9), with consistent vortex-shedding signals. The former
that resemble the 2S mode. The second node signal, denoted node
signal is periodic with distinct peaks that resemble the 2S mode. The second node signal,
with cluster number 1, indicates the 2PO mode with a subpeak in between the relative
denoted with cluster number 1, indicates the 2PO mode with a subpeak in between the
peaks generated by the weaker vortex structure. The samples in cluster 1 include samples
relative peaks generated by the weaker vortex structure. The samples in cluster 1 include
closer to that of cluster 5 located at the previous node, which indicates a level of overlap
samples closer to that of cluster 5 located at the previous node, which indicates a level of
between the vortex structures. A larger group of nodes with the same cluster number was
overlap between the vortex structures. A larger group of nodes with the same cluster num-
identified in the middle of the map, 3 < λ∗ < 7 and 0.2 < A∗ < 0.8. Cluster 3 follows
ber was identified in the middle of the map, 3 < 𝜆∗ < 7 and 0.2 < 𝐴∗ < 0.8. Cluster 3 fol-
differing variations of 2PO and 2P of dual peak and sub-peaks in oscillating actions. The
lows differing variations of 2PO and 2P of dual peak and sub-peaks in oscillating actions.
region defined by cluster 2 resembles the signal of the P + S mode determined from the
The region defined by cluster 2 resembles the signal of the P + S mode determined from
thesmoothed
smootheddualdualpeak
peakofofthe
theoppositely
oppositelysigned
signedvortices
vorticesofofthe
thePPmode
modeandandthe
thesingle
singlepeak of
the S mode passing the sensor.
peak of the S mode passing the sensor.
The cluster map produced using the Hybrid B method also outlines various regions
of similar fluid flow behaviour. The vortex-shedding map populated with the correspond-
ing clusters derived using the Hybrid B method is shown in Figure 25.
 Energies 2023, 16, 4440 27 of 30

Energies 2023, 16, x FOR PEER REVIEW 29 of 32

The cluster map produced using the Hybrid B method also outlines various regions of
similar fluid flow behaviour. The vortex-shedding map populated with the corresponding
clusters derived using the Hybrid B method is shown in Figure 25.

Figure 25. Vortex-shedding map regions using the Hybrid B method at Re = 10,000.
Figure 25. Vortex-shedding map regions using the Hybrid B method at Re = 10,000.
The coherent vortex-shedding patterns identified in the map first include the 2S
behaviour
The at the node
coherent (λ∗ , A∗ ) = (4, 0.9).
vortex-shedding Nearby,
patterns two strong
identified 2POmap
in the clusters
firstwere identified
include the 2S be-
∗ ∗
at (λ , A ) = (6, 0.9) and ( ∗ , A∗ ) = (4, 0.7), demonstrated by a weaker vortex being shed in
∗ ∗ λ
haviour at the node (𝜆 , 𝐴 ) = (4, 0.9). Nearby, two strong 2PO clusters were identified at
between the oscillations. ∗ The∗ former point of cluster 9 shares the similarity in shape but
(𝜆∗at ∗
, 𝐴high
) = (6, 0.9) and (𝜆 , 𝐴
peak amplitudes compared ) = (4, 0.7), demonstrated
to cluster 4. The largest byregion
a weaker vortex
of similar being shed
behaviour is in
between
identified thebyoscillations. The former
cluster 3, which dominatespointtheofhigher
cluster 9 shares the
wavelength similarity
portion of thein shape but at
subspace.
high peak amplitudes compared to cluster 4. The largest region
The signals in cluster 3 demonstrate the behaviour of 2P but appear to have less defined of similar behaviour is
identified
twin peaks. by cluster
Finally,3,a which
region in dominates the higher
low-amplitude nodeswavelength
was identified portion of thethe
resembling subspace.
2S
Themodesignalsat lower amplitudes
in cluster compared to
3 demonstrate that
the of node (λof∗ , 2P
behaviour A∗ )but
= (4,appear
0.9). to have less defined
twin peaks. From aFinally,
flow physics
a regionperspective, the vortex-shedding
in low-amplitude nodes was maps at high Reynolds
identified resemblingnum- the 2S
bers provide novel insights into the underlying vortex–structure
mode at lower amplitudes compared to that of node (𝜆 , 𝐴 ) = (4, 0.9). ∗ ∗ interactions. The higher
Reynolds number and associated higher flow energy seemed to produce more variation sig-
From a flow physics perspective, the vortex-shedding maps at high Reynolds num-
nals and a stronger dissipation effect. The dissipation effect is a product of a more turbulent
bers provide novel insights into the underlying vortex–structure interactions. The higher
flow, which increases flow mixing and creates a more homogenous flow. The signals ap-
Reynolds numberwith
peared smooth, andsamples
associated higher
with twin flow
peaks notenergy seemed
as defined as theto produce more variation
low-Reynolds-number
signals
case at and a stronger
4000. dissipation
Furthermore, the higheffect.
ReynoldsThenumber
dissipation effect is atoproduct
was observed create moreof anoise
more tur-
bulent
in theflow, which increases
vortex-shedding map with flowan mixing andnumber
increased createsof a nodes
more homogenous
with no visibleflow. The sig-
patterns
due to the coalescence of vortices. Overall, the signals of the
nals appeared smooth, with samples with twin peaks not as defined as the low-Reynolds-high-Reynolds-number case
were ofcase
number the larger amplitude
at 4000. Furthermore,of the y-component of velocity
the high Reynolds measurement.
number was observedSpecifically,
to create
more noise in the vortex-shedding map with an increased number of nodes with noinvisible
the 2S behaviour showed large amplitudes between peaks in the signals, which aided
identifying these clusters.
patterns due to the coalescence of vortices. Overall, the signals of the high-Reynolds-num-
The high-Reynolds-number case has clear regions of the 2PO mode demonstrated by
berthecase were of the larger amplitude of the 𝑦-component of velocity measurement. Spe-
signals of clusters 4 and 9, which show the two pairs of vortices being shed per cycle
cifically,
with one the 2S behaviour
vortex showedmuch
in each oscillation largeweaker.
amplitudes between
The transition peaks
mode inclose
is in the signals,
proximitywhich
aided
to the in observed
identifying P + these
S signal clusters.
behaviour of cluster 3 demonstrated by the dual peak of the P
modeThebeinghigh-Reynolds-number
separated with a single case hasfrom
peak cleartheregions
S mode. of The
the source
2PO mode of thedemonstrated
P + S mode by
the signals of clusters 4 and 9, which show the two pairs of vortices being shed per cycle
with one vortex in each oscillation much weaker. The transition mode is in close proximity
to the observed P + S signal behaviour of cluster 3 demonstrated by the dual peak of the
P mode being separated with a single peak from the S mode. The source of the P + S mode
Energies 2023, 16, 4440 28 of 30

may be attributed to the increased turbulent kinetic energy of a high Reynolds number,
which is decomposing 2P modes shed close to the cylinder into a single P and S mode.
The devolution of the 2P mode would have to be rapid as the sampling line of the dataset
is located at a distance of 4D in the wake of the cylinder. If the decomposition of the 2P
mode is the mechanism in which the P + S mode appears in the dataset, it would have a
negligible impact on the overall behaviour of the cylinder. The negligible impact is because
the observed rapid decay implies that the P + S mode dissipates quickly and does not
contribute significantly to the long-term vortex-shedding dynamics of the cylinder.
In conclusion, Hybrid A and B methods were found to be the most effective in gen-
erating clusters with high similarity of shape, outperforming the single-stage methods.
The vortex-shedding maps produced by these hybrid methods reveal valuable informa-
tion about the system’s dynamics. Notably, the Reynolds number case of 10,000 exhibits
similar vortex-shedding modes as the low-Reynolds-number dataset, including the 2S and
2PO modes. The region of the map previously inhabited with 2P modes was observed to
comprise the majority of P + S modes. Overall, the flow physics derived from the cluster
analysis demonstrates the increased dissipation effect in a high Reynolds flow, resulting in
a smoothing effect on the flow signals.

6. Conclusions
This study developed a data-driven approach for generating vortex-shedding maps of
a circular cylinder undergoing forced vibration. The unsupervised clustering of local flow
measurement time series subsequences reproduced the benchmark vortex-shedding map at
Reynolds number 4000. The clustering method developed herein was then applied to a high-
Reynolds-number flow to generate a vortex-shedding map that had not previously existed.
The procedure of generating vortex-shedding maps using novel unsupervised sub-
sequence clustering methods was presented and validated for the case of a low Reynolds
number of 4000 in Section 4. Several clustering methods were selected and compared
for the clustering task of subsequences extracted from the y-component of the velocity
(uy ) time series data. The clustering results of the proposed methods demonstrated their
ability to extract meaningful clusters that represented the underlying flow physics of the
varying modes. The application of the clustering analysis to produce regime maps provided
satisfactory agreement with the reference map by Morse and Williamson [5].
The proposed methods were then extended in Section 5 to quantify their performance
to produce vortex-shedding maps at high Reynolds numbers with more complex vortex
structures. The hybrid methods were the most effective in generating clusters with high
similarity of shape, outperforming the single-stage methods. The vortex-shedding maps
produced at a high Reynolds number provided novel insights into the underlying vortex
structure interactions and identified numerous regions of similar patterns despite increased
instability. The clustering procedure in the generation of vortex-shedding maps offers a
method that requires less data and supervision without resolving the entire flow field. The
method was implemented on the vortex-shedding patterns of an oscillating cylinder, but
the method could create value for any vortex-shedding behaviour.
Possible future work for this study includes a more comprehensive sampling of the
normalized amplitude–wavelength plane, specifically in the transition zones, to provide
a higher resolution of the cluster regions. Furthermore, incorporating real-world experi-
mental data or higher resolution numerical data derived from large-eddy simulations (LES)
to validate and expand upon our results would also be valuable, enabling a more robust
assessment of the procedure. Optimizing the computational efficiency of the proposed
clustering approach and exploring its generalizability to different geometric configurations,
flow conditions, and Reynolds numbers would enhance its applicability. Furthermore,
integrating the clustering results into system design and optimization processes would
provide insights into the impact of different flow regimes on the efficiency and stability of
VIV wind energy harvesting systems.
Energies 2023, 16, 4440 29 of 30

Author Contributions: Conceptualization, M.C.; methodology, M.C.; software, M.C.; validation,

M.C. and R.M.; formal analysis, M.C.; investigation, R.M.; resources, F.-S.L.; data curation, R.M.;
writing—original draft preparation, M.C.; writing—review and editing, M.C. and E.Y.; visualization,
M.C. and R.M.; supervision, F.-S.L. and W.M.; project administration, F.-S.L. and W.M.; funding
acquisition, F.-S.L. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the Natural Sciences and Engineering Research Council of
Canada (NSERC) Discovery Grants program Grant Nos. 50503-10234 and 50503-10925.
Data Availability Statement: Research data generated during this study that support the reported
results can be found at https://www.kaggle.com/datasets/ryleymcconkey/vortex-shedding-wake
(accessed on 16 October 2021).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Fan, D.; Wu, B.; Bachina, D.; Triantafyllou, M.S. Vortex-induced vibration of a piggyback pipeline half buried in the seabed. J.
Sound Vib. 2019, 449, 182–195. [CrossRef]
2. Kumar, N.; Kolahalam, V.K.V.; Kantharaj, M.; Manda, S. Suppression of vortex-induced vibrations using flexible shrouding—An
experimental study. J. Fluids Struct. 2018, 81, 479–491. [CrossRef]
3. Wang, W.; Wang, X.; Hua, X.; Song, G.; Chen, Z. Vibration control of vortex-induced vibrations of a bridge deck by a single-side
pounding tuned mass damper. Eng. Struct. 2018, 173, 61–75. [CrossRef]
4. Wu, Y.; Cheng, Z.; McConkey, R.; Lien, F.-S.; Yee, E. Modelling of Flow-Induced Vibration of Bluff Bodies: A Comprehensive
Survey and Future Prospects. Energies 2022, 15, 8719. [CrossRef]
5. Morse, T.L.; Williamson, C.H.K. Fluid forcing, wake modes, and transitions for a cylinder undergoing controlled oscillations. J.
Fluids Struct. 2009, 25, 697–712. [CrossRef]
6. Williamson, C.H.K.; Roshko, A. Vortex formation in the wake of an oscillating cylinder. J. Fluids Struct. 1988, 2, 355–381. [CrossRef]
7. Wu, W.; Bernitsas, M.M.; Maki, K. RANS Simulation Versus Experiments of Flow Induced Motion of Circular Cylinder With
Passive Turbulence Control at 35,000 < Re < 130,000. J. Offshore Mech. Arct. Eng. 2014, 136, 041802. [CrossRef]
8. Zhang, D.; Wang, W.; Sun, H.; Bernitsas, M.M. Influence of turbulence intensity on vortex pattern for a rigid cylinder with
turbulence stimulation in flow induced oscillations. Ocean Eng. 2021, 237, 109349. [CrossRef]
9. Bernitsas, M.; Ben-Simon, Y.; Raghavan, K.; Garcia, E. The VIVACE Converter: Model tests at high damping and Reynolds
number around 105. J. Offshore Mech. Arct. Eng. 2009, 131, 011102. [CrossRef]
10. Païdoussis, M.P.; Price, S.J.; de Langre, E. 3. Vortex-Induced Vibrations. In Fluid-Structure Interactions—Cross-Flow-Induced
Instabilities; Cambridge University Press: Cambridge, UK, 2010; Available online: https://app.knovel.com/hotlink/pdf/id:
kt008NA8H1/fluid-structure-interactions/two-dimensional-viv-phenomenology (accessed on 25 October 2021).
11. Techet, A. “13.42 Lecture: Vortex Induced Vibrations,” MIT OCW, Apr. 21, 2005. Available online: https://ocw.mit.edu/courses/
mechanical-engineering/2-22-design-principles-for-ocean-vehicles-13-42-spring-2005/readings/lec20_viv1.pdf (accessed on 18
October 2021).
12. Triantafyllou, M.S.; Bourguet, R.; Dahl, J.; Modarres-Sadeghi, Y. Vortex-Induced Vibrations In Springer Handbook of Ocean Engineering,
1st ed.; Springer International Publishing: Berlin/Heidelberg, Germany, 2016; pp. 819–850.
13. Yang, W.; Masroor, E.; Stremler, M.A. The wake of a transversely oscillating circular cylinder in a flowing soap film at low
Reynolds number. J. Fluids Struct. 2021, 105, 103343. [CrossRef]
14. Williamson, C.H.K.; Govardhan, R. Vortex-induced vibrations. Annu. Rev. Fluid Mech. 2004, 36, 413–455. [CrossRef]
15. Cann, M.; McConkey, R.; Lien, F.-S.; Melek, W.; Yee, E. Mode classification for vortex shedding from an oscillating wind turbine
using machine learning. J. Phys. Conf. Ser. 2021, 2141, 12009. [CrossRef]
16. Huera-Huarte, F.J.; Vernet, A. Vortex modes in the wake of an oscillating long flexible cylinder combining POD and fuzzy
clustering. Exp. Fluids 2010, 48, 999–1013. [CrossRef]
17. Menon, K.; Mittal, R. Quantitative analysis of the kinematics and induced aerodynamic loading of individual vortices in
vortex-dominated flows: A computation and data-driven approach. J. Comput. Phys. 2021, 443, 110515. [CrossRef]
18. Calvet, A.G.; Dave, M.; Franck, J.A. Unsupervised clustering and performance prediction of vortex wakes from bio-inspired
propulsors. Bioinspir. Biomim. 2021, 16, 046015. [CrossRef]
19. McConkey, R. Vortex Shedding in a Turbulent Wake. Available online: https://www.kaggle.com/ryleymcconkey/vortex-
shedding-wake (accessed on 16 October 2021).
20. ESI OpenCFD Releases OpenFOAM v2006. Available online: https://www.openfoam.com/news/main-news/openfoam-v20-06
(accessed on 27 April 2023).
21. Kinaci, O. 2-D URANS Simulations of Vortex Induced Vibrations of Circular Cylinder at Trsl3 Flow Regime. J. Appl. Fluid Mech.
2016, 9, 2537–2544. [CrossRef]
Energies 2023, 16, 4440 30 of 30

22. Khan, N.B.; Ibrahim, Z.; Nguyen, L.T.T.; Javed, M.F.; Jameel, M. Numerical investigation of the vortex-induced vibration of an
elastically mounted circular cylinder at high Reynolds number (Re = 104) and low mass ratio using the RANS code. PLoS ONE
2017, 12, e0185832. [CrossRef] [PubMed]
23. Zhao, M.; Cheng, L. Vortex-induced vibration of a circular cylinder of finite length. Phys. Fluids 2014, 26, 015111. [CrossRef]
24. Menter, F.R. Two-equation eddy-viscosity turbulence models for engineering applications. AIAA J. 1994, 32, 1598–1605. [CrossRef]
25. Colvert, B.; Alsalman, M.; Kanso, E. Classifying vortex wakes using neural networks. Bioinspir. Biomim. 2018, 13, 025003.
[CrossRef]
26. Wang, M.; Hemati, M.S. Detecting exotic wakes with hydrodynamic sensors. Theor. Comput. Fluid Dyn. 2019, 33, 235–254.
[CrossRef]
27. Alsalman, M.; Colvert, B.; Kanso, E. Training bioinspired sensors to classify flows. Bioinspir. Biomim. 2018, 14, 016009. [CrossRef]
[PubMed]
28. Keogh, E.; Lin, J. Clustering of time-series subsequences is meaningless: Implications for previous and future research. Knowl. Inf.
Syst. 2005, 8, 154–177. [CrossRef]
29. Mörchen, F.; Ultsch, A.; Hoos, O. Extracting interpretable muscle activation patterns with time series knowledge mining. Int. J.
Knowl. -Based Intell. Eng. Syst. 2005, 9, 197–208. [CrossRef]
30. Yeh, C.-C.M.; Zhu, Y.; Ulanova, L.; Begum, N.; Ding, Y.; Dau, H.A.; Silva, D.F.; Mueen, A.; Keogh, E. Matrix Profile I: All Pairs
Similarity Joins for Time Series: A Unifying View That Includes Motifs, Discords and Shapelets. In Proceedings of the 2016 IEEE
16th International Conference on Data Mining (ICDM), Barcelona, Spain, 12–15 December 2016; pp. 1317–1322. [CrossRef]
31. Liu, Y.; Li, Z.; Xiong, H.; Gao, X.; Wu, J. Understanding of Internal Clustering Validation Measures. In Proceedings of the 2010
IEEE International Conference on Data Mining, Sydney, Australia, 13–17 December 2010; pp. 911–916. [CrossRef]
32. Hennig, C.; Meila, M.; Murtagh, F.; Rocci, R. (Eds.) Chapter 26: Method-Independent Indices for Cluster Validation and Estimating
the Number of Clusters. In Handbook of Cluster Analysis; Chapman and Hall/CRC: New York, NY, USA, 2015. [CrossRef]
33. Rousseeuw, P.J. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Math. 1987,
20, 53–65. [CrossRef]
34. Dunn, J.C. Well-Separated Clusters and Optimal Fuzzy Partitions. J. Cybern. 1974, 4, 95–104. [CrossRef]
35. Aghabozorgi, S.; Shirkhorshidi, A.S.; Wah, T.Y. Time-series clustering—A decade review. Inf. Syst. 2015, 53, 16–38. [CrossRef]
36. Arthur, D.; Vassilvitskii, S. K-Means++: The Advantages of Careful Seeding. In Proceedings of the Eighteenth Annual ACM-SIAM
Symon Discrete Algorithms, New Orleans, LA, USA, 7 January 2007; pp. 1027–1035. [CrossRef]
37. Lai, C.-P.; Chung, P.-C.; Tseng, V.S. A novel two-level clustering method for time series data analysis. Expert Syst. Appl. 2010, 37,
6319–6326. [CrossRef]
38. Zhang, X.; Liu, J.; Du, Y.; Lv, T. A novel clustering method on time series data. Expert Syst. Appl. 2011, 38, 11891–11900. [CrossRef]
39. Aghabozorgi, S.; Wah, T.Y.; Herawan, T.; Jalab, H.A.; Shaygan, M.A.; Jalali, A. A Hybrid Algorithm for Clustering of Time Series
Data Based on Affinity Search Technique. Sci. World J. 2014, 2014, e562194. [CrossRef]
40. Aghabozorgi, S.; Teh, Y.W. Stock market co-movement assessment using a three-phase clustering method. Expert Syst. Appl. 2014,
41, 1301–1314. [CrossRef]
41. Ester, M.; Kriegel, H.-P.; Xu, X. A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise; AAAI Press:
Portland, OR, USA, 1996; pp. 226–231.
42. Fred, A.L.N.; Leitão, J.M.N. Partitional vs Hierarchical Clustering Using a Minimum Grammar Complexity Approach. In
Advances in Pattern Recognition; Ferri, F.J., Iñesta, J.M., Amin, A., Pudil, P., Eds.; Lecture Notes in Computer Science 1876; Springer:
Berlin/Heidelberg, Germany, 2000; pp. 193–202. [CrossRef]
43. Contreras, P.; Murtagh, F. (Eds.) Chapter 6: Hierarchical Clustering. In Handbook of Cluster Analysis; Chapman and Hall/CRC:
New York, NY, USA, 2015. [CrossRef]
44. Han, J. Data Mining Concepts and Techniques. In The Morgan Kaufmann Series in Data Management Systems, 3rd ed.; Elsevier:
Burlington, MA, USA, 2012.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

View publication stats