Applied Mathematics

to Mechanisms
and Machines
Edited by
Higinio Rubio Alonso, Alejandro Bustos Caballero,
Jesus Meneses Alonso and Enrique Soriano-Heras
Printed Edition of the Special Issue Published in Mathematics

www.mdpi.com/journal/mathematics
Applied Mathematics to Mechanisms
and Machines
Applied Mathematics to Mechanisms
and Machines

Editors
Higinio Rubio Alonso
Alejandro Bustos Caballero
Jesus Meneses Alonso
Enrique Soriano-Heras

MDPI • Basel • Beijing • Wuhan • Barcelona • Belgrade • Manchester • Tokyo • Cluj • Tianjin
Editors
Higinio Rubio Alonso Alejandro Bustos Caballero Jesus Meneses Alonso
Department of Mechanical Department of Mechanics, Department of Mechanical
Engineering, Universidad National University of Engineering, Universidad
Carlos III de Madrid, Distance Education, Carlos III de Madrid,
Leganés, Spain Madrid, Spain Leganés, Spain

Enrique Soriano-Heras
Department of Mechanical
Engineering, Universidad
Carlos III de Madrid,
Leganés, Spain

Editorial Office
MDPI
St. Alban-Anlage 66
4052 Basel, Switzerland

This is a reprint of articles from the Special Issue published online in the open access journal
Mathematics (ISSN 2227-7390) (available at: https://www.mdpi.com/journal/mathematics/special
issues/Applied Mathematics to Mechanisms Machines).

For citation purposes, cite each article independently as indicated on the article page online and as
indicated below:

LastName, A.A.; LastName, B.B.; LastName, C.C. Article Title. Journal Name Year, Volume Number,
Page Range.

ISBN 978-3-0365-5923-0 (Hbk)

ISBN 978-3-0365-5924-7 (PDF)

© 2022 by the authors. Articles in this book are Open Access and distributed under the Creative
Commons Attribution (CC BY) license, which allows users to download, copy and build upon
published articles, as long as the author and publisher are properly credited, which ensures maximum
dissemination and a wider impact of our publications.
The book as a whole is distributed by MDPI under the terms and conditions of the Creative Commons
license CC BY-NC-ND.
Contents

Preface to ”Applied Mathematics to Mechanisms and Machines” . . . . . . . . . . . . . . . . . vii

Alfonso Hernández, Aitor Muñoyerro, Mónica Urı́zar and Enrique Amezua

Hybrid Optimization Based Mathematical Procedure for Dimensional Synthesis of Slider-Crank
Linkage
Reprinted from: Mathematics 2021, 9, 1581, doi:10.3390/math9131581 . . . . . . . . . . . . . . . . 1

Oscar Altuzarra, David Manuel Solanillas, Enrique Amezua and Victor Petuya
Path Analysis for Hybrid Rigid–Flexible Mechanisms
Reprinted from: Mathematics 2021, 9, 1869, doi:10.3390/math9161869 . . . . . . . . . . . . . . . . 19

Zhiqi Yan, Shisheng Zhong, Lin Lin and Zhiquan Cui

Adaptive Levenberg–Marquardt Algorithm: A New Optimization Strategy for
Levenberg–Marquardt Neural Networks
Reprinted from: Mathematics 2021, 9, 2176, doi:10.3390/math9172176 . . . . . . . . . . . . . . . . 45

Minh Tuan Pham, Song Huat Yeo and Tat Joo Teo
Three-Legged Compliant Parallel Mechanisms: Fundamental Design Criteria to Achieve Fully
Decoupled Motion Characteristics and a State-of-the-Art Review
Reprinted from: Mathematics 2022, 10, 1414, doi:10.3390/math10091414 . . . . . . . . . . . . . . . 63

Alejandro Bustos, Jesus Meneses, Higinio Rubio and Enrique Soriano-Heras

Predictive Suspension Algorithm for Land Vehicles over Deterministic Topography
Reprinted from: Mathematics 2022, 10, 1467, doi:10.3390/math10091467 . . . . . . . . . . . . . . . 93

Zhiqi Yan, Ming Zu, Zhiquan Cui and Shisheng Zhong

A Novel Denoising Method for Retaining Data Characteristics Brought from Washing
Aeroengines
Reprinted from: Mathematics 2022, 10, 1485, doi:10.3390/math10091485 . . . . . . . . . . . . . . . 113

Gerardo Amato, Roberto D’Amato and Alessandro Ruggiero

Adaptive Rejection of a Sinusoidal Disturbance with Unknown Frequency in a Flexible Rotor
with Lubricated Journal Bearings
Reprinted from: Mathematics 2022, 10, 1703, doi:10.3390/math10101703 . . . . . . . . . . . . . . . 135

Marı́a T. Orvañanos-Guerrero, Mario Acevedo, Claudia N. Sánchez, Daniel U.

Campos-Delgado, Amir Aminzadeh Ghavifekr, Paolo Visconti and Ramiro Velázquez
Complete Balancing of the Six-Bar Mechanism Using Fully Cartesian Coordinates and
Multiobjective Differential Evolution Optimization
Reprinted from: Mathematics 2022, 10, 1830, doi:10.3390/math10111830 . . . . . . . . . . . . . . . 161

Isiah Zaplana, Hugo Hadfield and Joan Lasenby

Singularities of Serial Robots: Identification and Distance Computation Using Geometric
Algebra
Reprinted from: Mathematics 2022, 10, 2068, doi:10.3390/math10122068 . . . . . . . . . . . . . . . 185

Jesús Alan Calderón Chavarri, Carlos Gianpaul Rincón Ruiz, Ana Marı́a Gómez Amador,
Bray Jesús Martin Agreda Cardenas, Sebastián Calero Anaya, John Hugo Lozano Jauregui,
Alexandr Toribio Hinostroza and Juan José Jiménez de Cisneros y Fonfrı́a
Mathematical Analysis of a Low Cost Mechanical Ventilator Respiratory Dynamics Enhanced
by a Sensor Transducer (ST) Based in Nanostructures of Anodic Aluminium Oxide (AAO)
Reprinted from: Mathematics 2022, 10, 2403, doi:10.3390/math10142403 . . . . . . . . . . . . . . . 213

v
Emilio Lechosa Urquijo, Fernando Blaya Haro, Juan David Cano-Moreno, Roberto D’Amato
and Juan Antonio Juanes Méndez
Mechanical Model and FEM Simulations for Efforts on Biceps and Triceps Muscles under
Vertical Load: Mathematical Formulation of Results
Reprinted from: Mathematics 2022, 10, 2441, doi:10.3390/math10142441 . . . . . . . . . . . . . . . 245

Dylan C. Webb, Elissa Reynolds, Denise M. Halverson and Larry L. Howell

Miura-Ori Inspired Smooth Sheet Attachments for Zipper-Coupled Tubes
Reprinted from: Mathematics 2022, 10, 2643, doi:10.3390/math10152643 . . . . . . . . . . . . . . . 279

Ziyin Cui and Huilin Shang

Jump and Initial-Sensitive Excessive Motion of a Class of Relative Rotation Systems and Their
Control via Delayed Feedback
Reprinted from: Mathematics 2022, 10, 2676, doi:10.3390/math10152676 . . . . . . . . . . . . . . . 295

Leopoldo Prieto, Alejandro Quesada, Ana Marı́a Gómez Amador and Vicente Dı́az
The General Dispersion Relation for the Vibration Modes of Helical Springs
Reprinted from: Mathematics 2022, 10, 2698, doi:10.3390/math10152698 . . . . . . . . . . . . . . . 313

Ruben Medina, Mariela Cerrada, Shuai Yang, Diego Cabrera, Edgar Estupiñan and
René-Vinicio Sánchez
Fault Classification in a Reciprocating Compressor and a Centrifugal Pump Using Non-Linear
Entropy Features
Reprinted from: Mathematics 2022, 10, 3033, doi:10.3390/math10173033 . . . . . . . . . . . . . . . 331

Ester Olmeda, Enrique Roberto Carrillo Li, Jorge Rodrı́guez Hernández and Vicente Dı́az
Lateral Dynamic Simulation of a Bus under Variable Conditions of Camber and Curvature
Radius
Reprinted from: Mathematics 2022, 10, 3081, doi:10.3390/math10173081 . . . . . . . . . . . . . . . 361

vi
Preface to ”Applied Mathematics to Mechanisms and
Machines”
The science of mechanisms and machines has necessitated mathematics from the very moment
of its birth. The much celebrated renaissance figure Leonardo da Vinci already pointed out how
necessary mathematics is for the advancement of science, stating that “No human investigation can
be called real science if it cannot be demonstrated mathematically”. In fact, the first steps of the
development of mechanisms occurred hand in hand with geometry. At first, graphic methods, of
undoubted pedagogical value, provided results for the motion of mechanisms at a given instant,
which involved the arduous task of determining its kinematics and dynamics for the trajectory over
time; conversely, analytical methods were limited to simple case resolution. With the advent and
development of computing, numerical methods have made their way into all areas of knowledge
and technology. In recent decades, these methods have made possible a rapid development of
the analysis, optimization and synthesis of mechanical systems, mechanisms and machines in the
academic, scientific and industrial fields.
The 16 chapters of this book (all of them accepted and published articles in the Special Issue
”Applied Mathematics to Mechanisms and Machines” of the MDPI’s Mathematics journal) cover a
wide variety of topics, including the synthesis, design and optimization of mechanisms, robotics,
automotive, Maintenance 4.0, machine vibrations, control, biomechanics and medical devices, among
others, that combine mechanisms and machine science with mathematics in synergy.
The Guest Editors of the Special Issue are very grateful to the authors of these works for their
high quality contributions, as well as to their reviewers for the sharpness and insightfulness of their
comments. We also want to sincerely thank the administrative staff of MDPI publications for their
invaluable support throughout the process.

Higinio Rubio Alonso, Alejandro Bustos Caballero, Jesus Meneses Alonso, and Enrique
Soriano-Heras
Editors

vii
 mathematics

Article
Hybrid Optimization Based Mathematical Procedure for
Dimensional Synthesis of Slider-Crank Linkage
Alfonso Hernández 1 , Aitor Muñoyerro 2 , Mónica Urízar 1, * and Enrique Amezua 1

1 Faculty of Engineering in Bilbao, University of the Basque Country (UPV/EHU), Plaza Ingeniero Torres
Quevedo, 48013 Bilbao, Spain; a.hernandez@ehu.es (A.H.); enrique.amezua@ehu.es (E.A.)
2 SENER Aeroespacial, Avda. de Zugazarte 56, 48992 Getxo, Spain; aitor.munoyerro@aeroespacial.sener (A.M.)
* Correspondence: monica.urizar@ehu.es (M.U.)

Abstract: In this paper, an optimization procedure for path generation synthesis of the slider-crank
mechanism will be presented. The proposed approach is based on a hybrid strategy, mixing local and
global optimization techniques. Regarding the local optimization scheme, based on the null gradient
condition, a novel methodology to solve the resulting non-linear equations is developed. The solving
procedure consists of decoupling two subsystems of equations which can be solved separately and
following an iterative process. In relation to the global technique, a multi-start method based on a
genetic algorithm is implemented. The fitness function incorporated in the genetic algorithm will take
as arguments the set of dimensional parameters of the slider-crank mechanism. Several illustrative
examples will prove the validity of the proposed optimization methodology, in some cases achieving
an even better result compared to mechanisms with a higher number of dimensional parameters,
such as the four-bar mechanism or the Watt’s mechanism.

Keywords: path generation; dimensional synthesis; hybrid optimization; slider-crank mechanism

Citation: Hernández, A.; Muñoyerro,
A.; Urízar, M.; Amezua, E. Hybrid
Optimization Based Mathematical
Procedure for Dimensional Synthesis 1. Introduction
of Slider-Crank Linkage. Mathematics
Dimensional synthesis consists of finding a geometry that enables a mechanism
2021, 9, 1581. https://doi.org/
to generate certain motion characteristics, such as trajectories or positions of elements.
10.3390/math9131581
It is tough to solve this problem intuitively and often requires the implementation of
Academic Editor: Raimondas Ciegis
specific methods. Depending on the type and amount of prescribed motion characteristics,
it is not always possible to obtain an exact solution to this problem, forcing us to use
Received: 31 May 2021 optimization methods to find an approximation with minimal error. The target most
Accepted: 2 July 2021 commonly addressed in bibliographies is the synthesis type, known as path generation,
Published: 5 July 2021 where a point of a single degree of freedom mechanism is sought to run through a sequence
of prescribed positions. This motion may or may not be synchronized with the location of
Publisher’s Note: MDPI stays neutral the input element, resulting in prescribed or unprescribed timing problems, respectively.
with regard to jurisdictional claims in It should be noted that there also exist two other goals that are frequently studied, these
published maps and institutional affil- being beyond the scope of this paper. These are function generation, where the motion of
iations. two elements of the mechanism is synchronized, and motion generation, where a sequence
of locations for a certain element of the mechanism is prescribed.
Most of the literature on dimensional synthesis focuses on individual cases, the four-
bar hinged mechanism being the most widely studied. In the existing literature, both the
Copyright: © 2021 by the authors. exact synthesis, by means of graphic [1] or analytical methods [2], and the approximate syn-
Licensee MDPI, Basel, Switzerland. thesis, by means of dimensional optimization [3,4], have been studied. On the other hand,
This article is an open access article the slider-crank mechanism has been used to solve function generation problems [5–8], for
distributed under the terms and dynamic synthesis [9,10], and to serve as an adjustable mechanism [11–14]. However, no
conditions of the Creative Commons publications on dimensional optimization for path generation have been found. It should
Attribution (CC BY) license (https:// be noted that, in addition to the usual kinematic objectives, some papers include more
creativecommons.org/licenses/by/ specific characteristics within the error function. For example, in [15] a formulation is
4.0/).

Mathematics 2021, 9, 1581. https://doi.org/10.3390/math9131581 1 https://www.mdpi.com/journal/mathematics

Mathematics 2021, 9, 1581

developed by using exact differentiation that allows for establishing the position of the
instant center of rotation and the centrode.
An often-employed dimensional optimization procedure consists of minimizing the
error function, formulated as the sum of squared differences between the points of the
discretized prescribed path, and those belonging to the real generated path. The min-
imization process can be solved with different methods that can be classified into two
main groups, these being local and global methods, which are mentioned below. On the
other hand, there are also publications focused not so much on studying the mathematical
optimization techniques, but on proposing new ways of describing the output generated
by the mechanism. This may result in a more advantageous definition for the optimization
error function or for the elaboration of atlases and databases. In relation to the path gen-
eration problems discussed in this paper, there exist different approaches to describe the
trajectories, and probably the most typical ones are based on the Fourier series [16] or Haar
Wavelet transform [17]. Similarly, reference [18] describes a unified theory of the harmonic
characteristic parameter method for mechanism synthesis. Apart from the dimensional
synthesis, other publications focus on the phase that precedes it, i.e., structural synthesis,
the first step in the conceptual design of mechanisms. In this sense, there can be found
some proposals of automatic algorithms intended for the structural synthesis of robots and
closed-loop mechanisms [19].
In relation to the mathematical optimization techniques for dimensional synthesis,
the most effective and widely used local methods consist of applying the null gradient
condition, which leads to a non-linear system of equations. This system often includes
some passive variables that cannot be eliminated. To solve it, the function is linearized and
an iterative method, such as Gauss–Newton, is used, starting from an approximate initial
solution provided by the designer. The original reference for this type of method is a paper
published in 1966 by Chi-Yeh [20], which was dedicated to the four-bar linkage. From then
on, several papers related to dimensional synthesis of this mechanism by means of gradient
methods have been published, exploring different ways to improve the effectiveness
of optimization. As part of these alternative approaches, in [21] the authors proposed
modifying the set of variables to be optimized, considering the nodal coordinates instead
of the usual dimensional parameters, thereby allowing the elimination of some constraints
that were present in the original problem. It is also noteworthy that other publications
focus on reformulating the error function, such as [22,23]. The authors of those works
proposed to minimize the strain energy originated when the mechanism is forced to run
exactly through the prescribed trajectory. On the other hand, some authors have chosen to
estimate the error by avoiding its sensitivity to translation and rotation effects, such as [24],
where a system of relative coordinates between precision points is used. Following the
same idea, the authors proposed to perform a prior and independent phase to optimize the
translation, rotation, and scaling parameters [25]. In addition to the different alternatives
to characterize the design parameters of the mechanism and to estimate the resulting error,
a relevant aspect to achieve good performance in gradient methods is to carry out an exact
calculation of the partial derivatives, avoiding the numerical derivation, since it increases
the computational cost and results in a lower efficiency. Given the interest in solving
this problem, reference [26] presents a general method for calculating the exact partial
derivatives from the loop equations previously identified by the designer.
Despite the large number of existing publications devoted to optimal dimensional
synthesis by means of gradient methods, and to the improvement of their performance,
none of them are capable of solving the main limitation they have. Unfortunately, these
methods are highly sensitive to the chosen starting approximation, since they are local
in nature and hence converge to the nearest minimum, which will not necessarily be the
optimal overall solution. To overcome this drawback, global methods make it possible to
explore the entire space where solutions can be found. Metaheuristic methods are the most
common ones, and they have been covered in several references. These include genetic
algorithms [27–29], differential evolution [30–32], ant search [33], krill herd algorithm [34],

2
Mathematics 2021, 9, 1581

imperialist competitive algorithm [35], or neural networks [36]. Nevertheless, the weakness
of heuristic methods in comparison with gradient methods is their higher computational
cost and a lower convergence rate. Furthermore, there is no guarantee that they will
converge to a minimum, neither locally nor globally.
Hybrid optimization algorithms, such as [37], gain greater strength when a global
method generating seeds for starting mechanisms is combined with a good local method.
Normally, hybrid methods start by running a global method to obtain one or several designs
that will later be used as initial approximations in a local method to quickly converge to
the nearest relative minimum. In this paper, as described below, a hybrid optimization
approach is proposed.
In the case of the slider crank mechanism analyzed in this paper, thanks to the
simplicity of its kinematics, it is possible to express the synthesis variables directly as
a function of the dimensional and input parameters, and thus completely eliminating
the passive variables. A novel aspect of the proposed approach is the way in which the
resulting system of equations is solved. Considering the more general case of unprescribed
timing synthesis, the final system of equations associated with the null gradient condition
can be divided into two subsystems with different characteristics. As will be explained,
the procedure described in this article allows each subsystem to be solved separately
within an iterative process that connects them together. This makes it possible, in some
particularly simple cases, such as the two-parameter slider-crank, to solve all the equations
analytically, while in more complex cases, numerical methods must necessarily be adopted.
The optimum solution reached will be a relative minimum of the error function and will be
influenced by the initial approximation used to solve the system of equations numerically.
In this paper, reducing this dependency will be attempted by running the local optimization
algorithm from different starting points previously selected by a genetic algorithm. The
generation of the starting points could also be done by another type of heuristic method,
or through a sweeping process that generates random points within the entire search
space. Even so, the latter technique would not achieve the most promising regions as the
genetic algorithm does, but it would only generate a wide grid of different starting points.
Therefore, the genetic algorithm is the preferred choice.
It is important to highlight that the procedure described in this paper seeks to obtain
reliable solutions, not only from a mathematical vision, but also from a practical point of
view. For this reason, it will be stated how to avoid the circuit defect in the slider-crank
mechanism, relying on the concept of branch index. This concept was introduced in
reference [21] to analyze the kinematics of the four-bar hinged linkage. The incorporation
of design constraints will be addressed by means of penalty functions included in the error
function to be minimized. This is essential to impose the Grashof criterion and thus ensure
that the input element is able to fully rotate (crank input).
Finally, the effectiveness of the proposed methodology will be illustrated through
different examples. The final solutions obtained in this paper are as accurate as the ones
reached in other papers when solving the same problem by using more complex designs,
such as four-bar or Watt linkages.
The main novelties and highlights of this paper are:
• Deduction of the equations required for the optimal dimensional synthesis of the slider-
crank mechanism, which constitutes an alternative to the hinged four-bar linkage
usually used in the literature to solve this type of problem.
• Proposal of an original methodology to solve a non-linear system of equations re-
sulting from the null gradient condition, based on the decoupling of two subsystems
of equations. It facilitates the resolution of the system and, in some cases, allows to
obtain all the solutions in an analytical way.
• Integration of the local optimization methodology within a hybrid optimization
method, which uses a genetic algorithm to search for the best starting approximations.
The fitness function has been adapted to solve not only the prescribed timing problem,
but also unprescribed timing.

3
Mathematics 2021, 9, 1581

• Solving and comparison of examples proposed by other authors in the literature

dealing with the four-bar linkage. Thanks to the effectiveness of the method proposed
in this work, the slider-crank mechanism, though being simpler and more limited, is
able to provide similar performances (or even better in some cases) in path generation
problems.

2. Materials and Methods

In this section, the basis of the optimum synthesis procedure as well as the functioning
of the hybrid optimization strategy is presented.

2.1. Bases of the Optimum Synthesis Procedure

Before tackling the optimization process, the loop-closure equations of the mechanism
and the position equations of the coupler point that traces the trajectory will be obtained.
In addition, the variables that take part in the synthesis process and their functional
dependence law will be described. This enables obtaining the partial derivatives that
intervene in the minimization process.

2.1.1. Synthesis Equations for a General Design

First, the kinematic problem of the slider-crank mechanism with five dimensional
parameters, represented in Figure 1, is obtained.

Figure 1. Slider-crank mechanism with 5 dimensional parameters.

The loop-closure equations are the following:

a1 ·sinϕ = a5 − a2 ·sinθ (1)

s = a1 ·cosϕ + a2 ·cosθ (2)

From Equation (1) yields:

a5 − a1 ·sinϕ
sinθ = (3)
a2

Then, cosθ is given by Equation (4), where K = ±1.


 2
a5 − a1 ·sinϕ
cosθ = K 1− (4)
a2

The synthesis equations are as follows:

 
3π
x  = s − ( a2 − a3 )·cosθ + a4 ·cos θ − (5)
2

y = −( a2 − a3 )·sinθ + a5 + a4 ·cosθ (6)

4
Mathematics 2021, 9, 1581

Using the loop equations to solve the passive variables s (Equation (2)) and θ
(Equations (3) and (4)), and substituting them in Equations (5) and (6), the following
expressions are obtained for the synthesis variables, referring to the local system O’X’Y’:

 2
 a5 − a1 sinϕ a4 ·( a5 − a1 sinϕ)
x = a1 cosϕ + a3 ·K 1− − (7)
a2 a2

 2
 ( a3 − a2 )·( a5 − a1 sinϕ) a5 − a1 sinϕ
y = cosϕ + a5 + a4 ·K 1− (8)
a2 a2

Remark Regarding Branches and Circuits

The sign ± in Equations (7) and (8), which has been substituted for K = ±1 for
simplicity, is related to the two possible positions of the coupler point P for the same
input ϕ. This means that two branches exist, each one associated with the positive or
negative value of K. This circumstance is illustrated in Figure 2, in which the two possible
configurations of the mechanism, for a given value of the input ϕ, are represented.

Figure 2. Two possible configurations for the same input.

The two possible trajectories of point P associated with the different configurations
of the coupler element, commonly known as branches, can be connected or unconnected,
resulting in a unique circuit (a unicursal curve), or two circuits (a bicursal curve). In this
work, designs where the crank input is able to perform a 360◦ full rotation are considered,
meaning that the Grashof criterion must be fulfilled. Therefore, the two possible branches
will be two unconnected circuits. To avoid branch defects, all the selected points must have
the same value of K. This value will be the one corresponding to the branch that yields a
minimum error with respect to the desired path.
In the most general case, represented in Figure 3, the local reference system O’X’Y’ has
a rotation relative to the global system OXY, defined by the parameter a6 , and a translation
in the plane defined by the parameters a7 and a8 .
The equations that express the synthesis variables in the global reference system are
the following:
xi = xi  ·cos( a6 ) − yi  ·sin( a6 ) + a8 (9)
yi = xi  ·sin( a6 ) + yi  ·cos( a6 ) + a7 (10)

5
Mathematics 2021, 9, 1581

Figure 3. Slider-crank mechanism with 8 dimensional parameters.

Now that all the variables involved in the synthesis problem have been defined, the
following classification can be established:
• Dimensional variables: a1 , a2 , . . . , a8 . These are variables that define the lengths of
the bars and the translation or rotation parameters of the studied mechanism.
• Input variable: ϕ. This is an independent variable corresponding to the degree of
freedom of the mechanism under study.
• Passive variables: θ, s. These are not independent variables, but rather depend on the
input and the dimensional parameters.
• Output variables or synthesis variables: x, y. These correspond to the coordinates
of the coupler point P. In the case of path generation synthesis, these are indeed the
synthesis variables.

2.1.2. Optimal Design Based on the Error Function

The error function E, commonly used in synthesis problems, is defined as the sum of
the squared Cartesian distances between the prescribed points and those actually generated:

N  2  2 
E= ∑ xi − xid + yi − yid (11)
i =1

The error between prescribed and generated trajectory must be minimized to obtain
the optimal mechanism. There are two options for carrying out this minimization. In the
modality known as prescribed timing, only the dimensional parameters are optimized,
requiring solving the system shown in Equation (12). In this case, the input parameters
ϕi are not variables to be optimized, but constant values (prescribed values). However,
in an optimization known as unprescribed timing, both the dimensional parameters a j
and the set of input parameters ϕi are optimized, requiring solving the systems given by
Equations (12) and (13). This last option is more complex but its potential to obtain precise
solutions is greater, since the value of the input parameters is not being restricted.

N   ∂x   ∂y
∂E
= 0 → ∑ xi − xid i
+ yi − yid i
=0 ∀ j = 1, 2, . . . , n (12)
∂a j i =1
∂a j ∂a j

N   ∂x   ∂y 
∂E
= 0 → ∑ xi − xid i
+ yi − yid i
=0 ∀i = 1, 2, . . . , N (13)
∂ϕi i =1
∂ϕi ∂ϕi

6
Mathematics 2021, 9, 1581

2.2. Hybrid Optimization Procedure

Thanks to the simplicity of its kinematics, in the slider-crank mechanism it is possible
to eliminate the passive variables. In this way, the synthesis variables ( x, y) are expressed
as an explicit function of parameters a j and ϕ, facilitating the obtaining of the partial
derivatives that appear in Equations (12) and (13). After performing the substitution, it
can be verified that the system of Equations (12) and (13) is non-linear. This system can be
decomposed into two subsystems of equations. On the one hand, Equation (12) is composed
of 8 equations, as many as the dimensional parameters of the mechanism. Each of these
equations has the form g( a1 , a2 . . . a8 , ϕ1 , ϕ2 , . . . ϕ N ) = 0. On the other hand, Equation (13)
is made up of N equations, as many as points of precision has the trajectory. Each of these
equations has the form h( a1 , a2 . . . a8 , ϕi ) = 0. This circumstance is fundamental when
establishing the procedure for solving the system of equations.
In addition, in order to find the optimal initial approximation, the one that will be
the starting point of the local optimization process, and that plays a relevant role in the
process, a multi-start approach will be implemented. In the following sections, the hybrid
procedure proposed by the authors of this work will be explained.

2.2.1. Solving the Equation System for Local Optimization

In this section, the iterative algorithm to solve the non-linear system of equations
resulting from the null gradient condition is addressed. The procedure is based on the
decoupling of the two subsystems (12) and (13), as described below.
• First phase:
This starts from a certain mechanism, coming from the multi-start procedure that
will be explained in Section 3.2. The dimensions (a1 , a2 , . . . , a8 ) of this mechanism are
assumed constant in this phase. The values of these dimensions are substituted in each of
the equations of subsystem (13), resulting in a total of N equations, each with a unique ϕi .
In the simplest case of the slider-crank mechanism, with 2 dimensional parameters
(a1 = a2 = a3 and a5 = a6 = a7 = a8 = 0), and using the transformation of half-angle
ϕ
tangent, ti = tan 2i , Equation (13) becomes a fourth degree polynomial:
     
2a1 a4 + xid a4 t4i + 8a1 2 + xid a1 + 2yid a4 t3i − 12a1 a4 t2i + −8a1 2 + 4xid a1 + 2yid a4 ti + 2a1 a4 − xid a4 = 0 (14)

From the 4 possible values of ti ( ϕi ), the one that gives the minimum error is the
selected one.
For a design with 3 dimensional parameters, Equation (13) becomes a polynomial of
degree 10. Even so, it is easy to obtain the 10 roots and detect the correct one proceeding as
in the previous case. However, with 4 or more dimensional parameters, it is no longer as
easy to determine its corresponding univariate polynomial, nor is it really worth it. It is
more practical to operate as explained next.
The Equation (13) is solved numerically, starting from an initial approximation, and
arriving at a unique solution of the parameter ϕi . To guarantee that the global optimal
value of ϕi is obtained, it will be necessary to start from an initial approximation obtained
as follows: each sum of Equation (11), which represents the error made in each synthesis
position i, is evaluated as a function of the input parameter ϕi along the discretized domain
[0,2π). In this way, a graph similar to the blue curve represented in Figure 4 will be obtained,
where two minima of the error function appear, the one indicated on the right being the
one with the lowest value. The latter value of ϕi is taken as a starting approximation to
solve Equation (13) (in our case, using the MATLAB fsolve command).

7
Mathematics 2021, 9, 1581

Figure 4. Error function for a point i evaluated at [0,2π).

To illustrate this concept, in Figure 5 the prescribed point for a synthesis position i
and the trajectory generated by the mechanism of dimensions a j in the current iteration
are indicated. The effect of evaluating the sum of the error function (11) and choosing
the absolute minimum is equivalent to traversing the generated trajectory and selecting
the point of it (the black point) closest to the prescribed one (red point). In this example,
the black point indicated in Figure 5 corresponds to the absolute minimum, ϕ = 5.88 rad,
which is the one previously indicated in Figure 4.

Figure 5. Generated trajectory and prescribed point (red).

• Second phase:
In this phase, the values ϕi obtained in the previous phase will be assumed as
constants, and the unknowns (a1 , a2 , . . . , a8 ) will be calculated from the subsystem from
Equation (12).
As an example, in the simplified particular case of the slider-crank mechanism with
2 parameters, Equation (12) becomes the following linear system:
⎧ ⎫
⎡ ⎤ ⎪ a1 ⎪ ⎧ ⎫
N N ⎪
⎪ ⎪ ⎪
⎪ N
⎪
⎢ 4 i∑ cos2 ϕi ∑ sin2ϕi ⎥⎪⎨ ⎬ ⎪
⎪ ⎨ 2 ∑ cosϕi · xid ⎪
⎬
⎢ =1 i =1 ⎥ = i = 1
⎣ N ⎦⎪ ⎪ ⎪ N   ⎪
(15)
∑ sin2ϕi N ⎪
⎪ ⎪
⎪ ⎪
⎩ ∑ sinϕi · xid + cosϕi · yid ⎪
⎭
⎪
⎩ ⎪
⎭
i =1 a4 i =1

However, in the design cases with 3 or more dimensional parameters, the subsystem
of equations turns out to be non-linear, making it necessary to apply numerical solving

8
Mathematics 2021, 9, 1581

methods. As a starting approximation, the values (a1 , a2 , . . . , a8 ) that were assumed as

constants in the previous phase will be taken.
• Next steps:
With the new values of a j obtained in the second phase, we go back to the first phase,
and thus continue iteratively until convergence. The chosen stopping criterion consists
of comparing the current values of the optimization variables ( a1 , a2 . . . a8 , ϕ1 , ϕ2 , . . . ϕ N )
with those of the previous iteration, so that when the difference is less than a specified
tolerance, the iterative process will stop.

2.2.2. Implementing a Multi-Start Strategy

The method proposed in Section 3.1 has the disadvantage of being very sensitive to
the starting approximation. To avoid this drawback, a multi-start approach based on a
genetic algorithm will be used to guide the local optimization method towards a search for
the global optimal solution.
The proposed multi-start approach obtains a sufficiently large set of starting approxi-
mations, representative of a global sweep of the design space a j . In this article, a genetic
algorithm is used to locate 100 candidate solutions that serve as initial approximations. It is
decided to use the genetic algorithm incorporated in MATLAB, assigning as arguments of
the fitness function the values for the 8 dimensional parameters. The latter function assigns
the optimal correspondence of input parameters, according to the procedure outlined in
the first phase of Section 3.1, and returns a scalar value that quantifies the individual’s
fitness as the quadratic sum of the distances between generated and prescribed points.
It is important to bear in mind that, considering the way the fitness function is posed,
each starting solution for the unprescribed timing problem will not depend on 8 + N
variables (8 dimensional + N inputs), but only on the 8 dimensional parameters, which
facilitates the exploration of the search space by the genetic algorithm without the need to
restrict the input parameters.
The complete scheme of this hybrid procedure is shown in Figure 6. As can be seen,
it consists of executing a local optimization from the different starting points obtained
in the scan of the space of the dimensional parameters a j . In this way, different local
minimums will be obtained and the best of them will be selected.
To avoid excessive computational cost, the execution time of the genetic algorithm
is limited to an acceptable value (i.e., 1 min). Bear in mind that it will not be necessary to
obtain the optimal solutions, but that it will be enough to be close to them. On the other
hand, the execution time of the local method to be applied later will depend on the number
of iterations performed, but it generally consumes a few seconds for each starting point
used.

2.2.3. Incorporation of Design Constraints

Searching for an optimal design usually implies adapting the generated trajectory to
a prescribed one. In addition, some design requirements must be fulfilled. These design
constraints are related to several aspects, such as the maximum size of the mechanism, the
maximum and minimum lengths of certain bars, the Grashof criterion, the transmission
angle limits, and so on.

9
Mathematics 2021, 9, 1581

Figure 6. Flowchart of the hybrid optimization scheme.

In general, the optimization problem with certain inequality constraints can be formu-
lated as the minimum error subjected to

gk { a1 , . . . , an } ≥ 0 ∀k = 1, 2, . . . , l

A typical constraint equation, gk , is expressed as follows:

m
∑ a j Cj ≥ C0 (16)
j =1

When an inequality constraint, such as Equation (16), is included in the process, the
error function yields
 2
N  2 N  2 m
E= ∑ xi − xid + ∑ yi − yid + λ· ∑ a j Cj − C0 (17)
i =1 i =1 j =1

10
Mathematics 2021, 9, 1581

If, on the one hand, Equation (16) is not fulfilled during the iterations of the optimiza-
tion process, it is necessary to amplify the value of the error to invalidate the corresponding
design a j . This can be done by means of the λ parameter, which acts as a penalty factor,
proceeding as follows:
m
If ∑ a j Cj − C0 < 0 then λ = Prede f ined value
j =1
m
If ∑ a j Cj − C0 ≥ 0 then λ = 0
j =1

If, on the other hand, the condition to be imposed is an equality constraint, then λ
will always be a certain predefined value. In this paper, after reviewing similar works in
the literature and conducting several trials with different values, the value λ = 50 was
established, insofar as it provides good results for all the analyzed design cases.

3. Results
The following examples will demonstrate the validity of the proposed methodology,
showing the results of each design case.

3.1. Demonstrative Example 1

The desired trajectory, shown in Figure 7, is formed by two straight lines and two
circumference arcs tangent to these lines. Considering the simplicity of the slider-crank,
achieving a design that traces this trajectory is challenging. Moreover, this example is inter-
esting because the trajectory combines both straight lines and arcs, which are trajectories
commonly used in  the path planning synthesis of planar mechanisms. The coordinates of
the desired points, xid , yid , are given in Table 1. A total of 16 points are defined.

Figure 7. Example 1. Desired trajectory.

The eight dimensional parameters and the 16 input parameters (unprescribed timing)
will be optimized, that is, the most complete case possible. Following the previously
described technique, this example is solved using the proposed multi-start method, starting
from 100 different starting mechanisms. The conditions of the crank-input (based on
Grashof’s criterion) and a maximum 3:1 ratio between the longest and shortest bar are
imposed by means of penalty functions.

11
Mathematics 2021, 9, 1581

Table 1. Example 1. Initial data.

i xdi ydi
1 3.2500 1.7010
2 3.6294 1.6510
3 3.9330 1.8840
4 4.0995 2.1724
5 4.2665 2.4616
6 4.4330 2.7500
7 4.4829 3.1294
8 4.2500 3.4330
9 4.0000 3.5000
10 3.7143 3.5000
11 3.4286 3.5000
12 3.1429 3.5000
13 2.8571 3.5000
14 2.5714 3.5000
15 2.2857 3.5000
16 2.0000 3.5000

The best solution, shown in Figure 8, leads to an error equal to 0.0073. The correspond-
ing design parameters are included in Table 2. In the Supplementary Materials, a video
showing the motion of this optimum mechanism can be found.

Figure 8. Example 1: best multi-start solution for slider-crank, E = 0.0073.

To illustrate the influence that the starting point has on the final solution obtained,
Figure 9 shows, by way of example, the optimal mechanism obtained from another starting
point, which does not reach a solution as good as the one shown above, so it would be
excluded.

Table 2. Example 1: optimum design parameters and inputs.

Parameters Inputs
a1 1.292 ϕ1 4.877 ϕ9 2.992
a2 3.277 ϕ2 4.568 ϕ10 2.804
a3 1.292 ϕ3 4.409 ϕ11 2.634
a4 3.875 ϕ4 4.267 ϕ12 2.476
a5 1.970 ϕ5 4.104 ϕ13 2.323
a6 3.090 ϕ6 3.905 ϕ14 2.171
a7 1.356 ϕ7 3.579 ϕ15 2.014
a8 −0.583 ϕ8 3.192 ϕ16 1.846

12
Mathematics 2021, 9, 1581

Figure 9. Example 1: comparison to other multi-start solution (multi-start no. 10), E = 0.0227.

This same example has been used by the authors of this work in a recent work, in
particular reference [38], making use of another optimization method and applying it to
the four-bar mechanism. It should be taken into account that the four-bar mechanism has
an additional dimensional parameter with respect to the slider-crank and thus, in principle,
a greater potential to adjust to the prescribed trajectories. Nevertheless, comparing the
results of both achieved optimal mechanisms (the error in [38] with the four-bar was 0.0089),
it can be seen that they are almost of the same value, even having a slightly smaller error in
the current case.

3.2. Demonstrative Example 2

Figure 10a shows the results obtained by other authors [30,34,39,40] for an example of
path generation synthesis of a closed path formed by 25 points marked in red. Each of the
authors uses a different optimization method, but all of them use the four-bar mechanism.
On the other hand, Figure 10b shows the results obtained when solving the same example
with the slider-crank under the methodology proposed in this paper.
The comparison is made with other studies that analyze the four-bar because, as
mentioned in the introduction, the vast majority of path generation synthesis works make
use of the four-bar. Even so, as it has been shown in the previous example and it will
happen in this one, the method that we propose achieves very similar (or even better)
results, despite making use of a more limited mechanism such as the slider-crank.
The optimal solution obtained in this paper with the slider-crank mechanism (Figure 10b)
corresponds to an error of 0.103 and is of similar quality to Bulatović et al. [34]. They ob-
tained an error of 0.0392, which is slightly lower than the one we found, but is of the same
order of magnitude. On the other hand, in [34] they also indicated the error achieved by
Laribi et al. [39] and Smaili and Diab [40], these being 0.9022 and 0.5504, respectively, so
they would be considerably higher. On the other hand, the error obtained by Kafash and
Nahvi [30] is not directly comparable with the rest because they use a different error func-
tion. However, the trajectory generated by the authors in [30] is graphically represented
in Figure 10a. The corresponding optimal design parameters for the more general case of
slider-crank are shown in Table 3. As in the previous example, a video showing the motion
of this optimum mechanism is included as Supplementary Materials.

13
Mathematics 2021, 9, 1581

(a) (b)

Figure 10. Example 2: optimal solution for: (a) four-bar and (b) slider-crank mechanism.

Table 3. Example 2: Optimum design parameters and inputs.

Parameters Inputs
a1 2.309 ϕ1 4.703 ϕ9 −0.237 ϕ17 2.715 ϕ25 4.451
a2 48.819 ϕ2 4.92 ϕ10 0.049 ϕ18 2.933
a3 −3.304 ϕ3 5.041 ϕ11 0.672 ϕ19 3.17
a4 24.498 ϕ4 5.181 ϕ12 1.331 ϕ20 3.483
a5 46.315 ϕ5 5.362 ϕ13 1.698 ϕ21 3.762
a6 4.726 ϕ6 5.528 ϕ14 1.996 ϕ22 3.752
a7 −18.02 ϕ7 5.745 ϕ15 2.281 ϕ23 4.091
a8 15.293 ϕ8 5.875 ϕ16 2.494 ϕ24 4.265

3.3. Demonstrative Example 3

On this occasion, a path generation problem, previously solved in reference [41] for
the Watt mechanism, is presented. Their optimal solution is represented in Figure 11a,
together with the set of prescribed points represented in red. The optimal solution obtained
with the slider-crank mechanism, represented in Figure 11b, has an equal, and, it could
be said, even higher precision, despite being a simpler mechanism that has fewer design
parameters involved.
The optimal design parameters of the slider-crank mechanism are listed in Table 4.
Yet again, a video showing the motion of this optimum mechanism is included as Supple-
mentary Materials.

14
Mathematics 2021, 9, 1581

(a)

(b)

Figure 11. (a) Watt’s mechanism optimal solution from [41]; (b) optimal solution for slider-crank
mechanism.

Table 4. Example 3: Optimum design parameters and inputs.

Parameters Inputs
a1 6.719 ϕ1 5.580 ϕ9 3.502
a2 15.635 ϕ2 5.349 ϕ10 3.101
a3 5.217 ϕ3 5.116 ϕ11 2.708
a4 8.737 ϕ4 4.918 ϕ12 2.549
a5 −6.710 ϕ5 4.703
a6 −2.924 ϕ6 4.508
a7 42.513 ϕ7 4.124
a8 35.584 ϕ8 3.848

4. Discussion
In this paper, a hybrid optimization procedure has been presented to address the
dimensional synthesis of path generation using the slider-crank mechanism. On the one
hand, a novel approach has been proposed to solve the non-linear equations corresponding
to the null gradient condition, which is based on the decoupling of two subsystems of
equations, and which simplifies the resolution procedure without prejudice to the precision
obtained. On the other hand, the incorporation of the multi-start method manages to give
the local optimization method a global character, carrying out a previous sweep of the total
space of the dimensional variables to find the most promising starting approximations.
To carry out this search, a genetic algorithm is used where only the eight dimensional
parameters are involved, thanks to the fact that the fitness function has been programmed
to select the optimal input parameters based on the dimensions.
In addition, it should be noted that this methodology is aimed at achieving designs
that are valid from a practical point of view. Therefore, the non-existence of branch defects
is ensured, and different design restrictions can be incorporated in a simple way through the

15
Mathematics 2021, 9, 1581

use of penalty functions. In this way, it can be ensured that the crank input mechanism does
not exceed a maximum ratio between bar lengths or other additional design requirements
at the discretion of the designer.
Finally, the effectiveness of the method has been proved by verifying, through various
examples, that the slider-crank mechanism allows us to achieve solutions with a precision
comparable to other one degree of freedom mechanisms that have a greater number of
dimensional parameters, such as the four-bar or the Watt’s mechanism, with the additional
advantage of having simpler kinematics.

Supplementary Materials: The following are available online at https://www.mdpi.com/2227-739

0/9/13/1581/s1. The three videos corresponding to the motion of the optimum mechanisms are
available.
Author Contributions: Conceptualization, A.H., A.M., M.U. and E.A.; methodology, A.H., A.M. and
M.U.; software, A.M.; validation, A.H., A.M., M.U. and E.A.; formal analysis, A.H.; investigation,
A.H., A.M., M.U. and E.A.; resources, A.H. and A.M.; data curation, A.M.; writing—original draft
preparation, A.H. and A.M.; writing—review and editing, M.U. and E.A.; visualization, A.M.;
supervision, A.H.; project administration, A.H. and M.U.; funding acquisition, A.H., M.U. and E.A.
All authors have read and agreed to the published version of the manuscript.
Funding: The authors wish to acknowledge financial support received from the Spanish government
through the Ministerio de Economía y Competitividad (Project DPI2015−67626-P (MINECO/FEDER,
UE)), the support for the research group through Project Ref. IT949−16, provided by the Depar-
tamento de Educación, Política Lingüística y Cultura from the regional Basque Government, and
the Program BIKAINTEK 2020 (Ref. 012-B2/2020) provided by the Departamento de Desarrollo
Económico, Sostenibilidad y Medio Ambiente from the regional Basque Government.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Erdman, A.G.; Sandor, G.N. Mechanism Design: Analysis and Synthesis, 4th ed.; Pearson: London, UK, 2001.
2. Wampler, C.; Morgan, A.P.; Sommese, A.J. Complete Solution of the Nine-Point Path Synthesis Problem for Four-Bar Linkages. J.
Mech. Des. 1992, 114, 153–159. [CrossRef]
3. Lee, W.-T.; Russell, K. Developments in quantitative dimensional synthesis (1970-present): Four-bar motion generation. Inverse
Probl. Sci. Eng. 2017, 26, 133–148. [CrossRef]
4. Lee, W.-T.; Russell, K. Developments in quantitative dimensional synthesis (1970–present): Four-bar path and function generation.
Inverse Probl. Sci. Eng. 2018, 26, 1280–1304. [CrossRef]
5. Alizade, R.I.; Mohan Rao, A.V.; Sandor, G.N. Optimum Synthesis of Four-Bar and Offset Slider-Crank Planar and Spatial
Mechanisms Using the Penalty Function Approach with Inequality and Equality Constraints. J. Eng. Ind. 1975, 97, 785–790.
[CrossRef]
6. Rao, A.C. Optimum synthesis of a slider-crank mechanism using geometric programming. Int. J. Numer. Methods Eng. 1980, 15,
1595–1602. [CrossRef]
7. Plecnik, M.M.; McCarthy, J.M. Five position synthesis of a slider-crank function generator. In Proceedings of the ASME
International Conference IDETC/CIE 2011, Washington, DC, USA, 28–31 August 2011; pp. 317–324.
8. Almandeel, A.; Murray, A.P.; Myszka, D.H.; Stumph, H.E. A Function Generation Synthesis Methodology for All Defect-Free
Slider-Crank Solutions for Four Precision Points. J. Mech. Robot. 2015, 7, 031020–031021. [CrossRef]
9. Liniecki, A. Synthesis of a slider-crank mechanism with consideration of dynamic effects. J. Mech. 1970, 5, 337–349. [CrossRef]
10. Davidson, J.K. Analysis and synthesis of a slider-crank mechanism with a flexibly-attached slider. J. Mech. 1970, 5, 239–247.
[CrossRef]
11. Zhou, H.; Ting, K.-L. Adjustable slider–crank linkages for multiple path generation. Mech. Mach. Theory 2002, 37, 499–509.
[CrossRef]
12. Russell, K.; Sodhi, R. On the Design of Slider-Crank Mechanisms. Part I: Multi-Phase Motion Generation. Mech. Mach. Theory
2005, 40, 285–299. [CrossRef]
13. Russell, K.; Sodhi, R.S. On the design of slider-crank mechanisms. Part II: Multi-phase path and function generation. Mech. Mach.
Theory 2005, 40, 301–317. [CrossRef]

16
Mathematics 2021, 9, 1581

14. Zhou, H. Dimensional synthesis of adjustable path generation linkages using the optimal slider adjustment. Mech. Mach. Theory
2009, 44, 1866–1876. [CrossRef]
15. Sancibrian, R.; Sarabia, E.G.; Sedano, A.; Blanco, J.M. A general method for the optimal synthesis of mechanisms using prescribed
instant center positions. Appl. Math. Model. 2016, 40, 2206–2222. [CrossRef]
16. Sun, J.; Chu, J. Fourier series representation of the coupler curves of spatial linkages. Appl. Math. Model. 2010, 34, 1396–1403.
[CrossRef]
17. Liu, W.; Sun, J.; Zhang, B.; Chu, J. Wavelet feature parameters representations of open planar curves. Appl. Math. Model. 2018, 57,
614–624. [CrossRef]
18. Jianwei, S.; Jinkui, C.; Baoyu, S. A unified model of harmonic characteristic parameter method for dimensional synthesis of
linkage mechanism. Appl. Math. Model. 2012, 36, 6001–6010. [CrossRef]
19. Ding, H.; Huang, P.; Zi, B.; Kecskeméthy, A. Automatic synthesis of kinematic structures of mechanisms and robots especially for
those with complex structures. Appl. Math. Model. 2012, 36, 6122–6131. [CrossRef]
20. Chi-Yeh, H. A general method for the optimum design of mechanisms. J. Mech. 1967, 1, 301–313. [CrossRef]
21. Angeles, J.; Alivizatos, A.; Akhras, A. An unconstrained nonlinear least-square method of optimization of RRRR planar path
generators. Mech. Mach. Theory 1988, 23, 343–353. [CrossRef]
22. Avilés, R.; Navalpotro, S.; Amezua, E.; Hernández, A. An Energy-Based General Method for the Optimum Synthesis of Mecha-
nisms. J. Mech. Des. 1994, 116, 127–136. [CrossRef]
23. Vallejo, J.; Avilés, R.; Hernández, A.; Amezua, E. Nonlinear optimization of planar linkages for kinematic syntheses. Mech. Mach.
Theory 1995, 30, 501–518. [CrossRef]
24. Sancibrian, R.; Viadero, F.; García, P.; Fernández, A. Gradient-based optimization of path synthesis problems in planar mechanisms.
Mech. Mach. Theory 2004, 39, 839–856. [CrossRef]
25. Sancibrian, R.; De Juan, A.; Sedano, A.; Iglesias, M.; García, P.; Viadero, F.; Fernandez, A. Optimal Dimensional Synthesis of
Linkages Using Exact Jacobian Determination in the SQP Algorithm. Mech. Based Des. Struct. Mach. 2012, 40, 469–486. [CrossRef]
26. Mariappan, J.; Krishnamurty, S. A generalized exact gradient method for mechanism synthesis. Mech. Mach. Theory 1996, 31,
413–421. [CrossRef]
27. Cabrera, J.; Simon, A.; Prado, M. Optimal synthesis of mechanisms with genetic algorithms. Mech. Mach. Theory 2002, 37,
1165–1177. [CrossRef]
28. Acharyya, S.; Mandal, M. Performance of EAs for four-bar linkage synthesis. Mech. Mach. Theory 2009, 44, 1784–1794. [CrossRef]
29. Buśkiewicz, J.; Starosta, R.; Walczak, T. On the application of the curve curvature in path synthesis. Mech. Mach. Theory 2009, 44,
1223–1239. [CrossRef]
30. Kafash, S.H.; Nahvi, A. Optimal synthesis of four-bar path generator linkages using Circular Proximity Function. Mech. Mach.
Theory 2017, 115, 18–34. [CrossRef]
31. Gogate, G.R.; Matekar, S.B. Optimum synthesis of motion generating four-bar mechanisms using alternate error functions. Mech.
Mach. Theory 2012, 54, 41–61. [CrossRef]
32. Bulatović, R.R.; Ðord̄ević, S.R. Control of the optimum synthesis process of a four-bar linkage whose point on the working
member generates the given path. Appl. Math. Comput. 2011, 217, 9765–9778. [CrossRef]
33. Xiao, R.; Tao, Z. A Swarm Intelligence Approach to Path Synthesis of Mechanism. In Proceedings of the Ninth International
Conference on Computer Aided Design and Computer Graphics (CAD-CG’05), Hong Kong, China, 7–10 December 2005; Institute
of Electrical and Electronics Engineers (IEEE): Piscataway, NJ, USA, 2005; pp. 451–456.
34. Bulatovic, R.R.; Miodragovic, G.; Boskovic, M.S. Modified Krill Herd (MKH) algorithm and its application in dimensional
syn-thesis of a four-bar linkage. Mech. Mach. Theory 2016, 95, 1–21. [CrossRef]
35. Ebrahimi, S.; Payvandy, P. Efficient constrained synthesis of path generating four-bar mechanisms based on the heuristic
opti-mization algorithms. Mech. Mach. Theory 2015, 85, 189–204. [CrossRef]
36. Vasiliu, A.; Yannou, B. Dimensional synthesis of planar mechanisms using neural networks: Application to path generator
linkages. Mech. Mach. Theory 2001, 36, 299–310. [CrossRef]
37. Sedano, A.; Sancibrian, R.; De-Juan, A.; Viadero, F.; Egaña, F. Hybrid Optimization Approach for the Design of Mechanisms
Using a New Error Estimator. Math. Probl. Eng. 2012, 2012, 1–20. [CrossRef]
38. Hernández, A.; Muñoyerro, A.; Urízar, M.; Amezua, E. Comprehensive approach for the dimensional synthesis of a four-bar
linkage based on path assessment and reformulating the error function. Mech. Mach. Theory 2021, 156, 104126. [CrossRef]
39. Laribi, M.A.; Mlika, A.; Romdhane, L.; Zeghloul, S. A combined genetic algorithm-fuzzy logic method (GA-FL) in mechanism
synthesis. Mech. Mach. Theory 2004, 39, 717–735. [CrossRef]
40. Smaili, A.; Diab, N. Optimum synthesis of hybrid-task mechanisms using ant-gradient search method. Mech. Mach. Theory 2007,
42, 115–130. [CrossRef]
41. Fernández-Bustos, I.; Aguirrebeitia, J.; Avilés, R.; Angulo, C. Kinematical synthesis of 1-dof mechanisms using finite elements
and genetic algorithms. Finite Elements Anal. Des. 2005, 41, 1441–1463. [CrossRef]

17
 mathematics

Article
Path Analysis for Hybrid Rigid–Flexible Mechanisms
Oscar Altuzarra, David Manuel Solanillas *, Enrique Amezua and Victor Petuya

Department of Mechanical Engineering, University of the Basque Country UPV/EHU, 48013 Bilbao, Spain;
oscar.altuzarra@ehu.eus (O.A.); enrique.amezua@ehu.eus (E.A.); victor.petuya@ehu.eus (V.P.)
* Correspondence: david.m.solanillas.frances@gmail.com or davidmanuel.solanillas@ehu.eus

Abstract: Hybrid rigid–flexible mechanisms are a type of compliant mechanism that combines rigid
and flexible elements, being that their mobility is due to rigid-body joints and the relative flexibility
of bendable rods. Two of the modeling methods of flexible rods are the Cosserat rod model and its
simplification, the Kirchhoff rod model. Both of them present a system of differential equations that
must be solved in conjunction with the boundary constraints of the rod, leading to a boundary value
problem (BVP). In this work, two methods to solve this BVP are applied to analyze the influence of
external loads in the movement of hybrid compliant mechanisms. First, a shooting method (SM)
is used to integrate directly the shape of the flexible rod and the forces that appear in it. Then, an
integration with elliptic integrals (EI) is carried out to solve the workspace of the compliant element,
considering its buckling mode. Applying both methods, an algorithm that obtains the locus of all
possible trajectories of the mechanism’s coupler point, and detects the buckling mode change, is
developed. This algorithm also allows calculating all possible circuits of the mechanism. Thus, the
performance of this method within the path analysis of mechanisms is demonstrated.

Citation: Altuzarra, O.; Solanillas, Keywords: hybrid compliant mechanisms; path analysis; numerical methods; elliptic integrals;
D.M.; Amezua, E.; Petuya, V. Path kinematics
Analysis for Hybrid Rigid-Flexible
Mechanisms. Mathematics 2021, 9,
1869. https://doi.org/10.3390/
math9161869 1. Introduction
A compliant body is one whose motion depends on its geometry, its material properties
Academic Editors: Higinio Rubio
and the location and magnitude of the applied forces. If a body of this kind belongs to a
Alonso, Alejandro Bustos Caballero,
mechanism, it is known as a compliant element. When a mechanism is fully composed of
Jesus Meneses Alonso and Enrique
compliant elements, e.g., slender rods, it is named a compliant mechanism, and it gains
Soriano-Heras
all or part of its mobility, thanks to the relative flexibility of those compliant elements. On
another note, if the mechanism combines rigid and compliant elements, it takes the name
Received: 8 July 2021
Accepted: 2 August 2021
of a hybrid compliant or hybrid rigid–flexible mechanism [1].
Published: 6 August 2021
Regarding applications, such compliant—and, especially, hybrid-compliant—mechanisms
are an alternative to robotic systems of rigid bodies [2], or even cable robots [3] used in tasks
Publisher’s Note: MDPI stays neutral
where there is a human–machine interface. Additionally, such hybrid-flexible mechanisms
with regard to jurisdictional claims in
are a plausible alternative to sub-systems subjected to impacts, such as four-link legs in
published maps and institutional affil- bio-inspired mobile multipod robots [4]. Finally, its elasticity, if appropriately tuned, can
iations. play a beneficial role in the dynamic characteristics of systems used in high-speed link
motions, reducing the need for balancing [5].
For the complete kinematic characterization of a linkage with rigid bars, all the circuits
of the mechanism should be obtained [6]. Each circuit is the set of all possible orientations of
Copyright: © 2021 by the authors.
links that can be calculated without disconnecting any of the joints [7]. If the linkage needs
Licensee MDPI, Basel, Switzerland.
to be disassembled to move from one position to another, these positions lie on different
This article is an open access article
circuits. These mechanism’s circuits do not depend on the input link chosen. In [8,9] is
distributed under the terms and presented the circuit analysis for the Watt and Stephenson –like six-bar mechanisms. It
conditions of the Creative Commons should be noted that, in the case of compliant mechanisms, it is not needed to disassemble
Attribution (CC BY) license (https:// any link to obtain the different circuits but to deform a flexible rod through an external load.
creativecommons.org/licenses/by/ To solve the movement of these hybrid rigid–flexible mechanisms, it is needed to
4.0/). obtain the change of shape of compliant elements and the loads that these bodies withstand.

Mathematics 2021, 9, 1869. https://doi.org/10.3390/math9161869 19 https://www.mdpi.com/journal/mathematics

Mathematics 2021, 9, 1869

There are a few modeling methods to analyze flexible rods, such as the Chained Beam-
Constraint Model (CBCM) [10] or Pseudo-Rigid-Body Model (PRBM) [11,12]. Another model
is the Cosserat rod model, which produces the equations that describe the deformed shape
of a slender flexible element and the loads that appear in it [13,14]. This modeling method
leads to a set of differential equations that, in conjunction with the boundary constraints of
the compliant element, takes the form of a boundary value problem (BVP). This BVP can
be solved with the aid of a shooting method (SM) [15], which combines a direct integration
method, such as 4th order Runge–Kutta (R-K), and a numerical minimization method, e.g.,
Newton–Raphson (N–R), which refines the solutions given by the direct integration.
If the extension of the neutral line and the transverse shear strain are neglected in the
Cosserat rod model, the Kirchhoff model is obtained. The equations of this new model
can also be solved by applying elliptic integrals (EI) that make use of the two parameters
known as the modulus of the function (k) and the amplitude of the elliptic integral (φ) [16].
When a compliant element deforms, points of null moment can appear along its length.
These points of null moment are those where a change of curvature is produced, and they
are also known as inflection points [17,18]. Hence, the slender rod deforms changing
its curvature in each of these points, leading to what is known as buckling modes. These
buckling modes are ordered according to the number of inflection points. For instance, if a
rod has only one of these points, it is in its first buckling mode. If the same rod has two
points, it is in its second buckling mode, and so on. It is worth noting that the higher the
buckling mode, the more deformation energy is needed for it to appear. For this reason, the
compliant elements tend to deform in lower buckling modes, as they require less energy.
These buckling modes can be determined using EI, which can also be used to obtain the
rod workspace in each buckling mode [19].
As mentioned above, the motion of the compliant bodies, and hence that of the
compliant mechanisms, depends on the applied forces. Therefore, this paper makes use of
the SM and EI to check the loads’ influence in the movement of a compliant mechanism. In
addition, the work here developed shows how these methods allow obtaining the locus of
all possible paths of the mechanism’s coupler under any applied load.
A mechanism similar to that found in [20] was used to carry out the work in this paper,
and it is based on a mechanism of [21] (ch. 12). The mechanisms differ in the boundary
constrains of the compliant element. In the case of [20], a clamped-clamped rod is used,
but in this paper, a clamped-hinged rod is analyzed.
Thus, in the following section, an introduction to the general equations of flexible rods
and the different integration methods is given. Moreover, a way to obtain the workspace
of a slender rod using EI is also depicted here. In the third section, a closed-loop hybrid
mechanism is proposed, and it is used to analyze the load-dependent paths after explaining
its general equations. Finally, the potential of the developed algorithm to detect the
buckling mode changes of flexible rods and the possible circuits of the mechanism is
also shown.

2. Elastic Rod Fundamentals

2.1. General Equations
Among all of the modeling methods of flexible rods, because of its easy implemen-
tation within the control of a continuum robot, a simplification of the Cosserat model,
i.e., the Kirchhoff model, was used to develop the algorithms of this work. Recalling
the assumptions made in the Kirchhoff model, they are that the extension of the neutral
line and the transverse shear strain are neglected. It means that the deformation of the
rod is mainly due to bending and torsion. This is more likely to happen in rods such as
the one used in the studied mechanism because of its slenderness. All this reduces the
complexity of the equations, making more efficient algorithms without greatly increasing
calculation errors.

20
Mathematics 2021, 9, 1869

The derivation of the mathematics structure of this section draw on Antman’s work [13]
(chs. 4 and 8) on nonlinear elasticity. Nevertheless, the nomenclature used here are those
used in the robotics community, in the line of the work of Caleb [22].
The framework needed to describe the shape of a slender rod in space includes three
main aspects that, coupled, generate the nonlinear system of differential equations that has
to be solved in order to obtain the relationship between force and deformation. These are
as follows: a kinematic definition of the deformed rod, material constitutive laws and static
equilibrium equations.

2.1.1. Kinematics
A three-dimensional parametric Cartesian curve p(s) ∈ R3 that links the centroids
of each transverse section and the orthonormal rotation matrix R(s) ∈ SO(3) (SO(3) is
the special orthogonal subgroup in three dimensions, SO(3) = {R ∈ R3x3 | R T R =
I, and det(R) = 1}) that orients a local frame attached to the section are used to define
the deformed shape of a rod. The principal axes of each section are usually named the x-
and y-axes. The z-axis is perpendicular to the cross-section and tangent to the deformed
shape; see Figure 1. A scalar reference arc-length parameter s is used to position and orient
each cross-section. This parameter is within the finite interval s ∈ [0, L], being that L is
the length of the rod in the initial state. Due to the assumption that the sections are not
distorted, the whole deformation of the flexible bar can be described by mapping from s to
a homogeneous rigid-body transformation, T(s) ∈ SE(3) (SE(3) is the special Euclidean
subgroup in three dimensions, with p(s) ∈ R3 and R(s) ∈ SO(3)).
 
R(s) p(s)
T(s) = (1)
0T 1

where we will often refer to T(s) as a “frame” hereafter.

=


\

]>P@

]
 ;
[ <

  
 
  
\>P@  
 [>P@
 
Figure 1. A rod deformed in space, along with frames used.

For the unloaded pre-curved state of the rod, the magnitudes described above are
designated with the subscript ◦ .
A priori, R◦ (s) could be assigned arbitrarily. However, one can establish conventions
governing the assignment of reference orientations such that the mapping form R◦ (s)
to R(s) has an easily interpretable meaning in terms of material strains. As mentioned
above, we have chosen to assign reference orientation such that the z axis of the local

21
Mathematics 2021, 9, 1869

frame is parallel to the tangent curve to the reference curve (see Figure 1). Then, the
following applies:

dp◦ 
p
R◦ (s)e3 =  ds  =  ◦ 
  (2)
 dp◦  p◦
 ds 

where e1 , e2 and e3 are used for the standard basis vectors of the local frame: [1 0 0] T ,
[0 1 0] T and [0 0 1] T , respectively. The superscript  is used to indicate differentiation with
respect to the arc length s.
Now, let us consider the derivatives of p(s) and R(s) with respect to s. The position
and orientation evolve along the arc length according to the rates of change, linear v(s) ∈ R3
and angular u(s) ∈ R3 . These rates of change are defined in the local frame, and obtained
from the above mentioned derivatives, defined in the global frame, with the use of R(s) T
as follows:
dp(s) 
v(s) = R(s)T = R(s)T p (s)
ds (3)

U(s) = R(s)T R (s)

where U(s) is a skew-symmetric matrix of the following form:

⎡ ⎤
0 −uz (s) uy (s)
U(s) = ⎣ uz (s) 0 −u x (s)⎦ (4)
−uy (s) u x (s) 0

Because U(s) is defined through three independent values, u x (s), uy (s) and uz (s), it
 T
can be expressed with a vector u(s) = u x (s) uy (s) uz (s) such that u = U∨ and U = u ,
where the ∨ operator denotes conversion of an element of so(3) (so(3) is the Lie algebra of
SO(3)) to its corresponding element in R3 . The inverse operation, denoted by , maps R3 to
so(3), so that u ∨ = u. In the following, u
 (s) is used whenever a vector form is needed to
express U.
When v(s) and u(s) are those of the reference, i.e., the unloaded shape of the rod, they
are designated as v◦ and u◦ . Because of the convention of the rotation matrix stated in (2),
an initially straight rod has v◦ = [0 0 1] T and u◦ = 0.
The material strain can be defined by the variation of v(s), Δv(s) = v(s) − v◦ (s), and
u(s), Δu(s) = u(s) − u◦ (s). The local elongation is represented by Δvz ; values above 1
correspond to extension, values below 1 to compression and values equal to 1 mean that
the length of the curve has not changed. The shear deformation along the local axes is
depicted by Δv x and Δvy . The x and y components of Δu(s), Δu x and Δuy , measure the
bending along the local x- and y-axes. Concerning Δuz , it collects the value of the torsion.

2.1.2. Material Constitutive Relationships

The way the material behaves (it can be elastic or plastic, for instance) affects the
constitutive relationships between strain and stress. However, the Cosserat rod model only
is valid if the cross-sections do not deform.
Δv(s) and Δu(s) can be used to define the stress field since it is obtained from the
strain field. Additionally, Δv(s) and Δu(s) can also be used to express internal force n(s)
and moment m(s), due to the fact that they are calculated from the stress field.
Hereafter, the linear deformation v is related to internal forces by a linear constitu-
tive law, which is also used to relate the angular deformation u to the internal moment.
Diagonal stiffness matrices, i.e., KSE (s) and K BT (s), allow expressing the aforementioned

22
Mathematics 2021, 9, 1869

magnitudes in the local frame. It is worth noting that the subscript SE indicates shear and
extension, and BT, bending and torsion.

nloc (s) = KSE (s)Δv(s) (5)

mloc (s) = K BT (s)Δu(s) (6)

with
⎡ ⎤
GA(s) 0 0
KSE =⎣ 0 GA(s) 0 ⎦ (7)
0 0 EA(s)
⎡ ⎤
EIx (s) 0 0
K BT = ⎣ 0 EIy (s) 0 ⎦ (8)
0 0 GJz (s)

where A is the area of cross section, E is Young’s modulus, G is the shear modulus, Ix and
Iy are the second moment of area of the rod cross section about the principal axes x and y,
and Jz is the polar moment about the local axis z. The linear constitutive law applied here
is commonly used in mechanics for the small strain range. This happens for most metals
and for polymers in the range of small strain.
The rotation matrix is used to change n(s) and m(s) from the local axes to the fixed
global axes.

n(s) = R(s)KSE (s)Δv (9)

m(s) = R(s)K BT (s)Δu (10)

2.1.3. Static Equilibrium Equations

The forces and moments on an infinitesimal piece of the rod, [s, s + ds], are considered
to obtain the static equilibrium equations. The internal force and moment exerted on the
piece by the material of [0, s) are named −n(s) and −m(s). On the other side of the portion,
the material of (s + ds, L] exerts a force n(s + ds) and a moment m(s + ds). In addition,
distributed forces f(s) and moments l(s) can also be applied on the interval ds.
On the one hand, the equilibrium of forces yields the following:

−n(s) + n(s + ds) + f(s)ds = 0 (11)

Expanding the magnitudes in s + ds with the Taylor series until the element of first
order, simplifying terms, and dividing by ds gives the following:

dn(s)
+ f(s) = 0 (12)
ds
On the other hand, a balance of moments, expanding variables in s + ds with the
Taylor series until the element of first order, simplifying terms, and taking into account (12),
yields the following:

dm(s) dp(s)
+ × n + l(s) = 0 (13)
ds ds
So, the nonlinear ordinary differential equations for the equilibrium of a Cosserat rod
describing the evolution of the internal force n(s) and moment m(s) along arc length are
as follows:

n (s) + f(s) = 0 (14)
 
m (s) + p (s) × n(s) + l(s) = 0 (15)

23
Mathematics 2021, 9, 1869

2.1.4. Static Model

If the kinematic, constitutive, and equilibrium laws are combined, a static model of
the flexible rod can be achieved.
Thus, the Cosserat static model for a flexible rod can be obtained using m and n as
state variables, and combining (3), (14) and (15) to obtain a set of differential equations. If it
is considered that the rod is deformed only by bending and torsion, the Cosserat rod model
is simplified, leading to the Kirchhoff model. In the reference rod’s configuration, shear
and extension deformations do not usually appear, i.e., v = v◦ = e3 , where e3 is a unit
vector tangent to the deformed shape. In this way, a new system of differential equations
can be stated:

p = Re3

R = R
u
 (16)
n = −f
 
m = −p × n − l

where u = K−
BT R m + u◦ .
1 T

2.2. Kirchhoff Static Model Equations for the Planar Case

The three-dimensional deformation of a rod in space is defined by (16). However, if
the rod only deforms within a plane, the model can be further simplified and a shorter
formulation is obtained. This has an impact on the computational cost when the resolution
method is implemented within an algorithm.
The rod that appears in Figure 2 deforms within the plane OXY. The local frame
attached to each cross-section is oriented accordingly, so note that expressions in the
previous chapter vary slightly now.


\ [


V



V
\>P@


<
 
 5

 ;
2



        
[>P@
Figure 2. Deformed shape of a clamped-hinged planar rod under load R at extreme.

To orientate each cross-section the rotation matrix is used and it depends on the angle
θ = θ ( s ):
⎡ ⎤
cos θ − sin θ 0
R(s) = ⎣ sin θ cos θ 0⎦ (17)
0 0 1

24
Mathematics 2021, 9, 1869

The derivative of R with respect to arc length s is as follows:

⎡ ⎤
− sin θ − cos θ 0
 
R (s) = ⎣ cos θ − sin θ 0⎦ θ (18)
0 0 0

and allows finding vector u in the following:

⎡  ⎤ ⎡ ⎤
0 −θ 0 0 −uz 0

U = R R = ⎣θ 0⎦ = ⎣ u z 0⎦
T 
0 0 (19)
0 0 0 0 0 0

Thus, only the component uz of vector u is not null:


uz = θ (20)

This makes sense, as there is no bending moment with respect to the local y-axis and
no torsion (expressed through u x ).
Taking (20) into (16) considering this time v = v◦ = e1 , and taking n and m as
state variables, the expression of the system of nonlinear ordinary differential equations
in (16) yields the following:
⎧ ⎫
⎪ dx ⎪ ⎧ ⎫
⎪ ds ⎪
⎪ dy ⎪ ⎪ ⎪ cos θ ⎪
⎪
⎪ ⎪
⎪ ⎪ ⎪
⎪
⎪
⎪ ds ⎪
⎪ ⎪ ⎪ sin θ ⎪
⎪
⎪ ⎪
⎨ dθ ⎬ ⎨ ⎪ ⎪
⎬
mz
+ u◦,z
ds = EI (21)
⎪ dn x
⎪ ds ⎪ ⎪
⎪ ⎪ ⎪
⎪ − f x ⎪
⎪
⎪
⎪ ⎪ ⎪ ⎪
⎪
⎪
⎪
dn ⎪ ⎪ ⎪ − fy ⎪
⎪ ds ⎪ ⎪
y
⎪ ⎪ ⎩ ⎭
⎩ dmz ⎪⎭ n x sin θ − ny cos θ − lz
ds

If no distributed force and moment (f and l) are applied along the rod, we obtain
the following:
⎧ ⎫
⎪ dx ⎪ ⎧ ⎫
⎪
⎪ ds ⎪ cos θ
⎪ dy ⎪
⎪ ⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ ds ⎪
⎪ ⎪ ⎪ sin θ ⎪
⎪
⎪ ⎪
⎨ dθ ⎬ ⎨ ⎪ mz
⎪
⎬
ds
dn x = EI (22)
⎪
⎪ ds ⎪⎪ ⎪
⎪ 0 ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪
dn y
⎪ ⎪
⎪ ⎪ 0 ⎪
⎪
⎪ ds ⎪
⎪ dm ⎪ ⎩n sin θ − n cos θ ⎭
⎩ z⎭ x y
ds

Hence, in this particular case, the internal force n is constant. Such a rod with no
distributed loads is quite common in the applications.
The system of differential equations in (22) can be solved by using a direct integration
method, such as the 4th order R-K, or a method based on elliptic integrals.

2.3. Direct Integration Using Fourth-Order Runge–Kutta Method

Typical problems in robotics are the forward kinematic problem (FKP) and the inverse
kinematic problem (IKP). In the first one, the mechanism inputs are known, and the position
and orientation of the end-effector are unknown. In contrast, in the IKP, the end-effector is
wanted to describe a given trajectory, so its position is known, but the needed inputs to
accomplish the task must be calculated.
When a flexible rod is part of a mechanism, usually, it is under external loads coming
from that mechanism. Here, the aim should be to calculate how these loads deform the rod.
So, it could be considered as a kind of forward kinematic (FK) position problem because
the position and orientation of the extreme in which force and moment are applied are
not known.

25
Mathematics 2021, 9, 1869

So focusing on the FKP of a rod that is clamped at the proximal end s = 0 and
subjected to a known force next and moment mext at the distal end s = L (see Figure 2),
the rod undergoes a deformation that depends on these boundary conditions and the
geometric and mechanical characterization of the rod.
If the effect of gravity is not taken into account, there is not distributed force along the
length. Therefore, the shape and the load status of the rod are defined by (22).
This takes the form of a BVP, where the independent variable is s in [0, L], and the
vector of dependent variables, y, and the non-linear function, f(s, y), are the following:
⎧ ⎫ ⎧ ⎫
⎪
⎪ x⎪ ⎪ ⎪
⎪ cos θ ⎪
⎪
⎪
⎪ y⎪ ⎪ ⎪
⎪ sin θ ⎪
⎪
⎪
⎪ ⎪ ⎪ ⎪
⎨ ⎪ ⎬ ⎪
⎨ mz
⎪
⎬
θ
y= , f(s, y) = EI (23)
⎪
⎪ n x⎪
⎪ ⎪
⎪ 0 ⎪
⎪
⎪
⎪ ⎪ ⎪ ⎪
⎪ ny ⎪
⎪ ⎪ ⎪
⎪ 0 ⎪
⎪
⎩ ⎪ ⎭ ⎪
⎩ ⎪
⎭
mz n x sin θ − ny cos θ

The boundary conditions are at s = 0 of the kinematic kind, and at s = L of the load
kind. Regarding the kinematic variables, the position and orientation at s = 0 are data,
and unknown at s = L. In relation to internal forces, and because no distributed force is
considered, its value does not change along the length of the rod. At s = 0, and because of
the static equilibrium, the internal force takes the same value that of the external load, next ,
at s = L, but in opposite direction (although the internal forces could be taken out of the
integration, in this work, they are included because they are needed to solve the FKP of the
whole mechanism). With respect to the internal moment, the value is a datum at s = L (in
fact, m( L) = mext ), and unknown at s = 0.
Then, y at both ends are the following:
⎧ ⎫ ⎧ ⎫ ⎧ ⎫ ⎧ ⎫
⎪
⎪ x (0) ⎪ ⎪ ⎪
⎪ x0 ⎪
⎪ ⎪
⎪ x ( L) ⎪ ⎪ ⎪
⎪ ? ⎪ ⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪ ⎪ ⎪ ⎪ ⎪
⎪
⎪ y ( 0 ) ⎪
⎪ ⎪ y 0⎪
⎪ ⎪
⎪ y( L) ⎪ ⎪ ⎪
⎪ ? ⎪ ⎪
⎨ ⎬ ⎪ ⎨ ⎪ ⎬ ⎪
⎨ ⎪
⎬ ⎪ ⎨ ⎪
⎬
θ (0) θ0 θ ( L) ?
y (0) = = , y( L) = = (24)
⎪ n x (0) ⎪
⎪
⎪ ⎪ ⎪
⎪ ⎪
⎪
?⎪ ⎪
⎪ ⎪ n x ( L) ⎪
⎪
⎪ ⎪ ⎪
⎪ ⎪
⎪
next,x ⎪
⎪
⎪
⎪
⎪ n y (0) ⎪⎪ ⎪
⎪ ?⎪ ⎪ ⎪
⎪ ny ( L) ⎪⎪ ⎪
⎪ n ⎪
⎪
⎪
⎩ ⎪
⎭ ⎩ ⎪
⎪ ⎭ ⎪
⎩ ⎪ ⎪
⎭ ⎩ ext,y ⎪ ⎭
m z (0) ? mz ( L) mext

being Dirichlet conditions.

Since there are three unknowns at the initial of the range of integration s = 0, three
!T
y
guess values (one for each variable) are needed: guess = n xguess n guess mzguess . Once
these values are chosen, the initial value problem is solved using the 4th R-K method.

dy
= f(s, y) , y ( s0 ) = y0 (25)
ds
where y is a vector that contains the unknown variables at each s, function f is given, s is
the independent variable, and subscript 0 indicates values at initio.

2.4. Integration Using Elliptic Integrals

The system of differential equations in (22) can be further developed looking for an
analytical integration. If only end-point load is applied (expressed in terms of magnitude
R and orientation ψ as in Figure 2), we can express the internal force as the following:

n x = R cos ψ (26)
ny = R sin ψ (27)

26
Mathematics 2021, 9, 1869

For constant E and I, with a stress-free reference straight rod, considering the well-
known Bernoulli–Euler law, and upon substitution of (26) and (27) into (22), we obtain
the following:

d2 θ 1 R
= ( R cos ψ sin θ − R sin ψ cos θ ) = sin (θ − ψ) (28)
d2 s EI EI
Therefore, the system of differential Equation (22) can be simplified to the following:
⎧ ⎫ ⎧ ⎫
dx
⎨ ds ⎪
⎪ ⎬ ⎨ cos θ ⎬
dy
= sin θ (29)
⎪
⎩ d2 θ ⎪
ds
⎭ ⎩ R sin (θ − ψ)⎭
2d s EI

Elliptic integrals is the classical mathematical tool to solve (29) because of its rapid
computation. The following equations are used in this work to solve for the coordinates of
an arbitrary point along the beam, and they are found in [21] (ch. 4).
"
EI
x (φi ) = − cos ψ[2E(k, φi ) − 2E(k, φ1 ) − F (k, φi ) + F (k, φ1 )]
R
"
EI
− 2k sin ψ[cos φi − cos φ1 ] (30)
R

"
EI
y(φi ) = − sin ψ[2E(k, φi ) − 2E(k, φ1 ) − F (k, φi ) + F (k, φ1 )]
R
"
EI
+ 2k cos ψ[cos φi − cos φ1 ] (31)
R

where the functions F (k, φ) and E(k, φ) are the incomplete elliptic integrals of the first and
second kind, respectively. Moreover, the following holds:
"
EI L
= (32)
R [ F (k, φ2 ) − F (k, φ1 )]

Thus, the value of the end-tip force R in terms of k and φ can be obtained from (32) as:
√
√ EI
R= [ F (k, φ2 ) − F (k, φ1 )] (33)
L
In addition, the bending moments along the rod are given by the following:
√
mi = 2k REI cos φi (34)

The functions F (k, φ) and E(k, φ) depend on two parameters:

• The non-dimensional parameter k
It is known as the modulus of the function and it can vary between −1 and 1. In this
application, k corresponds roughly but non-linearly to the magnitude of the force R.
• The variable φ
It is called the amplitude of the elliptic integral. It is measured in radians, and it
varies continuously along the beam from φ1 on the left edge to φ2 on the right. These
amplitudes, φ1 and φ2 , are defined by the boundary conditions of the rod extremes.
The variable φ is related to the beam angle θ by the following relation:
 
ψ−θ
k sin(φ) = cos (35)
2

27
Mathematics 2021, 9, 1869

Hence, (29) is integrated to yield (30), (31), and (33), i.e., a parametric system that
uniquely defines a deformed rod in terms of parameters k and ψ for known boundary
conditions and a given buckling mode.

2.4.1. Boundary Conditions for a Clamped-Pinned Rod

The studied mechanism in this work has a clamped-pinned rod (as in Figure 2) that
substitutes one of the rigid bars of the four-bar linkage.
The first extreme clamped implies a given slope θ1 for the rod at s = 0; hence,
using (35), we can state the following:
 
1 ψ − θ1
φ1 = arcsin cos (36)
k 2

A pinned end-tip implies that no moment is acting at that point. Upon application
on (34), we obtain the following:
√
0 = 2k REI cos φ2 (37)

This means that the end-tip is a point of null curvature, then an inflection point of the
deflected curve. Additionally, the second part of the equation implies the following:

φ2 = (2q − 1)π/2 for q = 1, 2, ... (38)

being that q is the order of the Buckling Mode shape.

Note that feasible valuesof k may depend on ψ as, for example, is the case of the
ψ−θ
limiting value of k min = ||cos 2
1
|| that is necessary to produce real solutions in (36).
This condition is indicated with the gridded area in Figure 3a. Therefore k is in the
range k = (k min , 1) or k = (−1, −k min ). Provided that range for k, φ1 will have values in
[−π/2, π/2]. On the other hand, φ2 will have discrete values π/2, 3π/2, and so on, for
buckling modes 1, 2 · · · , respectively. The variable φ will vary continuously from φ1 to
φ2 . As shown in Figure 3, if we set a duple of values to k and ψ, inside the feasible range
(Figure 3a), we will obtain the deformed shape of the rod for every buckling mode chosen
(Figure 3b).
From a practical point of view, the data will come from physical quantities that should
be used to find those parameters k and ψ. For example, if end-tip force is given, R and
ψ, along with boundary conditions, (33) should be used to find k, and then the deformed
shape with (30) and (31). If the end-tip position, XP and YP , is given for some boundary
conditions, again the objective will be to find the corresponding parameters k and ψ
from (30) and (31). In most occasions, this process is numerical.
The different problems that can be stated regarding the shape of a rod accomplishing
given conditions may produce a multiplicity of solutions for each buckling mode consid-
ered. For clamped-pinned bars those multiple solutions do not have to be symmetrical, but
refer to completely different configurations. For example, given a same end-point position,
we can obtain different solutions for given buckling modes as shown in Figure 4.

2.4.2. Clamped-Pinned Rod’s Workspace Analysis by Modes

The rod’s end-tip position can be obtained upon substitution of φ2 from (38) into (30)
and (31). These coordinates can be calculated for a range of values of ψ = [0, 2π ] and
k = (−1, −k min ) or k = (k min , 1). This calculation results in a set of curves that is the
workspace of the rod’s end-tip, for positive values of k (Figure 5a) and negative values of k
(Figure 5b). This workspace is obtained by calculating the rod’s end-tip position for given
ψ’s and varying the value of k in the corresponding interval.

28
Mathematics 2021, 9, 1869





 
<


\>P@
N>@
  ;

 5

 


5


             

>GHJ@ [>P@
(a) (b)

Figure 3. (a) Given values of k and ψ. (b) Deformed shapes for buckling modes 1 and 2 of a
clamped-hinged rod.

 

 
\>P@

\>P@

 
< <

 ;  ;

 
         
[>P@ [>P@
(a) (b)

Figure 4. (a) Deformed shape for buckling mode 1. (b) Deformed shape for buckling mode 2. Both
correspond to a clamped-pinned rod for a given end-tip position.


5 






 <

< 2

 ;
\>P@

\>P@

2
 ;








 5

       
[>P@ [>P@

(a) (b)

Figure 5. Rod’s end-tip workspace for buckling mode 1. (a) Positive k. (b) Negative k.

29
Mathematics 2021, 9, 1869

The curves plotted in Figure 5 were calculated for the first buckling mode of the
rod. To obtain the workspace for other buckling modes, it is necessary to execute the
calculation again, changing φ2 in (38), for the desired mode. For instance, Figure 6a shows
the workspace for positive k and buckling mode 2 of the rod, and Figure 6b represents that
for negative k and the same mode.










 5 5
<
< 

2
\>P@

\>P@
2  ;
 ;









       

[>P@ [>P@

(a) (b)

Figure 6. Rod’s end-tip workspace for buckling mode 2. (a) Positive k. (b) Negative k.

As for the orange lines in Figures 5 and 6, they come from the calculation of the rod’s
end-tip position for the higher limit of the range of k and varying ψ. The green lines are
calculated with the lower limit of the interval of k and varying ψ. To close the workspace
area, the purple lines are plotted. They are related to the calculation with the higher limit
of the interval of ψ and varying k. In contrast, the cyan lines derive from the computation
with the lower limit of the interval of ψ and varying k.

3. Closed-Loop Hybrid Mechanisms

Once the fundamentals of elastic rods are understood, this type of element can be
used in conjunction with rigid bars to create new hybrid mechanisms. These mechanisms
are characterized by combining rigid and flexible bars to perform a task, such as describing
a trajectory or performing a given force. In this paper, one of the mechanisms proposed
in [21] (ch. 12), the four-bar linkage with one flexible clamped-pinned rod (see Figure 7), is
the subject of study.

Figure 7. Analysed mechanism in an arbitrary position.

30
Mathematics 2021, 9, 1869

The dimensional data and the description of each bar are given in Table 1. Each angle
showed in Figure 7 is explained in Table 2.

Table 1. Dimensional data of the analyzed mechanism.

Segment Length (m) Description

AB 0.3 Input Link (Crank)
BC 1.2207 Coupler
CD 1 Output Link (Flexible clamped-pinned rod)
BE 0.6103 Distance to place the end-effector bar
EF 1 End-effector bar

Table 2. Angles data of the analyzed mechanism.

Angle Description
α Input angle, which positions the crank
θ1 Orientation of the rigid coupler BC
θ2 Relative angle between BC and EF segments. It has a constant
value of 90º through the cycle.

The flexible segment CD is a slender rod, made of nitinol, with a circular cross-section;
its properties are shown in Table 3.

Table 3. Properties of the flexible rod.

∅ (m) E (Pa) I (m4 )

3 × 10−3 63 × 109 3.9761 × 10−12

3.1. General Equations for the 4-Bar Linkage

The problem to solve is the FKP since the inputs to the mechanism, i.e., the α angle and
the external forces, Fext , and moment Mext , are known, but the position and orientation of
the coupler BC and the shape of the flexible rod CD are unknown.
A shooting method that combines the direct integration with a 4th order R-K method
(explained in Section 2.3) and a N-R scheme is used.
The resolution begins from a known position of the mechanism for an α angle of
0 degrees.
In the next step, a given α is introduced in the calculation. Thus, the crank AB rotates
and the coordinates of B point can be calculated. Then, the coordinates of point C must
be obtained to position the coupler BC. However, this calculus requires a deeper analysis
because the rod CD deforms, so one needs to integrate (22) to obtain the position of C point.
Here, the SM is used to solve the position of the C point and, thus, that of the
mechanism, for each step. This method is focused on the minimization, with the aid of the
N-R scheme, of a residue function that represents all the constraints of the mechanism, and
is given by (39).
⎧ ⎫ ⎧ ⎫
⎪
⎨ m∗z,C ⎪
⎬ ⎨ 0 ⎬
residue = # ∑ Mz  B −  0  (39)
⎪ ⎭ ⎩ BC ⎭
⎩ ( x ∗ − x )2 + ( y ∗ − y )2 ⎪
C B C B


where m∗z,C is the bending moment at C, ∑ Mz  B the set of all moments about B produced
by the external loads, and the superscript * indicates that the variables are not the exact
solutions; instead, they are the approximate solutions obtained numerically in each step of
the SM.

31
Mathematics 2021, 9, 1869

The first two components of (39) are restrictions of the loads kind, and state that the
moments about points B and C must be null because in those points there are hinges. The
other restriction is of the geometric kind and imposes that the distance between points B
and C is the length of the rigid coupler BC.
To find the solution of the mechanism’s position, the SM needs two inputs: the
!T
y
residue function (39) and a vector of guess values, guess = n xguess n guess mzguess , to start
the integration of the rod CD. These guess values are needed to solve the BVP shown in
Section 2.3. The 4th order R-K method explained in that section can be particularized for
the case of the rod CD if mext is equal to 0 since there is a hinge in the point C.
For the first α step, these guess variables can take arbitrary values. Despite this,
previous calculations were made in this work to obtain guess values that improve the
convergence of the SM. For the next α steps, the vector of guess values is composed by the
 T
n x ny mz values of the solution of the previous α step.
With these two inputs to the SM, the process to obtain the position of point C and the
shaped of rod CD is as follows:
!T
1. For each guess, apply 4th order R-K to obtain y∗ = xC∗ yC ∗ θ ∗ n∗ n∗ m∗
C x,C y,C z,C ; see
Section 2.3.
2. Evaluate the residue (39). To do this, xC∗ and yC ∗ should be used to calculate θ , and
 1
∗ ∗  ∗
θ1 , n x,C and ny,C to solve ∑ Mz B . mz,C from the direct integration of the rod CD is
directly used to compare its value with 0.
3. If all terms are below a tolerance, exit: the final solution is obtained.
4. If not, calculate variation of residue with guess: J = ddresidue
guess .
5. Modify the guess values, guessi+1 = guessi − J−1 residue.
6. Start the loop again.
The results are obtained when the three components of the residue are below a tolerance.
With this, the position of the mechanism is obtained for the current α step, and the results
are the guess values for the next α step.

3.2. Load-Dependent Paths

Applying the aforementioned SM, the trajectory of point F can be calculated, but it
depends on external loads because these affect the deformed shape of the flexible rod CD.
Depending on the load applied, point C takes different positions and, in consequence, the
position and the orientation of the coupler BC varies. This influence of the loads in the
path described by point F can be seen in Figure 8 where a set of trajectories for different
load cases is plotted.
If these trajectories are plotted as a function of the moment that the external forces
and moments generate in point B, it results in Figure 9. In Figure 9a, the trajectories plotted
for constant external moments applied to the coupler with no external forces result in
level curves. However, if external forces are applied, the relative position between points
B and F varies during the movement, so the moment of external loading about B takes
different values. For this reason, the trajectories produced are not flat when forces are
applied. Despite this, they belong to the same three-dimensional surface that contains the
previous level set in Figure 9a, as can be seen in Figure 9b.
Hence, by applying this method to solve the FKP of mechanisms of this kind, it is
possible to obtain the trajectories of the end-effector points. Moreover, this method predicts
the three-dimensional surface in which the paths under different load cases are.

32
Mathematics 2021, 9, 1869

Figure 8. Trajectory of point F after a complete turn of the crank AB and under different load cases.

)[)\0
)[)\0 )[)\0
)[)\0 )[)\0
 )[)\0
)[)\0 

>1P@
>1P@



%H[W


%H[W


0




  
 \>P@  
  
[>P@

[>P@ \>P@

(a) (b)

Figure 9. Trajectories for different load cases. (a) External constant moments. (b) External con-
stant forces.

3.3. Buckling Mode Change

Next, if the SM is combined with the integration using EI, explained in Section 2.4, a
powerful algorithm that solves the FKP and shows the workspace of the flexible rod CD is
obtained. In this way, it is possible to know in which buckling mode rod CD is working at
each instant of the work cycle. The process should follow these steps:
1. Integrate the deformed shape using the SM explained in Section 3.1. By doing this,
the distribution of internal moments along the bar is obtained.
2. Search for sign changes of the internal moments along the bar. Each sign change
means a point of curvature change. The number of points of changes of the curvature
corresponds to the buckling mode of the flexible rod. For instance, if there are two
curvature changes, the rod is in the second buckling mode. It must be borne in mind

33
Mathematics 2021, 9, 1869

that, in the case of this mechanism, there is always a point of null moment, i.e., null
curvature, in the distal end of the slender rod CD because of the hinge.
3. Once the buckling mode is known, solve the workspace of the rod CD, applying the
concepts of Section 2.4.2, and plot it.
An example of the application of this algorithm is shown in Figures 10–15 (an anima-
tion is attached in the Supplementary Materials: Video S1), in which a complete turn of
the crank AB is simulated while an external horizontal force of 5 N at point F is applied.
Taking into account that the points of null moment of the rod CD are plotted in red, the
flexible rod CD is working in its first buckling mode when the cycle starts (see Figure 10).
Moreover, k parameter of the EI that allows calculating the workspace is positive. The rod
CD is working in this mode (see Figure 11) until it reaches an α of 159 degrees, where its
buckling mode changes from 1 to 2 (see Figure 12). In addition, k also changes from positive
values to negatives ones. The rod CD remains in its second buckling mode (see Figure 13)
until the input α acquires a value of 285 degrees (see Figure 14). From this pose on, the rod
CD works in its first buckling mode, with negatives k, until the crank AB completes the
turn (see Figure 15).

 ž




)H[W[ 1
$ )
 %
\>P@

 (
<
 ;
' &




      
[>P@

Figure 10. Cycle with buckling mode changes. Initial position.

 ž


)H[W[ 1

 %

$

\>P@


&
<
 ;
'




      
[>P@

Figure 11. Cycle with buckling mode changes. α = 90◦ .

34
Mathematics 2021, 9, 1869

 ž

)H[W[ 1
)


% $


(
\>P@


&
<
 ;
'




      
[>P@
Figure 12. Cycle with buckling mode changes. Mode 1 to Mode 2 change.

 ž


 )H[W[ 1

)
$


%
\>P@


(
<
&
 ;
'




      
[>P@
Figure 13. Cycle with buckling mode changes. α = 235◦ .

35
Mathematics 2021, 9, 1869

 ž




$ )  1
H[W[
 )
\>P@

 %

< (
 ;
'
&



      
[>P@
Figure 14. Cycle with buckling mode changes. Mode 2 to Mode 1 change.

 ž




)  1
$ H[W[

 % )
\>P@


(
<
 ;
' &




      
[>P@
Figure 15. Cycle with buckling mode changes. Final position.

36
Mathematics 2021, 9, 1869

If the trajectory of point F is plotted against the moment of external loading about
B for this load case, Figure 16 is obtained. It can be seen that the path is contained in the
locus exposed on the previous Section 3.2.
It is worth noting that a cut appears on the surface for moments less than −2 Nm. The
potential of this method to obtain the locus in which trajectories can lie for different load
cases is highlighted here. Looking at the shadowed surface of Figure 16, it can be detected
that load cases that generate moments about B less than −2 Nm cannot produce complete
closed trajectories of the point F.
Thus, it is demonstrated that the SM proposed in Section 3.1 can detect and solve
buckling mode changes as long as the trajectory of the end-effector point is contained in its
corresponding locus.

)[)\0


>1P@


%H[W


0

 
  \>P@

[>P@ 
Figure 16. Trajectory for an external applied load of 5 N in the positive x-axis.

3.4. Multiplicity of Mechanism Circuits

Besides buckling mode changes, the method here developed also can find all mecha-
nism circuits. The process to obtain all circuits is the following:
1. Integrate the deformed shape with a 4th order R-K method in points along a cir-
cumference with the center at point B and a radius equal to the length of bar BC. In
this way, all potential solutions of rod CD that satisfy geometric conditions of the
mechanism are calculated. Once the forces in the distal end of the rod CD are known,
continue with the next step.
2. Find the points of the circumference where there is a sign change in the summation of
moments about B. These points are the potential circuits of the mechanism.
3. Apply a SM to refine the potential initial solutions. The function to minimize is (39).
4. Run the SM explained in Section 3.1 to solve the cycle for each calculated circuit.
An example of the application of this process is shown in Figures 17–22. For the case
of the mechanism analyzed in this paper, three circuits appear when external forces of
Fext,x = −4 N and Fext,y = −4 N, and a external moment of Mext,z = 4 Nm are applied
on point F. In the initial position of the first circuit, the rod CD is straight (Figure 17),
and completes a turn without changing the buckling mode, but with a sign change of EI’s
parameter k. As for the second circuit, the rod CD starts and carries out the whole cycle in
its second buckling mode (Figure 19). Finally, if the mechanism moves on its third circuit,
the rod CD works in its first buckling mode (Figure 21). The complete trajectories for the
first, second and third circuits are plotted in Figures 18, 20 and 22, respectively. In these
figures, it can be seen that each circuit generates a specific trajectory.
For a better understanding, Videos S2–S4 are available as Supplementary Materials.

37
Mathematics 2021, 9, 1869

 ž



0 H[W] 1P
$ )  1
H[W[
 % )
)H[W\ 1
\>P@


(
<
 ;
' &




      
[>P@
Figure 17. First mechanism circuit under external loading of Fext,x = −4 N, Fext,y = −4 N, and
Mext,z = 4 Nm. Initial position.

Figure 18. First mechanism circuit under external loading of Fext,x = −4 N, Fext,y = −4 N, and
Mext,z = 4 Nm. Final position and trajectory.

38
Mathematics 2021, 9, 1869

 ž




$
 %
\>P@

 )  1
H[W[ 0  1P
H[W]
( )
< )H[W\ 1
 ;
'
&



      
[>P@
Figure 19. Second mechanism circuit under external loading of Fext,x = −4 N, Fext,y = −4 N, and
Mext,z = 4 Nm. Initial position.

 ž




$
 %
\>P@

 )H[W[ 1 0 H[W] 1P

( )
< )H[W\ 1
 ;
'
&



      
[>P@
Figure 20. Second mechanism circuit under external loading of Fext,x = −4 N, Fext,y = −4 N, and
Mext,z = 4 Nm. Final position and trajectory.

39
Mathematics 2021, 9, 1869

 ž




$
 %

(
\>P@



<
)H[W[ 1 0  1P
 ; H[W]
& ' )
)H[W\ 1




      
[>P@
Figure 21. Third mechanism circuit under external loading of Fext,x = −4 N, Fext,y = −4 N, and
Mext,z = 4 Nm. Initial position.

 ž




$
 %

(
\>P@



<
)H[W[ 1 0  1P
 ; H[W]
& ' )
)  1
H[W\




      
[>P@
Figure 22. Third mechanism circuit under external loading of Fext,x = −4 N, Fext,y = −4 N, and
Mext,z = 4 Nm. Final position and trajectory.

40
Mathematics 2021, 9, 1869

3.5. Algorithm’s Flowchart

To summarize all steps of the entire algorithm, its flowchart is given in Figure 23.

Start End

Inputs: Yes
- Initial unloaded
position
- Fext, Mext
No
Is the last circuit?
Integrate rod in
points along a
circumference with centre
Yes
in B and radius

No
Find points of that Is the final
circumference where there step?
are sign changes of

Refine potential solutions Search for sign changes Solve and plot the
with a SM similar to that of of internal moment of rod
workspace of rod
section 3.1 but with a
to obtain its applying EI and the
condition of buckling mode concepts of section 2.4.2
Yes
Inputs:
- angle No
- Is residue < tolerance?
- Initial position of the
circuit for the first step
- Residue tolerance

Evaluate residue (39)

No - With and Calculate variation of

Is the first
step? - With and residue with guess:

Yes - With

Give arbitrary values to

Integrate deformed shape of rod

with R-K method to get Modify guess values:

equal to the solutions of the

previous step

Figure 23. Flowchart of the entire algorithm.

4. Discussion
In this paper, the performance of the SM that combines the 4th order R-K method
with a numerical minimization method, i.e., N-R scheme, to solve the position of hybrid
compliant mechanisms is shown. This method not only allows solving the FKP, but also

41
Mathematics 2021, 9, 1869

allows predicting the locus of all possible paths of the mechanism’s coupler under different
load cases. In this manner, this method can be used in the synthesis process of mechanisms
to know whether the concerned mechanism can describe the desired trajectory supporting
the external loading.
Furthermore, the SM is combined with the use of EI to obtain the workspace of the
flexible rod, allowing detecting changes of buckling modes.
Future work could look for an algorithm able to detect abrupt changes or loops on
the deformed shape of the compliant element. If this information is known, a highly
detailed knowledge of the mechanism motion can be acquired. All this would allow
designing mechanisms able to foresee the possible problems that could appear in the
prototyping phase.
Further research can be conducted in the multiplicity of circuits of this kind of mecha-
nism, applying the method here proposed. An ambitious goal could be an algorithm to
obtain the loci of all trajectories for different load cases and all the mechanism circuits since,
as can be seen in Figures 17–22, diverse circuits lead to different trajectories’ loci. Detecting
the influence of external loading in these circuits also should be studied.

5. Conclusions
The contribution of this paper is an algorithm that combines two numerical methods
(a shooting method and elliptic integrals) to solve the movement of a hybrid rigid–flexible
mechanism, taking into account the influence of the external load in that movement. The
external load can affect the coupler trajectory by changing the deformed shape of the
slender rod since, depending on the load, the flexible rod can work in different buckling
modes, or even the whole mechanism can operate in different circuits. For this reason, the
proposed algorithm is also developed to detect the circuit and buckling mode in which the
mechanism works.
With this work, the potential of this method to study the behavior of compliant
mechanisms is demonstrated. Its potential is such that it can be used, for instance, in the
modeling process of humanoid robots, such as the one that appears in [23].
Supplementary Materials: The following are available online at https://www.mdpi.com/2227-7
390/9/16/1869/s1: Video S1: Buckling mode change; Video S2: First mechanism circuit; Video S3:
Second mechanism circuit; Video S4: Third mechanism circuit.
Author Contributions: Conceptualization, O.A., D.M.S., E.A. and V.P.; methodology, O.A., D.M.S.,
E.A. and V.P.; software, D.M.S.; validation, O.A., D.M.S., E.A. and V.P.; formal analysis, O.A. and
D.M.S.; investigation, O.A., D.M.S., E.A. and V.P.; resources, O.A., E.A. and V.P.; data curation, O.A.
and D.M.S.; writing—original draft preparation, O.A. and D.M.S.; writing—review and editing, O.A.
and D.M.S.; visualization, O.A.; supervision, O.A., E.A. and V.P.; project administration, O.A. and
V.P.; funding acquisition, O.A. and V.P. All authors have read and agreed to the published version of
the manuscript.
Funding: This work was possible thanks to the funding of Project DPI2015-64450-R and PID2020-
116176GB-100 (MINECO/FEDER, UE), and Project Ref. IT949-16 (Departamento de Educación,
Política Lingüística y Cultura from the regional Basque Government).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Acknowledgments: The authors wish to acknowledge the financial support received from the
Spanish government through the Ministerio de Economía y Competitividad (Project DPI2015-64450-
R and PID2020-116176GB-100 (MINECO/FEDER, UE)), and the support for the research group
through Project Ref. IT949-16, provided by the Departamento de Educación, Política Lingüística y
Cultura from the regional Basque Government.
Conflicts of Interest: The authors declare no conflict of interest.

42
Mathematics 2021, 9, 1869

Abbreviations
Abbreviations
The following abbreviations are used in this manuscript:

BVP Boundary Value Problem

EI Elliptic Integrals
FK Forward Kinematics
FKP Forward Kinematic Problem
IKP Inverse Kinematic Problem
N-R Newton–Raphson
R-K Runge–Kutta
SM Shooting Method

References
1. IFToMM Terminology—Compliant Mechanisms. Available online: http://www.iftomm-terminology.antonkb.nl/2057/frames.
html (accessed on 24 June 2021).
2. Görgülü, İ.; Can Dede, M.İ.; Carbone, G. Experimental Structural Stiffness Analysis of a Surgical Haptic Master Device Manipula-
tor. J. Med. Devices 2021, 15, 011110. [CrossRef]
3. Hamida, I.B.; Laribi, M.A.; Mlika, A.; Romdhane, L.; Zeghloul, S.; Carbone G. On the Optimal Design of LAWEX for a Safe
Upper Arm Rehabilitation Exercising. In Mechanism Design for Robotics; MEDER 2021; Zeghloul, S., Laribi, M.A., Arsicault, M.,
Mechanisms and Machine Science, Eds.; Springer: Cham, Switzerland, 2021; Volume 103, pp. 313–321. [CrossRef]
4. Velázquez, R.; Garzón-Castro, C.L.; Acevedo, M.; Orvañanos-Guerrero, M.T.; Ghavifekr, A.A. Design and Characterization of a
Miniature Bio-Inspired Mobile Robot. In Proceedings of the 2021 12th International Symposium on Advanced Topics in Electrical
Engineering (ATEE), Bucharest, Romania, 25–27 March 2021; pp. 1–5. [CrossRef]
5. Orvañanos-Guerrero, M.T.; Acevedo, M.; Sánchez, C.N.; Giannoccaro, N.I.; Visconti, P.; Velázquez, R. Efficient Balancing Optimiza-
tion of a Simplified Slider-Crank Mechanism. In Proceedings of the 2020 IEEE ANDESCON, Quito, Ecuador, 13–16 October 2020;
pp. 1–6. [CrossRef]
6. Myszka, D.H.; Murray, A.P.; Wampler, C.W. Mechanism branches, turning curves and critical points. In Proceedings of
the IDETC/CIE 2012 ASME 2012 International Design Engineering Technical Conferences & Computers and Information in
Engineering Conference, Chicago, IL, USA, 12–15 August 2012; pp. 1513–1525. [CrossRef]
7. Chase, T.; Mirth, J. Circuits and Branches of Single-Degree-of-Freedom Planar Linkages. ASME J. Mech. Des. 1993, 115, 223–230.
[CrossRef]
8. Mirth, J.; Chase, T. Circuit Analysis of Watt Chain Six-Bar Mechanisms. ASME J. Mech. Des. 1993, 115, 214–222. [CrossRef]
9. Wantanabe, K.; Katoh, H. Identification of Motion Domains of Planar Six-Link Mechanisms of the Stephenson-Type. Mech. Mach.
Theory 2004, 39, 1081–1099. [CrossRef]
10. Ma, F.; Chen, G. Modeling Large Planar Deflections of Flexible Beams in Compliant Mechanisms Using Chained Beam-Constraint-
Model. J. Mech. Robot. 2016, 8, 021018. [CrossRef]
11. Kuo, C.H.; Chen, Y.C.; Pan, T.Y. Continuum Kinematics of a Planar Dual-Backbone Robot Based on Pseudo-Rigid-Body Model:
Formulation, Accuracy, and Efficiency. In Proceedings of the ASME 2017 International Design Engineering Technical Conferences
and Computers and Information in Engineering Conference, Cleveland, OH, USA, 6–9 August 2017; ASME: New York, NY, USA,
2017. [CrossRef]
12. Midha, A.; Bapat, S.G.; Mavanthoor, A.; Chinta, V. Analysis of a Fixed-Guided Compliant Beam With an Inflection Point Using
the Pseudo-Rigid-Body Model Concept. J. Mech. Robot. 2015, 7, 031007. [CrossRef]
13. Antman, S.S. Nonlinear Problems of Elasticity, 2nd ed.; Springer: New York, NY, USA, 2005; Volume 107. [CrossRef]
14. Rucker, D.C.; Webster, R.J., III. Statics and dynamics of continuum robots with general tendon routing and external loading. IEEE
Trans. Robot. 2011, 27, 1033–1044. [CrossRef]
15. Black, C.B.; Till, J.; Rucker, D.C. Parallel Continuum Robots: Modeling, Analysis, and Actuation-Based Force Sensing. IEEE Trans.
Robot. 2018, 34, 29–47. [CrossRef]
16. Altuzarra, O.; Caballero, D.; Campa, F.J.; Pinto, C. Position analysis in planar parallel continuum mechanisms. Mech. Mach.
Theory 2019, 132, 13–29. [CrossRef]
17. Gere, J.M.; Timoshenko, S.P. Mechanics of Materials, 4th ed.; PWS Publishing Company: Boston, MA, USA, 1997.
18. Kimball, C.; Tsai, L.W. Modeling of Flexural Beams Subjected to Arbitrary End Loads. J. Mech. Des. 2002, 124, 223–235. [CrossRef]
19. Holst, G.L.; Teicher, G.H.; Jesen, B.D. Modeling and Experiments of Buckling Modes and Deflection of Fixed-Guided Beams in
Compliant Mechanisms. J. Mech. Des. 2011, 133, 051002. [CrossRef]
20. Zhang, A.; Chen, G. A Comprehensive Elliptic Integral Solution to the Large Deflection Problems of Thin Beams in Compliant
Mechanisms. J. Mech. Robot. 2013, 5, 021006. [CrossRef]
21. Howell, L.L.; Magleby, S.P.; Olsen, B.M. Handbook of Compliant Mechanisms; John Wiley & Sons, Ltd.: Hoboken, NJ, USA, 2013.
[CrossRef]

43
Mathematics 2021, 9, 1869

22. Rucker, D.C.; Jones, B.A.; Webster, R.J. A geometrically exact model for externally loaded concentric-tube continuum robots. IEEE
Trans. Robot. Publ. IEEE Robot. Autom. Soc. 2010, 26, 769–780. [CrossRef] [PubMed]
23. Cafolla, D.; Ceccarelli, M. Design and FEM analysis of a novel humanoid torso. Multibody Mechatron Syst. Mech. Mach. Sci. 2015,
25, 477–488. [CrossRef]

44
 mathematics

Article
Adaptive Levenberg–Marquardt Algorithm: A New
Optimization Strategy for Levenberg–Marquardt
Neural Networks
Zhiqi Yan, Shisheng Zhong *, Lin Lin and Zhiquan Cui *

Department of Mechanical Engineering, Harbin Institute of Technology, Harbin 150000, China;

cumtmaris@163.com (Z.Y.); waiwaiyl@163.com (L.L.)
* Correspondence: zhongss@hit.edu.cn (S.Z.); xiaocui2002yan@163.com (Z.C.)

Abstract: Engineering data are often highly nonlinear and contain high-frequency noise, so the
Levenberg–Marquardt (LM) algorithm may not converge when a neural network optimized by the
algorithm is trained with engineering data. In this work, we analyzed the reasons for the LM neural
network’s poor convergence commonly associated with the LM algorithm. Specifically, the effects of
different activation functions such as Sigmoid, Tanh, Rectified Linear Unit (RELU) and Parametric
Rectified Linear Unit (PRLU) were evaluated on the general performance of LM neural networks,
and special values of LM neural network parameters were found that could make the LM algorithm
converge poorly. We proposed an adaptive LM (AdaLM) algorithm to solve the problem of the LM
algorithm. The algorithm coordinates the descent direction and the descent step by the iteration
Citation: Yan, Z.; Zhong, S.; Lin, L.;
number, which can prevent falling into the local minimum value and avoid the influence of the
Cui, Z. Adaptive Levenberg–Marquardt
parameter state of LM neural networks. We compared the AdaLM algorithm with the traditional LM
Algorithm: A New Optimization
algorithm and its variants in terms of accuracy and speed in the context of testing common datasets
Strategy for Levenberg–Marquardt
and aero-engine data, and the results verified the effectiveness of the AdaLM algorithm.
Neural Networks. Mathematics 2021,
9, 2176. https://doi.org/10.3390/
math9172176
Keywords: Levenberg–Marquardt algorithm; convergence; neural networks; local minima; optimiza-
tion
Academic Editors: Higinio
Rubio Alonso, Alejandro
Bustos Caballero, Jesus
Meneses Alonso and 1. Introduction
Enrique Soriano-Heras When applied to real-world data interspersed with high nonlinearity and high-
frequency noise, LM neural networks have irreplaceable advantages. They reduce the
Received: 16 August 2021 requirement of computing resources and guarantee high convergence speed, making their
Accepted: 1 September 2021 performance superior to other existing regression methods. The LM neural network is
Published: 6 September 2021 an artificial neural network with an LM optimizer. The LM algorithm is a second-order
method for solving general least squares problems. As an optimizer, the LM algorithm has
Publisher’s Note: MDPI stays neutral wide engineering applications in LM neural networks because of its fast convergence and
with regard to jurisdictional claims in
small memory occupation [1–3].
published maps and institutional affil-
However, the disadvantage of the LM algorithm makes LM neural networks have
iations.
potential symptoms of divergence or convergence to a bad local minimum. The paper [4]
reported that the LM algorithm may become stuck and fail to degrade the cost function.
For a nonconvex optimization problem, such as neural network optimization, there can be
many local minima. The cost function may end up in the “bad” local minima when it is
Copyright: © 2021 by the authors. difficult to guarantee the nonsingularity of the Jacobian matrix and Lipschitz continuity [5].
Licensee MDPI, Basel, Switzerland. LM neural networks show higher errors than do regular neural networks in practical
This article is an open access article engineering application [6].
distributed under the terms and
Most of the recent work have tried to help the LM algorithm find one of the “good”
conditions of the Creative Commons
local minima by combining with other algorithms for better engineering applications,
Attribution (CC BY) license (https://
because it is generally acceptable in the modern neural network community that searching
creativecommons.org/licenses/by/
for a global minimum is often an unnecessary endeavor, e.g., the genetic algorithm (GA)
4.0/).

Mathematics 2021, 9, 2176. https://doi.org/10.3390/math9172176 45 https://www.mdpi.com/journal/mathematics

Mathematics 2021, 9, 2176

can be used to optimize the initial weight of LM neural networks [7,8], and Bayesian
estimation can be used to calibrate model variables. In Refs. [9,10], the wavelet method was
used to preprocess data and the LM neural networks method was used to analyze them. In
Ref. [11], the least squares support vector machine (LSSVM) and LM were combined into a
hybrid model. The above hybrid methods have been proved to be effective by experiments.
Several other research groups have looked into improving the LM algorithm. Yang [12]
presented a high-order LM method that has biquadratic convergence. Chen [13] presented
a fourth-order method called the accelerated modified LM method. Derakhshandeh [14]
presented a new three-step LM (TSLM) algorithm based on fuzzy logic theory (FLT). In
Ref. [15], an adaptive Levenberg–Marquardt (LM) algorithm-based echo state network was
proposed by adding a new adaptive damping term to the LM algorithm. These works
proved the validity of the modified method algorithm by a trust region technique.
However, hybrid methods indicate a lot of computation, which weakens the advantage
of the LM algorithm. The problem with improved algorithms is that their performance
improvement is conditionally limited. These improved algorithms prove to be effective
only under specific prerequisites:
(1). F(w) and f (w) are nonlinearly continuous and differentiable. The f (w) is a neural
network model, w is a parameter of the neural network model, and F(w) is the cost
function of the neural network. In Refs. [12–15], the global optimization ability of
the proposed algorithm based on nonlinearly continuous and differential F(w) and
f (w) was verified. In detail, the activation functions in the neural networks are not
necessarily continuous and differentiable, so the above references are only valid in
the current setting without proving that the global optimization can be achieved in
the neural network model.
(2). J = ∂F(w)/∂w. J is a Jacobian matrix. The Jacobian matrix in Refs. [16–18] was obtained
by directly solving the system of nonlinear equations. The calculation results were
relatively accurate, but the calculation process was very complex. In the neural
networks, the Jacobian matrix participating in the proof process was obtained by
back-propagation rather than by the derivation of the objective function. Therefore,
the theory of the above references may be not suitable for the field of neural networks.
Rather, a new adaptive LM algorithm is proposed for neural networks in this paper.
We explain in detail the specific factors that cause the cost function to fall into bad local
minima by the original LM algorithm, by analyzing the output performance of several
activation functions. In view of these factors, the new algorithm makes up for the deficiency
of the LM algorithm to train a network efficiently.
The remaining parts are organized as follows. Section 2 introduces the LM algorithm.
Section 3 proves that the original LM algorithm has the possibility of falling into local
minima in neural networks. Section 4 gives the optimization of the LM algorithm. Section 5
verifies the correctness of this theory about experiments, and the conclusions are drawn in
last section.

2. Explanation and Problem of LM Algorithms

2.1. LM Algorithm
The LM algorithm was first proposed for nonlinear least squares problems by scholars
Levenberg and Marquardt [19]. Before introducing the LM and carrying out research, some
parameters need to be defined, which are the same as those of [20].

Definition 1. Given a neural network model f(w), the cost function is the least squares problem:

F(w) = Σ(ylabel − f (w))2 (1)

where w represents neural network parameters and ylabel represents label data.

46
Mathematics 2021, 9, 2176

Definition 2. A model L is assumed to describe the behavior of F in the current iteration

L(h) ≈ F(w + h) ≡ F(w) + hT J T f (w) + 1 /2 hT J T Jh (2)

Definition 3. There is a method to evaluate the accuracy of model L, called gain ratio q:

F (w) − F (w + h)
q= (3)
L (0) − L ( h )

LM method minimizes F(w) by iteration. h is introduced to describe the change

direction of F(w).

Definition 4. h is a descent direction for F(w) at w, then F(w + h) < F(w).

h = − (J T J + μI)−1 Jr (4)

where J is the Jacoby matrix, r is the cost function error, and μ is the damping parameter.

Given a variable v, the damping parameter μ is adjusted by v in every iteration. The μ

and v updating formulas are illustrated below [20]:
(1). If q < 0, μ = μ × v, v = v × 2;
(2). If q > 0, μ = μ × max{1 /3 , 1 − (2q − 1)3 }.

2.2. Problem of LM Algorithm

The problem of the LM algorithm is from the gain ratio q. A positive q means L(h)
is a good approximation to F(w + h), and vice versa. The L(0)−L(h) has proved to be
positive [20], and its proof is:

L(0) − L(h) = − hT J T f − 12 hT J T Jh
$ %
= − 12 hT 2g + ( J T J + μI − μI )h
= 12 hT (μh − g)
where, g = J T f
 $ (5)
% −1 T
hT h > 0, − hT g = − hT ( J T f ) = − − J T J + μI g g
$ % T
− 1
= gT J T J + μI g>0
L(0) − L(h) = 12 hT (μh − g) > 0

In Definition 3, we know q = (F(w) − F(w + h))/(L(0)−L(h)), which indicates whether

L(h) is a good approximation to F(w + h) dependences on F(w) − F(w + h). As just said, if
q > 0, L(h) is a good approximation. If q < 0, the damping parameter μ in Equation (4) is
adjusted and the current iteration is recalculated.
There is a situation where q will always remain negative regardless of how the damp-
ing parameter μ in Equation (4) is adjusted, while the cost function F(w) in Equation (1) is
still very large.
That situation indicates that iteration cannot continue, and that LM neural networks
have fallen into the “bad” local minima or had divergence.

3. Analysis of Falling into “Bad” Local Minima

In the previous section, the possibility of the LM algorithm was analyzed. In this
section, we focus on the case of where q would always remain negative and analyze the
possible situation. If we find the condition where gain ratio q < 0 regardless of how the

47
Mathematics 2021, 9, 2176

parameters are adjusted, we can prove that LM neural networks may diverge or converge
to the poor local minima.
The proof process includes three steps: First, given the inequality f (w) + J(w) < r(w2 ),
the conditions satisfying the inequality are analyzed. Secondly, we prove that “∃f (w), when
f (w) + J(w) < r(w2 ), in Equation (3), the gain ratio q < 0”. The conditions leading to q < 0 can
be obtained. Thirdly, based on the above analysis, we discuss the conditions under which
LM neural networks may diverge or converge to a worse local minimum.

Definition 5. According to the structure of the neural network, the residual function is as follows:

f (w) = σ(wx + b)−ylabel (6)

where σ is an activation function, x is the input of the neural network, the weight of the neural
network is w, and the bias is b.

Definition 6. The first-order Taylor expansion for the output function f(w + h) of the neural
network at h is:
f (w + h) = f (w) + J(w)h + r(w2 ) (7)
where r(w2 ) is the remainder.

Lemma 1. ∃{w,b,h}, when the absolute values of training data x and label data y of the neural
network are large enough, f(w) + J(w) < r(w2 ).

Proof. According to Equation (7):

r(w2 ) = f (w + h)−f (w)−J(w)h (8)

Substituting Equation (8) into f (w) + J(w) < r(w2 ) gives:

f (w + h)−2f (w)−2J(w)h > 0 (9)

Thus, to prove f (w) + J(w) < r(w2 ), you only need to prove Equation (9). By Equation
(6), we obtain:

f (w + h) = σ(wx + hx + b) − ylabel
(10)
J (w) = σ (wx + b)
Substituting Equation (6) and Equation (10) into Equation (9) gives:

σ (wx + hx + b) − 2σ (wx + b) − 2σ (wx + b)h + ylabel > 0 (11)

Let:

G (w, h) = σ(wx + hx + b) − 2σ(wx + b) − 2σ (wx + b)h + ylabel (12)

Thus, to prove f (w) + J(w) < r(w2 ), you only need to prove:

G (w, h) = σ(wx + hx + b) − 2σ (wx + b) − 2σ (wx + b)h + ylabel > 0 (13)

In order to simplify the calculation, define:

w = wx + b, h = hx (14)
Then:

48
Mathematics 2021, 9, 2176

σ(w ) = σ (wx + b)
σ(w + h ) = σ(wx + hx + b)

∂w 
σ (w )h = σ (wx + b) ∂w  
 h = σ ( wx + b ) x = σ ( wx + b ) h
h
(15)
G (w , h ) = σ(w + h ) − 2σ(w ) − 2σ (w )h + ylabel =
σ (wx + hx + b) − 2σ (wx + b) − 2σ (wx + b)h + ylabel
= G (w, h)
The frequently used activation functions include the sigmoid function, tanh function,
ReLU function, and PReLU function. Discussion:
(1). When σ is sigmoid:

σ (w + h ) = 1
 
1+ e − w − h
σ (w ) = 1
 (16)
1+ e − w

e−w
σ (w ) = 
1+ e − w

Then, the function G(w ,h ) is:


1 2 2he−w
G (w , h ) = −w − h
−  −  + y label (17)
1+e 1 + e−w 1 + e−w
Due to the arbitrariness of the value of ylabel , the first three terms of the function
G(w ,h ) are considered and expressed by g(w ,h ):

1 2 2he−w
g(w , h ) = G (w , h ) − ylabel =   −  −  (18)
1 + e−w −h 1+e − w 1 + e−w
The image of g(w ,h ) in Equation (18) drawn by OCTAVE is shown in Figure 1:

Figure 1. Regardless of the value of w , the g(w ,h ) curve monotonically decreases and intersects the
x-axis at the point (0.5, 0). This means that when the conditions (σ = sigmoid, h ≤ 0.5, and ylabel > 0)
are met, the algorithm has the possibility of divergence.

From Figure 1, it can be seen that the g(w ,h ) curve is monotonously decreasing and
converges at point (h = 0.5, g(w ,h ) = 0); when h ≤ 0.5, g(w ,h ) ≥ 0.That is:

49
Mathematics 2021, 9, 2176

g(w , h )|h ≤0.5 = G (w , h )|h ≤0.5 − ylabel ≥ 0

G (w , h )|h ≤0.5 ≥ ylabel (19)
G (w, h) = G (w , h )|h ≤0.5 ≥ ylabel
Thus, when h ≤ 0.5 and ylabel > 0, G(w,h) > 0. As h = hx, for the arbitrariness of h,
when the absolute values of x and ylabel are large enough, we have G(w,h) > 0, that is, f (w) +
J(w) < r(w2 ).
(2). When σ is tanh:
Similarly, let: w = wx + b, h = hx, then:
   
ew +h −e−w −h
σ (w + h ) =    
ew +h +e−w −h
 
ew −e−w
σ (w ) =  
(20)
ew +e−w
σ (w ) = 4
 
2+e−2w +e2w

Then, the function G(w ,h ) is:

2 4 4h
G (w , h ) =   −  −   − 1 + y label (21)
1 + e 2( w + h ) 1 + e2w 2 + e2w + e−2w
For the arbitrariness of the value of ylabel , the first three terms of the function G(w ,h )
are considered and expressed by g(w ,h ):

g(w , h ) = G (w , h ) − ylabel

(22)
= 2
  − 4
 − 4h
  −1
1+ e2( w + h ) 1+e2w 2+e2w +e−2w

The image of g(w ,h ) in Equation (21) drawn by OCTAVE is shown in Figure 2:

Figure 2. When σ = tanh, the curve g(w ,h ) monotonically decreases: (a) when h < 0.3 and w < 0, lim g(w , h ) = −3;
w→∞
(b) when h < −0.3 and w > 0.5, lim g(w , h ) = −1. When the conditions (σ = tanh, h < 0.3, w < 0, and ylabel > 3) are met,
w→∞
the algorithm has the possibility of divergence.

In Figure 2, the curve g(w ,h ) is also monotonically decreasing. When h < 0.3,
g(w ,h )
≥ −3, and so:

50
Mathematics 2021, 9, 2176

g(w , h )|h ≥−3 = G (w , h )|h ≥−3 − ylabel ≥ −3

G (w , h )|h ≥−3 ≥ ylabel − 3 (23)
G (w, h) = G (w , h )|h ≥−3 ≥ ylabel − 3
Thus, when h’< 0.3 and ylabel > 3, G(w,h) > 0. As h’ = hx, for the arbitrariness of h, when
the absolute values of x and ylabel are large enough, we have G(w,h) > 0, that is, f (w) + J(w)
< r(w2 ).
(3). When σ is ReLU:
Similarly, let: w’= wx + b, h’ = hx, so:

σ (w + h ) = max(0, w + h )
σ(w ) = max(0, w ) (24)
σ (w ) = max(0, 1)
then, G(w ,h ) is:
$ % $ %
G (w , h ) = max 0, w + h − max 0, 2w − max(0, 2h) + ylabel (25)
Due to the arbitrariness of the value of ylabel , the first three terms of the function
G(w ,h ) are considered and expressed by g(w ,h ). The image of g(w ,h ) is shown in
Figure 3:

Figure 3. The curve g(w ,h ) at σ = ReLU: (a) w > 0; (b) w < 0. When the conditions (σ = ReLU, h < 0.3, w < 0, and ylabel >
0) are met, the algorithm has the possibility of divergence.

In Figure 3, when w < 0 and h < 0, g(w , h ) ≥ 0, and so:

 
g(w , h )w <0,h <0 = G (w , h )w <0,h <0 − ylabel ≥ 0

G (w , h )w <0,h <0 ≥ ylabel (26)

G (w, h) = G (w , h )w <0,h <0 ≥ ylabel
Thus, when w < 0, h < 0, and ylabel > 0, G(w,h) > 0. As h = hx and w = wx + b, for the
arbitrariness of h, w, b, when the absolute values of x and ylabel are large enough, we have
G(w,h) > 0, that is, f(w) + J(w) < r(w2 ).
(4). When σ is PReLU:
Similarly, let: w = wx + b and h = hx, and draw the image of g(w ,h ) as shown in
Figure 4:

51
Mathematics 2021, 9, 2176

Figure 4. The curve g(w ,h ) at σ = PReLU. (a) w’ > 0; (b) w < 0. When the conditions (σ = PReLU, h’ < 0, w’ < 0, and ylabel >
0) are met, the algorithm has the possibility of divergence.

In Figure 4, when w < 0, g(w ,h ) > 0, and so:

 
g(w , h )w <0,h <0 = G (w , h )w <0,h <0 − ylabel > 0

G (w , h )w <0,h <0 > ylabel (27)

G (w, h) = G (w , h )w <0,h <0 > ylabel
Thus, when w < 0 and ylabel > 0, G(w,h) > 0. As h’ = hx and w’= wx + b, for the
arbitrariness of h,w,b, when the absolute values of x and ylabel are large enough, we have
G(w,h) > 0, that is, f (w) + J(w) < r(w2 ).
To sum up, when the absolute values of x and ylabel are large enough, there are f (w) +
J(w) < r(w2 ). 

Lemma 2. ∃ f(w), when f(w) + J(w) < r(w2 ), in Equation (3), the gain ratio q < 0.

Proof. The cost function is: F(w + h) = 12 ||f (w + h)||2 . The first-order Taylor expansion for
F(w + h) at w is:

1
F (w + h) = F (w) + hT J T f + hT J T Jh + R(w) = L(h) + R(w) (28)
2
As f (w) + J(w) < r(w2 ), the r(w2 ) term of the residual function f (w + h) after Taylor’s first-
order expansion cannot be omitted. Therefore, there exists f (w) that makes it impossible to
omit the R(w) term of Taylor’s first-order expansion for F(w + h). This means: L(w) < R(w).

∵ L(h) = L(0) + ΔL (29)

In [19], it has been proven that:

ΔL > 0, L(0)  F (w) = 12  f (w)2 > 0

∴ ΔL < L(h) < R(w)
F (w)− F (w+ h)
∴q= L(0)− L(h) (30)
F (w)− F (w)− h T J T f − 12 h T J T Jh− R(w)
= ΔL
ΔL− R(w)
= ΔL < 0

52
Mathematics 2021, 9, 2176

Lemma 3. In the case of Lemma 1 and Lemma 2, the LM algorithm causes cost functions to fall
into “bad” local minima.

Proof. In the kth iteration, we have qk < 0; in the case of Lemma 2, according to the LM
algorithm in the original, when qk < 0, the trust region will be reduced, that is:

v = v×2
μ = μ×v (31)
$ % −1
h = − J T J + μI Jr
At this point, the absolute value of step h decreases, and its sign remains unchanged.
There is a situation where q is always negative when h is at any point in the positive or
negative half-axis, and h will approach zero infinitely and the cost function cannot decrease.
At this time, there is:
q < 0, v → +∞, h → 0 (32)
According to Equation (14), h’ = hx, we have:

h
q < 0, v → +∞, →0 (33)
x
We analyze the common activation functions one by one to find out whether they
have the possibility of this situation:
For the activation function σ = sigmoid, when ylabel > 1, the area covered by the
function is shown in Figure 5a. It can be seen that g(w ,h ) will remain positive when h is
at any point on the negative half of the x-axis. In the case of Lemma 1 and Lemma 2, it is
known that “g(w ,h ) remains positive” means “q remains negative.” According to Equation
(14), h’ = hx, so when ylabel > 1 and hx < 0, q remains negative, and the LM algorithm will
stop running, and the LM neural networks will also diverge or converge poorly.
Similarly, for the activation function σ = tanh, when w ≤ 0 and ylabel > 3, the area
covered by the function is shown in Figure 5b. When w > 0 and ylabel > 1, the area covered
by the function is shown in Figure 5c. Thus, when ylabel > 3, x > 0, and hx < 0, the LM
neural network will diverge or converge poorly.
Similarly, for the activation function σ = ReLU, when w < 0 and ylabel > 0, the area
covered by the function is shown in Figure 5d. When w < 0, ylabel > 0, x > 0, and hx < 0, the
LM neural networks will diverge or converge poorly.
Finally, for the activation function σ = PReLU, when w < 0 and ylabel > 0, the area
covered by the function is shown in Figure 5d. When ylabel > 0 and w < 0, the LM neural
network will diverge or converge poorly. 

Figure 5. Cont.

53
Mathematics 2021, 9, 2176

Figure 5. The gray part is the coverage area of g(w ,h ): (a) The grey area covered by the g(w ,h ) function when σ = sigmoid
and ylabel > 1. If h ≤ 0.5, g(w ,h ) is always positive; (b) The grey area covered by the g(w ,h ) function when σ = tanh, ylabel
> 3, and w ≤ 0. If h ≤ 0.3, g(w ,h ) is always positive; (c) The grey area covered by the g(w ,h ) function when σ = tanh, ylabel
> 1, and w > 0. If h ≤ 0, g(w ,h ) is always positive; (d) The grey area covered by the g(w ,h ) function when σ = ReLU, ylabel
> 0, and w < 0. If h ≤ 0, g(w ,h ) is always positive; (e) The grey area covered by the g(w ,h ) function when σ = PReLU, ylabel
> 0, and w < 0. g(w ,h ) is always positive.

4. The Proposed Algorithm: AdaLM

It can be seen from Section 3 that the parameter state of the model will affect the
stability of the LM algorithm. In order to reduce the influence of the state of the model
on the algorithm, the improved scheme starts from other algorithms to participate in the
iterative operation, and then slowly transitions to the LM algorithm. Therefore, a more
stable algorithm is needed, and the steepest descent method meets this requirement.
With this idea, this paper proposed the AdaLM algorithm that is close to the steepest
descent method at the beginning and close to the LM algorithm after several iterations.
The core step calculation of the AdaLM algorithm can be expressed by Equation (34):

A = JT J
g = JT e (34)
hk = − 1+e1k−10 ( A + μI )−1 g − 1 g
1+e10−k  g

54
Mathematics 2021, 9, 2176

where A = JT J, α is the inertia parameter, V is the accumulated value of h, and k is the

current number of iterations. The damping coefficient μ is calculated from q and v:
(1). If q < 0, μ = (μ + 1) × v, v = v × 2;
(2). If q > 0, μ = 0, v = 2.
If AdaLM reaches the global minimum: k → ∞ and F(x∞ ) = g = 0. Given ε1 > 0, the
first stop criteria can be set to:
g ∞ ≤ ε1 (35)
Consistent with Reference [20], there is a second stop condition if the step size is very
small:
 hk  ≤ ε2 (x +ε2 ) (36)
Finally, in all iterations, we need a protection against infinite loops, so there is a third
stop condition:
k ≥ kmax (37)
Thus, the AdaLM algorithm can be described in Algorithm 1:

Algorithm 1 AdaLM algorithm.

Begin
k = 0; v = 2; x = x0 ; μ = 0;
A = JT J; g = JT f (x);
While (||g||∞ ≥ ε1 ) and (k < kmax )
{
k = k + 1;
hk = − (A +μI) − 1 g/(1 + exp(k −10)) − g/||g||(1 + exp(k −10));
if (|| hk || ≤ ε2 (||x|| + ε2 ))
break;
else
{
xnew = x + hk ;
q = (F(x) − F(xnew ))/(L(0) − L(hk ));
if (q > 0)
{
x = x new ;
A = JT J; g = JT f (x);
if (||g||∞ ≤ ε1 )
break;
μ =0; v = 2;
}
else
{
}μ = (μ + 1) × v; v = v × 2;
}
}
}
End

5. Case Studies
5.1. Experimental Preparation
In order to evaluate the performance of AdaLM, it was arranged to compare with
the original LM algorithm, as well as other popular LM variants, including High order
Levenberg–Marquardt (HLM) and Three Step Levenberg–Marquardt (TSLM). Each al-
gorithm is described in Table 1. In this paper, each algorithm and neural network were
combined to build a prediction model, and the performance level of the algorithm was
reflected by testing the prediction ability of the model.

55
Mathematics 2021, 9, 2176

Table 1. Description table of other LM family algorithms participating in the comparative test.

Algorithm Full Name Description

− 1 J(x )e(x )
LM LM LM(Δxk ) = k )J(xk ) + μI]
[JT (x k k
HLM High order LM HLM(Δxk ) = LM(Δxk ) + LM(Δxk +1 ) + LM(Δxk +2 )
TSLM Three step LM TSLM(Δxk ) = LM(Δxk ) + αLM(Δxk +1 ) + LM(Δxk +2 )

As for the neural network model, it has been proven that the three-layer neural
network can simulate any complex nonlinear mapping if there are enough neurons in the
hidden layer [21–23]. Therefore, the neural network models involved in this paper were
all three-layer neural networks. We found that all the models involved in this paper will
converge before 50 steps of iteration. Thus, we set the maximum number of iterations of
the model to 50. In this study, the mean absolute error (MAE) was used to give the accuracy
of the experimental results. MAE is the average value of absolute error, which can reflect
well the actual situation of the predicted value error. Its formula is as follows:

1 n
n i∑
MAE( f ( x ), y) = | f ( xi ) − yi | (38)
=1

The number of nodes in the model affects the quality of the neural network. The input
and output nodes of the model need to be set according to the requirements of the task.
The number of hidden nodes of the model can be calculated according to the empirical
formula given in [24]: √
m = n×l (39)
where m is the number of hidden layer nodes, n is the number of input layer nodes, and l is
the number of output layer nodes.
The neural network model is shown in Figure 6. The model uses the square difference
formula to calculate the error between the prediction data and the label data.

Figure 6. Neural network model.

56
Mathematics 2021, 9, 2176

In order to test the engineering application ability of each algorithm, the real prediction
task of aero-engine performance parameters was arranged in this paper. In this paper, fuel
flow data of an aero-engine were selected as training data. There is a large noise in the
fuel flow data of the aero-engine due to environmental conditions, operation conditions,
flight tasks, maintenance measures, etc. [25]. In Figure 7, the fuel flow data collected from
the CFM56-5B engine records the fuel consumption rate of the A320 aircraft in pounds per
hour.

Figure 7. Aeroengine fuel flow data.

It is shown from Figure 7 that the data were highly nonlinear with many outliers. The
data sample size was small. Therefore, the selection of the data can reflect the robustness
and convergence of the algorithm. The specific implementation steps of the data prediction
task for the fuel flow of the aeroengine are as follows:
Step 1: Obtain fuel flow data of the aeroengine. Set the total amount of data to N.
Use normalization and data smoothing to preprocess the fuel flow data of the aeroengine,
and then obtain the data X = (x1 , x2 , x3 , . . . . . . , xN ). The purpose of normalization is to
avoid numerical problems such as overflow and underflow, reducing the impact of model
initialization, and improving the prediction accuracy. The purpose of data smoothing is to
reduce the influence of noise and outliers on the model. The data smoothing method is
given by Equation (40):
1 n
MAE( f ( x ), y) = ∑ | f ( xi ) − yi | (40)
n i =1
Step 2: Organize data into datasets. In this step, the dataset is generated according
to the number of input layer nodes n of neural network, and the determination method
of n is given in step 5. The fuel flow data are divided into subsequences with length of n
as the input data of the model and the adjacent element of the subsequence as the label
data. Because this task is a one-step forecast, the label data dimension is 1. Input data X
and label data Y are, respectively:

{x1 , x2 , . . . . . . , xn } →{xn+1 }
{x2 , x3 , . . . . . . , xn+1 } → {x n+2 }
...
(41)
{xi , xi+1 , . . . . . . , xi+n −1 } → {xi+n }
...
{xN −n , xN −n+1 , . . . . . . , xN −1 } → {xN }

Step 3: Split the dataset {X,Y} into training set {X train , Y train } and test set {X test , Y test }.
According to experience, the first 95% of the data are used as training sets. The rest are the
test set.
Step 4: Build the neural network model and use the training set {X train , Y train } to train
the model. The model uses the square difference formula to calculate the error between the
prediction data and the label data.
Step 5: Determine the number of nodes in the neural network model to make the
model optimal. The enumeration method is used to determine the input node n of the

57
Mathematics 2021, 9, 2176

neural network, with the range of 5−12. In detail, steps 2 to 4 are looped 8 times, and the
input node n of the neural network is set to 5 to 12 in step 2 in each loop. The number
of hidden layer nodes is determined by Equation (39), and the number of output nodes
is 1. The MAE error in Equation (38) is used to describe the prediction accuracy of the
experimental results. The number of input nodes and the corresponding model MAE are
shown in Figure 8.

Figure 8. MAE of models with different neural network nodes.

Figure 8 shows that when the number of input nodes n is 9, the output errors of the
algorithm LM reach the minimum, as well as HLM, TSLM, and AdaLM.
Step 6: Test the model. Input the test set X test into the input model to obtain f (X test ).

5.2. The Influence of Each Algorithm on Activation Function

The purpose of this section is to evaluate the impact of the four algorithms on the
activation function to evaluate the performance of each algorithm. In this paper, the AdaLM
algorithm was compared with LM, HLM, and TSLM. As a contrast, two frameworks were
set-up in this study to investigate the performance of these algorithms under different
conditions.
√
The test results are listed in Table 2. “ ” means convergence, “good” means local
minima, “” means “bad” local minima, and “×” means divergence or the term does not
exist.

Table 2. Performance comparison of different activation functions.

Algorithm Year Sigmoid Tanh ReLU PReLU

√
LM 1978 √ × ×
√ ×
HLM 2013 √ 
√ √ 
√
TSLM 2018 √ √ √ √
AdaLM 2020

Table 2 shows that the traditional LM model diverges when Tanh, ReLU, and PReLU
are used as activation functions, which proves that the LM problem exists. HLM is a higher-
order descent method. The HLM model could not converge when the activation functions
were Tanh and PReLU, which indicates that over-high-order descent may produce negative
effects. The TSLM model is an improved version of HLM. In the test, TSLM converged
with four activation functions. The AdaLM algorithm proposed in this paper could change
the descent direction in time and has good stability, so it achieved convergence.

58
Mathematics 2021, 9, 2176

5.3. Evaluation of Prediction Effect

From Figure 9a, the generated curve amplitude was very small with a large error.
From the perspective of the prediction effect, the curve predicted by LM could not describe
the future trend of real data. From Figure 9b,c, it could hardly reproduce the original data.
After training, HLM and TSLM only achieved a line with slight fluctuation. Obviously,
they could not predict the trend of future data in this test. From Figure 9d, the curves
generated by AdaLM could fit the real data well. From the perspective of the prediction
effect, the curve predicted by AdaLM could describe the future trend of real data.

Figure 9. The prediction effect of each algorithm: (a) LM; (b) HLM; (c) TSLM; (d) AdaLM.

In order to quantify the prediction effect of each algorithm, the time consumption in
50 iterations and the prediction MAE of algorithms are shown in Table 3.

Table 3. MAE and time consumption of each algorithm.

Algorithm LM HLM TSLM AdaLM

Time/s 140 181 173 62
MAE 41.4847 38.2276 42.6715 32.9175

From Table 3, the neural network model with the original LM algorithm was trained
in 140 s. The data predicted by the model after training were basically consistent with the
original data, with an error of 41.4847. This shows that the original LM algorithm could
basically make the neural network predict the fuel consumption data of the aeroengine,
but the training took a lot of time because of the falling into a “bad” local minimum.

59
Mathematics 2021, 9, 2176

The neural network model with HLM algorithm completed the training in 181 s. The
data predicted by the model after training fit the original data with an error of 38.2276. The
HLM algorithm took the longest time to train in the prediction of fuel consumption data
of an aeroengine. The HLM algorithm increased the unnecessary complexity and did not
improve the performance.
The neural network model of the TSLM algorithm was trained in 173 s. The data
predicted by the trained model fit the original data with an error of 42.6715. TSLM
algorithm has the highest error in this task, and the complexity of the algorithm affects the
prediction performance.
The neural network model with the AdaLM algorithm proposed in this paper was
trained in 62 s, and the data predicted by the trained model fit the original data with an
error of only 32.9175, which indicates that the AdaLM algorithm took into account high
performance and high speed in predicting the fuel consumption data of aeroengine, with
the least time consumption and the highest accuracy.

6. Conclusions
This paper pointed out that the LM neural networks may converge poorly or diverge
because the training data are large and the weight is not selected properly. This work can
guide researchers to carry out some appropriate strategies such as training data preprocess-
ing and weights intervention at the beginning of training to avoid problems in the use of
the LM neural networks when they still insist on using this model. Furthermore, this work
can expand the application scope of LM neural networks and make LM neural networks
play its role better in more situations.
This study proposed a new solution to the problem of LM neural networks: the
AdaLM algorithm. The algorithm adds the weight term and error direction on the original
algorithm, which makes the algorithm close to the steepest descent method at the beginning
to preliminarily adjust the weight of the neural network. Then, the AdaLM algorithm
gradually approaches the LM algorithm, which improves the stability based on inheriting
the efficient optimization ability of LM. This paper compared the performance of traditional
LM, HLM, TSLM, and AdaLM algorithms. We found that the performance improvement
of the higher-order algorithm was very limited. Among many high-order algorithms, the
third-order algorithm had the better performance. The AdaLM algorithm could not only
make the neural network converge to a “good” local minimum, but also greatly shorten
the operation time without losing the prediction accuracy.
The application of the AdaLM algorithm focused on a small sample and highly non-
linear engineering data. It was especially effective for the analysis and prediction of engine
performance parameters in the aviation field. In future research, we will collect engine
data from more dimensions to comprehensively evaluate AdaLM’s ability of analysis. The
application scenarios of AdaLM will also be further explored.

Author Contributions: Writing—original draft preparation, Z.Y., S.Z., Z.C.; writing—review and
editing, L.L., Z.C; funding acquisition, Z.C. All authors have read and agreed to the published version
of the manuscript.
Funding: This research was funded by the National Natural Science Foundation of China, grant
number U1733201.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

60
Mathematics 2021, 9, 2176

References
1. Luo, G.; Zou, L.; Wang, Z.; Lv, C.; Ou, J.; Huang, Y. A novel kinematic parameters calibration method for industrial robot based
on Levenberg-Marquardt and Differential Evolution hybrid algorithm. Robot. Comput. Integr. Manuf. 2021, 71, 102165. [CrossRef]
2. Kumar, S.S.; Sowmya, R.; Shankar, B.M.; Lingaraj, N.; Sivakumar, S. Analysis of Connected Word Recognition systems using
Levenberg Mar-quardt Algorithm for cockpit control in unmanned aircrafts. Mater. Today Proc. 2021, 37, 1813–1819. [CrossRef]
3. Mahmoudabadi, Z.S.; Rashidi, A.; Yousefi, M. Synthesis of 2D-Porous MoS2 as a Nanocatalyst for Oxidative Desulfuriza-tion of
Sour Gas Condensate: Process Parameters Optimization Based on the Levenberg–Marquardt Algorithm. J. Environ. Chem. Eng.
2021, 9, 105200. [CrossRef]
4. Transtrum, M.K.; Machta, B.B.; Sethna, J.P. Why are nonlinear fits to data so challenging? Phys. Rev. Lett. 2010, 104, 060201.
[CrossRef]
5. Amini, K.; Rostami, F. A modified two steps Levenberg–Marquardt method for nonlinear equations. J. Comput. Appl. Math. 2015,
288, 341–350. [CrossRef]
6. Kim, M.; Cha, J.; Lee, E.; Pham, V.H.; Lee, S.; Theera-Umpon, N. Simplified Neural Network Model Design with Sensitivity
Analysis and Electricity Consumption Prediction in a Commercial Building. Energies 2019, 12, 1201. [CrossRef]
7. Zhao, L.; Otoo, C.O.A. Stability and Complexity of a Novel Three-Dimensional Envi-ronmental Quality Dynamic Evolution
System. Complexity 2019, 2019, 3941920. [CrossRef]
8. Zhou, W.; Liu, D.; Hong, T. Application of GA-LM-BP Neural Network in Fault Prediction of Drying Furnace Equipment. Matec.
Web Conf. 2018, 232, 01041. [CrossRef]
9. Jia, P.; Zhang, P. Type Identification of Coal Mining Face Based on Wavelet Packet Decomposition and LM-BP. In Proceedings
of the 2018 IEEE 9th International Conference on Software Engineering and Service Science (ICSESS), Beijing, China, 23–25
November 2018.
10. Hua, L.; Bo, L.; Tong, L.; Wang, M.; Fu, H.; Guo, R. Angular Acceleration Sensor Fault Diagnosis Based on LM-BP Neural
Network. In Proceedings of the 37th Chinese Control Conference, Wuhan, China, 25–27 July 2018; pp. 6028–6032.
11. Hossein, A.M.; Nazari, M.A.; Madah, R.G.H.; BehshadShafii, M.; Ahmadi, M.A. Thermal conductivity ratio prediction of
Al2O3/water nanofluid by applying connectionist methods. Colloids Surf. A Physicochem. Eng. Asp. 2018, 541, 154–164.
12. Yang, X. A higher-order Levenberg–Marquardt method for nonlinear equations. Appl. Math. Comput. 2013, 219, 10682–10694.
[CrossRef]
13. Chen, L. A high-order modified Levenberg–Marquardt method for systems of nonlinear equations with fourth-order convergence.
Appl. Math. Comput. 2016, 285, 79–93. [CrossRef]
14. Derakhshandeh, S.Y.; Pourbagher, R.; Kargar, A. A novel fuzzy logic Leven-berg-Marquardt method to solve the ill-conditioned
power flow problem. Int. J. Electr. Power Energy Syst. 2018, 99, 299–308. [CrossRef]
15. Qiao, J.; Wang, L.; Yang, C.; Gu, K. Adaptive levenberg-marquardt algorithm based echo state network for chaotic time series
prediction. IEEE Access 2018, 6, 10720–10732. [CrossRef]
16. Ma, C.; Tang, J. The quadratic convergence of a smoothing Levenberg–Marquardt method for nonlinear complementarity problem.
Appl. Math. Comput. 2008, 197, 566–581. [CrossRef]
17. Du, S.Q.; Gao, Y. Global convergence property of modified Levenberg-Marquardt meth-ods for nonsmooth equations. Appl. Math.
2011, 56, 481. [CrossRef]
18. Zhou, W. On the convergence of the modified Levenberg–Marquardt method with a non-monotone second order Armijo type
line search. J. Comput. Appl. Math. 2013, 239, 152–161. [CrossRef]
19. Moré, J.J. The Levenberg-Marquardt algorithm: Implementation and theory. In Numerical Analysis; Springer: Berlin/Heidelberg,
Germany, 1978; pp. 105–116.
20. Madsen, K.; Nielsen, H.B.; Tingleff, O. Methods for Non-Linear Least Squares Problems, 2nd ed.; Technical University of Denmark:
Lyngby, Denmark, 2004.
21. Zhang, Z.; Ma, X.; Yang, Y. Bounds on the number of hidden neurons in three-layer binary neural networks. Neural Netw. 2003,
16, 995–1002. [CrossRef]
22. Liang, X.; Chen, R.C. A unified mathematical form for removing neurons based on or-thogonal projection and crosswise
propagation. Neural Comput. Appl. 2010, 19, 445–457. [CrossRef]
23. Chua, C.G.; Goh, A.T.C. A hybrid Bayesian back-propagation neural network approach to multivariate modelling. Int. J. Numer.
Anal. Methods Geomech. 2003, 27, 651–667. [CrossRef]
24. Sequin, C.H.; Clay, R.D. Fault tolerance in artificial neural networks. In Proceedings of the 1990 IJCNN International Joint
Conference on Neural Networks, San Diego, CA, USA, 17–21 June 1990; pp. 703–708.
25. Li, Y.G.; Nilkitsaranont, P. Gas turbine performance prognostic for condition-based maintenance. Appl. Energy 2009, 86, 2152–2161.
[CrossRef]

61
 mathematics

Article
Three-Legged Compliant Parallel Mechanisms: Fundamental
Design Criteria to Achieve Fully Decoupled Motion
Characteristics and a State-of-the-Art Review
Minh Tuan Pham 1, *, Song Huat Yeo 2 and Tat Joo Teo 3

1 Faculty of Mechanical Engineering, Ho Chi Minh City University of Technology, VNU-HCM,

Ho Chi Minh City 700000, Vietnam
2 School of Mechanical and Aerospace Engineering, Nanyang Technological University,
Singapore 639798, Singapore; myeosh@ntu.edu.sg
3 Robotics, Automation and Unmanned Systems of Home Team Science and Technology Agency,
Singapore 138507, Singapore; daniel_teo@htx.gov.sg
* Correspondence: pminhtuan@hcmut.edu.vn

Abstract: A three-legged compliant parallel mechanism (3L-CPM) achieves fully decoupled motions
when its theoretical 6 × 6 stiffness/compliance matrix is a diagonal matrix, which only contains
diagonal components, while all non-diagonal components are zeros. Because the motion decoupling
capability of 3L-CPMs is essential in the precision engineering field, this paper presents the fundamen-
tal criteria for designing 3L-CPMs with fully decoupled motions, regardless of degrees-of-freedom
and the types of flexure element. The 6 × 6 stiffness matrix of a general 3L-CPM is derived based on
the orientation of each flexure element, e.g., thin/slender beam and notch hinge, etc., and its relative
Citation: Pham, M.T.; Yeo, S.H.; Teo,
position to the moving platform. Based on an analytical solution, several requirements for the flexure
T.J. Three-Legged Compliant Parallel elements were identified and needed to be satisfied in order to design a 3L-CPM with a diagonal
Mechanisms: Fundamental Design stiffness/compliance matrix. In addition, the developed design criteria were used to analyze the
Criteria to Achieve Fully Decoupled decoupled-motion capability of some existing 3L-CPM designs and shown to provide insight into the
Motion Characteristics and a motion characteristics of any 3L-CPM.
State-of-the-Art Review. Mathematics
2022, 10, 1414. https://doi.org/ Keywords: three-legged parallel mechanism; compliant mechanism; flexure-based mechanism;
10.3390/math10091414 flexure; compliant joint; decoupled motion; coupled motion; stiffness; compliance
Academic Editors: Higinio Rubio
Alonso, Alejandro Bustos Caballero, MSC: 70-10
Jesus Meneses Alonso and
Enrique Soriano-Heras

Received: 24 February 2022 1. Introduction

Accepted: 18 April 2022
Compliant (or flexure-based) mechanisms have been widely used to develop pre-
Published: 22 April 2022
cise devices, such as micro-scale manipulation grippers [1,2], nano-scale positioning
Publisher’s Note: MDPI stays neutral actuators [3,4], alignment systems [5–7], etc., due to their frictionless elastic deformation
with regard to jurisdictional claims in behavior. Precise positioning is considered to be one of the most popular applications
published maps and institutional affil- for a compliant mechanism, because it has the ability to deliver repeatable motion with
iations. high positioning resolution, which its traditional counterparts failed to achieve [8]. In the
past few years, compliant mechanisms have been preferred in biological/surgical applica-
tions [9,10], micro-machining systems [11–13], and bulk lithography and microlithography
three-dimensional fabrication processes [14,15]. It is seen that the abilities of producing
Copyright: © 2022 by the authors.
precise and repeatable motions of compliant mechanisms are essential in these advanced
Licensee MDPI, Basel, Switzerland.
This article is an open access article
applications. Thus, the motion property of compliant mechanisms needs to be thoroughly
distributed under the terms and
investigated to further improve their performance.
conditions of the Creative Commons Compliant mechanisms in these applications can be designed through various tech-
Attribution (CC BY) license (https:// niques, such as the pseudo-rigid-body modeling approach [5,16–41], the exact constraint
creativecommons.org/licenses/by/ design approach [42–52], and topological/structural optimization methods [6,7,53–56].
4.0/). In addition, parallel kinematic configuration is commonly adopted to design compliant

Mathematics 2022, 10, 1414. https://doi.org/10.3390/math10091414 63 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 1414

positioning stages, due to its closed-loop and compact architecture. As a result, numerous
three-legged compliant parallel mechanisms (3L-CPMs) with different numbers of degrees-
of-freedom (DOF) have been developed over the past two decades, e.g., the 3-DOF in-plane
motions (X-Y-θ Z ) 3L-CPMs [18–20,22,27,29,35,53,55], the 3-DOF out-of-plane motions (X-Y-
Z) [24,25] and (θ X -θ Y -Z) [5–7] 3L-CPMs, and the 6-DOF motions 3L-CPMs [28,31,42,43,56].
These 3L-CPMs all adopted a three-legged parallel kinematic configuration as a base archi-
tecture. Due to their popularity, this paper focuses on the 3L-CPM with detailed studies on
its motion decoupling capability.
In general, a 3L-CPM consists of three legs, wherein each leg is formed by either one
or a series of flexure elements, e.g., thin/slender beams or notch hinges, etc., connected
together. Hence, each leg can be partially compliant if a rigid-link is used to connect two
flexure elements, or is fully compliant; or if there is no rigid-link between two flexure
elements. Depending on the structure of the leg, 3L-CPMs can be classified into two types,
i.e., single flexure serial chain and double reflecting flexure serial chains in a leg. For
a 3L-CPM or any compliant mechanism, DOF represents the number of possible output
motions that the moving platform can deliver, i.e., three translation motions along and three
rotation motions about the respective X, Y, and Z axes. In an ideal case, the output motions
have to be fully decoupled, i.e., delivering the desired DOF in the actuating directions
and without any parasitic motion in the non-actuating directions. Based on Hooke’s Law,
the motion property (coupled or decoupled) of a compliant mechanism is governed by a
6 × 6 stiffness matrix where the diagonal components represent the stiffness characteristics
of all six possible actuation directions, while the non-diagonal components are responsible
for the off-axes (or non-actuating) stiffness characteristics.
Past works in the literature have shown that the motion properties of many existing 3L-
CPMs were generally neglected [18–22,24,25,27–29,31,33,35,41–43,54]. Based on the derived
6 × 6 stiffness/compliance matrices, only few recent 3L-CPMs demonstrated decoupled
motions [7,52,55,56], while most 3L-CPMs could only deliver coupled motions [5,6]. The
main reason is that the existing 3L-CPMs were synthesized with the aim of achieving the
desired DOF. As a result, they were able to deliver the motions in the desired actuating
directions, but they also produced undesired parasitic motions in the non-actuating di-
rections. More recent efforts have mainly focused on synthesizing 3L-CPMs with a high
ratio between the non-actuating stiffness and actuating stiffness [6,7,53,55], so as to keep
the undesired parasitic motions to a very small percentage as compared to the actuating
motions. In addition, several design criteria for achieving 3L-CPMs with decoupled motion
capability were recently presented in Reference [7]. Such criteria were obtained by substitut-
ing a number of discrete parameters (orientation and position) of flexure elements into the
mathematical model of a 3L-CPM used in a specific structural optimization method [7,56].
Thus, these criteria are not general and cannot be applied in different design methods.
Because motion decoupling is an important performance indicator for any positioning
system, the criteria that can be used to design decoupled-motion 3L-CPMs regardless of
design method and DOF are essential. This paper presents the fundamentals for designing
any 3L-CPMs with fully decoupled motion characteristic. This includes several design
criteria that need to be fulfilled in order to completely eliminate parasitic motions. The
findings that arise from this work suggest that parametric features of flexure elements,
such as the orientation and relative position to the end effector, will have a direct impact
on the performance of any 3L-CPM in terms of the DOF, the constrained motions, and the
parasitic motions.
The remainder of this paper is organized as follows: Section 2 describes the stiffness
modeling of a typical 3L-CPM and the stiffness property of each leg. The criteria of flexure
elements to design a 3L-CPM with fully decoupled motion are presented in Section 3, and
a special case of 3L-CPMs having two reflecting flexure chains in a leg is discussed in
Section 4. Section 5 presents a review on the decoupled-motion capability of an existing
3L-CPM, and Section 6 provides discussions about the findings in this work. Lastly, some
conclusions are offered in Section 7.

64
Mathematics 2022, 10, 1414

2. Stiffness Modeling of a Decoupled-Motion 3L-CPM

In this work, a CPM is represented by a mechanism having three compliant legs that
are distributed symmetrically about the center of the end effector. The legs are fixed at one
end, while the free ends are connected with the end effector. Each legs contains a serial
chain of flexure elements and rigid links, as illustrated in Figure 1. Here, the global frame,
XYZ, is attached to the center of the end effector, and the local frame of each leg, X Y Z , is
attached at the free end of each leg. Note that the X Y plane of the local frame of each leg
lies on the same plane as the XY plane of the global frame.

Figure 1. Construction of a general 3L-CPM.

The stiffness property of a 3L-CPM is governed by the stiffness of the legs and the
moving stage (end-effector). The stiffness matrix of the leg along the Y axis is represented
by Kl with respect to (w.r.t.) the local frame, i.e., at point E, as shown in Figure 1. With D as
the vector that represents the distance between the local frame of the leg and the global
frame of the 3L-CPM, the stiffness matrix of the entire 3L-CPM, Km , is expressed as follows:

3   −1   T
Km = ∑ Jil Ril Kl Ril Jil (1)
i =1

where i = 1, 2, and 3 denotes the three legs in the CPM, as illustrated in Figure 1; Jil is the
translation matrix from the local frame of the ith leg to the global frame; and Ril is the
rotation matrix about the Z axis of the ith leg. As three legs are symmetrical and 120◦ apart,
Ril and Jil are written as follows:
⎡ ⎤
  cos θi − sin θi 0
Rz ( θ i ) 0
Ril = where Rz (θi ) = ⎣ sin θi cos θi 0 ⎦ (2)
0 Rz ( θ i )
0 0 1
⎡ ⎤
  0 Dz i − Dy i
I 0
Jil = where Til = ⎣ − Dz i 0 D xi ⎦ (3)
Til I
Dy i − D xi 0
In Equation (2), the values of θ1 , θ2 , and θ3 are 0◦ , 120◦ , and 240◦ , respectively. In
Equation (3),Dxi , Dyi , and Dzi are three components of Di and represent the projections of
the distance from each local frame to the global frame onto the X, Y, and Z axes, respectively.

65
Mathematics 2022, 10, 1414

Note that Dzi = 0, since the X Y plane of the local frames lies on the same plane as the XY
plane of the global frame.
Here, the stiffness of a general leg, Kl , is represented as follows:
⎡ ⎤
kl11
⎢ kl l ⎥
⎢ 21 k22 SYM ⎥
⎢ l l l ⎥
⎢ k k k ⎥
Kl = ⎢ l31 32 33 ⎥ (4)
⎢ k41 k42 k43 k44
l l l ⎥
⎢ l ⎥
⎣ k51 kl52 kl53 kl54 kl55 ⎦
kl61 kl62 kl63 kl64 kl65 kl66

where the non-diagonal components are symmetrical. By substituting Equations (2)–(4)

into Equation (1), the general form of the stiffness matrix of a 3L-CPM is as follows:
⎡ ⎤
km
11
⎢ 0 km SYM ⎥
⎢ 22 ⎥
⎢ 0 0 km ⎥
K =⎢
m
⎢
33 ⎥
⎥ (5)
⎢ km
41 km
42 0 km
44 ⎥
⎣ km km 0 0 km ⎦
51 52 55
0 0 km
63 0 0 km
66

with d denoting the size of the end effector, as illustrated in Figure 1, the expressions of the
non-zero components within the stiffness matrix of Equation (5) are expressed as follows:
 
11 = k 22 = 2 k 11 + k 22
3
km m l l

33 =
km 3kl33
 
44 = 55 = 2 d k 33 − 2dk 43 + k 44 + k 55
3 2 l
km km l l l
 
66 =
km 3 d2 kl11 + 2dkl61 + kl66 (6)
 
41 = 52 = − 2 dk 31 − k 41 − k 52
3
km km l l l
 
51 = −km42 = 2 k 51 + dk 32 − k 42
3
km l l l
 
63 =
km 3 dkl31 + kl63

By referring to Equation (6), the five non-diagonal components within Km (km m

41 , k 42 ,
km , k m , and k m ) are represented by seven components within Kl (k l , k l , k l , k l , k l ,
51 52 63 31 32 41 42 51
kl52 , and kl63 ). To fulfill the requirements of a fully decoupled motion, all non-diagonal
components in Km must be zeros. In this work, the 6 × 6 stiffness matrix of a 3L-CPM with
fully decoupled motion characteristic is termed as a diagonal stiffness matrix, as shown
in Equation (7).
⎡ m ⎤
k11
⎢ 0 km SYM ⎥
⎢ 22 ⎥
⎢ 0 0 k m ⎥
Km = ⎢⎢ 0
33
m
⎥
⎥ (7)
⎢ 0 0 k44 ⎥
⎣ 0 0 0 0 km ⎦
55
0 0 0 0 0 km
66
In order to obtain a diagonal stiffness matrix as shown in Equation (7), the five non-
diagonal components in Equation (5), for which the expressions are written in Equation (6),
must be zeros. With km 41 = k 52 and k 51 = − k 42 , the relationship between seven variables
m m m

(kl31 , kl32 , kl41 , kl42 , kl51 , kl52 , and kl63 ) in km

41 , k m , k m , k m , and k m can only be represented
42 51 52 63
by three equations, and this condition led to a multiple-solutions problem. Hence, it is
important to note that this work only uses a special case (or solution) to demonstrate how

66
Mathematics 2022, 10, 1414

the presented mathematical models can be used to synthesize a 3L-CPM with the aim
of achieving fully decoupled motion capability. This special case is to make those seven
components within Kl be zeros, as shown in Equation (8).
⎡ ⎤
kl11
⎢ kl l ⎥
⎢ 21 k22 SYM ⎥
⎢ l ⎥
⎢ 0 0 k ⎥
Kl = ⎢ 33 ⎥ (8)
⎢ 0 l
0 k43 k44 l ⎥
⎢ ⎥
⎣ 0 0 kl53 kl54 kl55 ⎦
kl61 kl62 0 kl64 kl65 kl66

In this work, the two popular-choice flexure elements that are used to synthesize the
compliant mechanisms are the beam type and the notch type, as shown in Figure 2. Both
elements have thin features which permit elastic bending in a specific direction.

Figure 2. Original orientation of the flexure elements (solid lines): (a) beam type and (b) notch type,
with the local frames, X”Y”Z”, attached at the free end and the arbitrary orientation of the flexure
elements (dotted-lines) about these local frames.

Assuming that each leg being formed by a serial chain of flexure elements and rigid
links where a rigid link has infinite stiffness (non-compliance) property, the compliance
of each leg (Cl ) is governed by the compliance of each flexure element, Cej , expressed
as follows:
n   −1   T
Cl = ∑ Jej Rej Cej Rej Jej (9)
j =1

where n denotes the number of flexure elements, and Rej and Jej are the rotation matrix, and
translation matrix of the jth flexure element respectively. Referring to Reference [57], the
compliance matrix of each original flexure element, Cej , with respect to the local frame, as
illustrated in Figure 2, is defined as follows:
⎡ e ⎤
c11
⎢ 0 e
c22 SYM ⎥
⎢ ⎥
⎢ 0 0 e
c33 ⎥
Cj = ⎢
e
⎢ e
⎥
⎥ (10)
⎢ 0 0 0 c44 ⎥
⎣ 0 0 e
c53 0 e
c55 ⎦
0 e
c62 0 0 0 e
c66

Equation (10) is applicable for both the beam type and notch type flexure elements, as
illustrated in Figure 2 [57]. In addition, the geometry of each flexure element type can vary
without changing the form of the compliant matrix expressed in Equation (10). Several
kinds of flexure elements which have a similar form of compliance matrix are presented in

67
Mathematics 2022, 10, 1414

Appendix A. Note that the X”Y” plane of the local frame of each flexure element (Figure 2)
lies on the parallel plane with the XY plane of the global frame, as illustrated in Figure 1.
By referring to Equation (9), we see that the rotation matrix, Rej , is a 6 × 6 matrix and is
defined as the multiplication of the rotation matrices about the X”, Y”, and Z” axes (Rex j ,
Rey j and Rez j ). Hence, it is written as follows:
⎡ ⎤
Rez j Rey j Rex j 0 cos γ j − sin γ j 0
Rej = where Rez j = ⎣ sin γ j cos γ j 0 ⎦,
0 Rez j Rey j Rex j
0 0 1
⎡ ⎤
cos β j 0 sin β j
Rey j = ⎣ 0 1 0 ⎦, (11)
− sin β j 0 cos β j
⎡ ⎤
1 0 0
Rx j = ⎣ 0 cos α j − sin α j
e ⎦
0 sin α j cos α j

Here, α j , β j , and γ j represent the rotation angles about the X”, Y”, and Z” axes,
respectively. The geometries of a flexure element before and after orientation are also
illustrated in Figure 2. As for the 6 × 6 translation matrix, Jej , it represents the projected
distances onto the three axes (r x j , ry j , and rz j ) from the jth flexure element to the local frame
of the leg that are indicated by vector r j , as shown in Figure 1, written as follows:
⎡ ⎤
0 rz j −r y j
I rej ⎢ ⎥
Jej = where rej = ⎣ −rz j 0 rxj ⎦ (12)
0 I
ry j −r x j 0

Using Equations (9)–(12), the compliance matrix of a leg, Cl , can be obtained, and the
  −1
stiffness matrix of each leg is given as Kl = Cl .
As mentioned earlier, the stiffness matrix of each leg must follow the exact form shown
in Equation (8), and this requirement applies to its corresponding compliance matrix too.
The detailed derivation of the compliance matrix of each leg with the aim of achieving
that requirement is presented in Appendix B. To summarize the results obtained from
Appendix B, one condition which allows for the compliance matrix of a leg to become the
exact same form as Equation (8) is for seven components within the compliance matrix
l = cl = cl = cl = cl = cl = cl = 0. This condition
of a leg to be zeros, i.e., c31 32 41 42 51 52 63
offers simplicity during the design stage and can be used as the standard approach to
synthesize 3L-CPMs with the aim of achieving fully decoupled motion capability. However,
it also introduces a multiple-solutions problem to solve the corresponding components
within the stiffness matrix, Kl . Among a number of possible solutions, kl64 = kl65 = 0 is a
unique solution that is used to fulfill the condition in this work. By adopting this unique
solution, both the stiffness matrix and the compliance matrix of a leg will have the same
form as expressed in Equation (13), and the expression of each component within Kl is
given in Appendix C.
⎡ ⎤ ⎡ ⎤ −1
kl11 l
c11
⎢ kl21 kl22 ⎥ ⎢ l l ⎥
⎢ SYM ⎥ ⎢ c21 c22 SYM ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 0 kl33 ⎥ ⎢ 0 0 l
c33 ⎥
⎢ ⎥=⎢ ⎥ (13)
⎢ 0 0 kl43 kl44 ⎥ ⎢ 0 0 l
c43 l
c44 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0 kl53 kl54 kl55 ⎦ ⎣ 0 0 l
c53 l
c54 l
c55 ⎦
kl61 kl62 0 0 0 kl66 l
c61 l
c62 0 0 0 l
c66
} }
Kl −1
(Cl )

68
Mathematics 2022, 10, 1414

3. Characteristics of Flexure Elements in Decoupled-Motion 3L-CPMs

In this section, the desired rotation angles (α, β, and γ) and distances (r x , ry , and rz )
of flexure elements in a leg to achieve decoupled-motion capability are analyzed. First,
Equation (9) is re-expressed as follows:
  −1   −1
  −1 n   −1   T n
e
Kl = Cl = ∑ Jej Rej Cej Rej Jej = ∑ Cj (14)
j =1 j =1

e
where C j indicates the compliance matrix of the jth oriented flexure element referring to
the local frame X Y Z attached to the free end of the leg, as illustrated in Figure 1. In order
to achieve fully decoupled motion, the nine non-diagonal components (c31 l , cl , cl , cl , cl ,
41 51 32 42
l , cl , cl , and cl ) of the compliance matrix, Cl , must be zeros, as shown in Equation (13).
c52 63 64 65
e
As Cl is the sum of n sub-components, C j , there could be numerous solutions, because all
e
components within C j can have arbitrary values in general cases. In this work, a special
e e
case where all Cj have the same form (C ) is considered and yields the following:
⎡ e ⎤
c11
⎢ e
c21 e
c22 SYM ⎥
⎢ ⎥
⎢ e
c31 =0 e
c32 = 0 c33
e ⎥
C =⎢ ⎥
e
⎢ ⎥ (15)
⎢
e
c41 =0 e
c42 = 0 c43
e e
c44 ⎥
⎣ e
c51 =0 e
c52 = 0 c53
e e
c54 e
c55 ⎦
e
c61 e
c62 e
c63 = 0 c64
e
= 0 c65
e
= 0 c66
e

e
Note that C in Equation (15) is different from Ce in Equation (10), since
C = Je Re Ce (Re )−1 (Je ) T , as expressed in Equation (14). Re and Je are similar to
e

Equations (11) and (12) with the subscript, j, being removed.

Equation (15) describes the condition to obtain a 3L-CPM with fully decoupled motion
capability, expressed by nine equations that can be obtained based on those nine zero
components. There are six unknowns in these equations, i.e., the rotation angles (α, β, and
γ) and the distances (r x , ry , and rz ) measured from the moving end of the flexure element
to the free end of the leg.
$ %$ %
e
c31 = −c11 e sin β cos β cos γ + ce cos β r cos γ − r sin γ r sin β + r cos β sin γ −
44 y x y z
$ e %  $ % 
cos β sin α cos γ sin α c62 rz − c22e sin β + cos α ce r cos β + ce − ce r sin β sin γ +
62 y 22 62 z
$ e %  $ % 
cos α cos β cos α cos γ c53 rz + c33 e sin β + sin α − ce r cos β + ce + ce r sin β sin γ +
53 y 33 53 z
 $ % $ %
sin α sin β ry cos γ − r x sin γ − cos α r x cos γ + ry sin γ · (16)
$ e %  $ % 
cos α cos γ c55 rz + c53
e sin β + sin α − ce r cos β + ce + ce r sin β sin γ +
55 y 53 55 z
 $ % $ % 
cos γ r x sin α + ry cos α sin β + ry sin α − r x cos α sin β sin γ ·
$ e %  $ % 
cos γ sin α −c66 rz + c62e sin β − cos α ce r cos β + ce − ce r sin β sin γ
66 y 62 66 z
$ %
e
c41 = c44
e cos β cos γ r sin β + r cos β sin γ + (cos γ sin α sin β − cos α sin γ )·
y z
$ e %  $ % 
cos α cos γ c55 rz + c53e sin β + sin α − ce r cos β + ce + ce r sin β sin γ +
55 y 53 55 z
(17)
(cos α cos γ sin β + sin α sin γ)·
$ e %  $ % 
cos γ sin α −c66 rz + c62
e sin β − cos α ce r cos β + ce − ce r sin β sin γ
66 y 62 66 z
$ %
e
c51 = c44
e cos β sin γ r sin β + r cos β sin γ + (cos α cos γ + sin α sin β sin γ )·
y z
$ e %  $ % 
cos α cos γ c55 rz + c53e sin β + sin α − ce r cos β + ce + ce r sin β sin γ +
55 y 53 55 z
(18)
(− cos γ sin α + cos α sin β sin γ)·
$ e %  $ % 
cos γ sin α −c66 rz + c62
e sin β − cos α ce r cos β + ce − ce r sin β sin γ
66 y 62 66 z

69
Mathematics 2022, 10, 1414

$ %
e
c32 = −c11 e cos β sin β sin γ − ce cos β (r cos β cos γ + r sin β ) r cos γ − r sin γ +
44 z x y x
 e $ e % $ e %
cos β sin α cos α c62 r x cos β + cos γ c22 − c62
e r sin β
z + sin α −c62 rz + c22
e sin β sin γ +
 e $ e % $ % 
cos α cos β c53 r x cos β sin α − cos γ sin α c33 + c53 e r sin β + cos α ce r + ce sin β sin γ +
z 53 z 33
 e $ e % $ e %
cos α c66 r x cos β + cos γ c62 − c66
e r sin β
z + sin α −c66 rz + c62
e sin β sin γ · (19)
 $ % $ % 
cos γ r x sin α + ry cos α sin β + ry sin α − r x cos α sin β sin γ +
 e $ e % $ % 
c55 r x cos β sin α − cos γ sin α c53 + c55e r sin β + cos α ce r + ce sin β sin γ ·
z 55 z 53
 $ % $ %
sin α sin β ry cos γ − r x sin γ − cos α r x cos γ + ry sin γ
e
c42 e cos β cos γ (r cos β cos γ + r sin β ) + (cos α cos γ sin β + sin α sin γ )·
= −c44 z x
 e $ e % $ e %
cos α c66 r x cos β + cos γ c62 − c66 rz sin β + sin α −c66
e rz + c62
e sin β sin γ +
(20)
(cos γ sin α sin β − cos α sin γ)·
 e $ e % $ % 
c55 r x cos β sin α − cos γ sin α c53 + c55
e r sin β + cos α ce r + ce sin β sin γ
z 55 z 53
e
c52 = −c44
e cos β (r cos β cos γ + r sin β ) sin γ + (− cos γ sin α + cos α sin β sin γ )·
z x
 e $ e % $ e %
cos α c66 r x cos β + cos γ c62 − c66
e r sin β
z + sin α −c66 rz + c62
e sin β sin γ +
(21)
(cos α cos γ + sin α sin β sin γ)·
 e $ e % $ % 
c55 r x cos β sin α − cos γ sin α c53 + c55
e r sin β + cos α ce r + ce sin β sin γ
z 55 z 53
$ e %  e $ % 
e
c63 = 12 cos β c62 + c53
e cos β sin 2α + 2c44 − c55
e − ce + ce − ce
66 55 66 cos 2α ·
$ % $ e % $ % (22)
sin β −ry cos γ + r x sin γ − c55 − c66 e sin 2α r x cos γ + ry sin γ
  !
e
c64 = cos β cos γ −c44 e
+ c66
e
cos2 α + c55e
sin2 α sin β + (−c55 e
+ c66
e
) cos α sin α sin γ (23)
  !
e
c65 = cos β (c55 e
− c66
e
) cos α cos γ sin α + c66e
cos2 α sin β sin γ + −c44 e
+ c55
e
sin2 α sin β sin γ (24)

This set of equations can be solved by considering Equation (22) first, because β is the
only dominant angular variable, and the results are given as follows:
⎡
β = 90◦
e
c63 ⎣
= 0 ⇔ α = 0◦ and β = 0◦ , 180◦ ; ∀γ, r x , ry , rz (25)
α = 90◦ and β = 0◦ , 180◦

Here, the rotation angle about the X” axis, α, varies from 0◦ to 90◦ , because of the
symmetrical structure of the flexure elements, as illustrated in Figure 2. Equation (25)
shows that there are three possible cases for, c63e
= 0 with the four remaining variables (γ,
r x , ry , and rz ) being arbitrary values.
First, the case with β = 90◦ is considered. With every component within the compli-
ance matrix of the flexure element having a specific value, as shown in Equation (10), six
Equations, from (16) to (21), are always different from zero with any value of γ, r x , ry , and
rz . Hence, β = 90◦ is not a feasible solution.

70
Mathematics 2022, 10, 1414

Next, consider the second case with α = 0◦ and β = 0◦ , 180◦ ; the following results can
be obtained:
 $ % 
e
c31 = ±rz ∓c55 e r cos2 γ ∓ ce r sin2 γ + cos γ ce + ± ce ∓ ce r sin γ
x 44 x 53 44 55 y
$ e %
e
c41 = c44 − c55
e r cos γ sin γ
z
$ %
e
c51 = rz c55
e cos2 γ + ce sin2 γ
44
 e $ e % $ %
e
c32 = ±rz ∓c44 ry cos2 γ + ±c44 e r cos γ sin γ + sin γ ce ∓ ce r sin γ
∓ c55 x 53 55 y (26)
$ %
c42 = −rz c44 cos γ + c55 sin γ
e e 2 e 2
$ e %
e
c52 = −c44 + c55
e r cos γ sin γ
z
e
c64 = c65
e
=0

Here, the upper signs of “±” and “∓” in Equation (26) represent the case of α = 0◦
and β = 0◦ , while the lower signs represent the case of α = 0◦ and β = 180◦ . To solve
e
Equation (26), c51 is first considered to be equal to zero, a unique solution can be obtained
rz = 0 by, and γ can be any value. With rz = 0, all the remaining equations in Equation
(26) will also be equal to zeros. Hence, rz = 0, α = 0◦ , β = 0◦ or 180◦ are solutions used to
obtain a 3L-CPM with any DOF and decoupled motions.
Similarly, for the case with α = 90◦ and β = 0◦ or 180◦ , it can be shown that this is
also a possible solution. In summary, the two feasible solutions found from Equation (25)
with rz = 0 are as follows:

α = 0◦ and β = 0◦ , 180◦
; ∀γ, r x , ry (27)
α = 90◦ and β = 0◦ , 180◦

Equation (27) provides the design criteria for the orientations and positions of the
flexure elements that need to be satisfied to design a 3L-CPM with fully decoupled motion
capability. Figure 3 illustrates the desired orientation of the flexure elements about the X”
and Y” axes, with various orientations about the Z” axis. With flexure elements in a leg
having these orientations and distributing in the X Y plane (rz = 0), a 3L-CPM is able to
achieve fully decoupled motions.

Figure 3. Orientations of the flexure elements in fully decoupled motion 3L-CPMs: (a) beam-type and
(b) notch-type with α = 0◦ , β = 0◦ or 180◦ , and rz = 0, respectively. (c) Beam-type and (d) notch-type
with α = 90◦ , β = 0◦ , 180◦ , and rz = 0 respectively.

71
Mathematics 2022, 10, 1414

4. Stiffness Analysis of 3L-CPMs Containing Two Serial Flexure Chains in a Leg

In Section 2, the stiffness modeling of a general 3L-CPM, with each leg consisting
of a single serial flexure chain, is presented. However, there are many existing 3L-CPMs
synthesized by constraint-based and optimization methods that have two reflecting (or
symmetrical) serial flexure chains, as shown in Figure 4 [6,7,28,31,42,43,53,55,56,58]. In this
section, the analysis of such a leg configuration is presented.

Figure 4. Construction of a 3L-CPM leg containing two reflecting serial flexure chains.

From the literature [6,7,28,31,42,43,53,55,56,58], we can see that the double flexure
chains are either on the same plane or have an offset distance of 2Δ along the Z axis, as
shown in Figure 4. The stiffness matrix of each leg is expressed as follows:
  −1   −1
n n
Kl = ∑ J(+Δ) Csc J(+
T
Δ) + ∑ J(−Δ) MCsc MT J(−
T
Δ) (28)
j =1 j =1

where M is the reflection matrix about the Y Z plane given in Equation (29); and J(+Δ) and
J(−Δ) represent the offset matrices used to shift the original flexure chain and the reflecting
flexure chain along the Z -axis distances of +Δ and −Δ, respectively, as given in Equation (30).
⎡ ⎤
−1 0 0 0 0 0
⎢ 0 1 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 1 0 0 0 ⎥
M=⎢
⎢
⎥
⎥ (29)
⎢ 0 0 0 −1 0 0 ⎥
⎣ 0 0 0 0 1 0 ⎦
0 0 0 0 0 1
⎡ ⎤
1 0 0 0 0 0
⎢ 0 1 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 1 0 0 0 ⎥
J(±Δ) =⎢
⎢
⎥ (30)
⎢ 0 ±Δ 0 1 0 0 ⎥
⎥
⎣ ∓Δ 0 0 0 1 0 ⎦
0 0 0 0 0 1

72
Mathematics 2022, 10, 1414

Here, Csc is the compliance matrix of a serial flexure chain that can be calculated
by Equation (9). The results from Section 3 are used to analyze the stiffness property of
3L-CPMs with two reflecting flexure chains. After substituting Equations (9), (14), (29) and
(30) into Equation (28), the results show that the decoupled motion capability can only be
achieved when the offset distance 2Δ = 0 (or Δ = 0). The offset distance can be considered
as the translation component along the Z axis of each flexure elements (rz j ) that can lead
coupled motions, as mentioned before. Most important, to achieve fully decoupled output
motion, Csc must be in the following form:
⎡ sc ⎤
c11
⎢ sc
c21 sc
c22 SYM ⎥
⎢ ⎥
⎢ 0 0 sc
c33 ⎥
C =⎢
sc
⎢ sc sc
⎥
⎥ (31)
⎢ 0 0 c43 c44 ⎥
⎣ 0 0 sc
c53 sc
c54 sc
c55 ⎦
sc
c61 sc
c62 0 0 0 sc
c66

5. Review and Analysis of the Motion Characteristics of Existing 3L-CPMs

In Sections 2–4, 3L-CPMs constructed with single or double serial flexure chains in a
leg together with their corresponding stiffness modeling and decoupled design criteria were
presented. In this section, the proposed criteria are used to analyze the decoupled-motion
capability of some existing 3L-CPM designs. It can be seen that the analysis provides insight
into the motion characteristics of 3L-CPMs. With the design criteria for synthesizing 3L-
CPMs with fully decoupled motions being provided and the motion-decoupling capability
of popular 3L-CPM designs being clearly defined, designers can select or synthesize any 3L-
CPM with suitable motion property (coupled or decoupled) for their specific applications.
This section presents a short review on the motion characteristics from some of the
existing 3L-CPMs. For each CPM, either the compliance matrix of a leg/serial flexure
chain or the stiffness/compliance matrix of the entire mechanism was derived and used
to determine the theoretical motion characteristics. For 3L-CPMs having a single serial
flexure chain in each leg, fully decoupled motion can be achieved if the compliance matrix
of the leg is in the form of Equation (13). The full stiffness matrix of the entire 3L-CPMs
can then be calculated by using Equations (5), (6) and (14). For 3L-CPMs containing two
reflecting serial flexure chains in each leg, fully decoupled motion can be obtained if the
compliance matrix of a serial flexure chain is similar to Equation (31), and the full stiffness
matrix of entire CPM can be calculated based on Equations (5), (6) and (28). Note that the
analyses were conducted for general cases where most parameters, such as rotations angles
and distances, are symbolic. Therefore, the obtained results are in general forms, while the
results for some specific designs were obtained by substituting values into the parameters.

5.1. Three-Legged Revolute–Revolute–Revolute and Three-Legged Prismatic–Revolute–Revolute CPMs

The Three-legged Revolute–Revolute–Revolute (3RRR) configuration has been the
most popular design for developing 3-DOF 3L-CPMs with (X-Y-θ Z ) planar motions, which
are widely used in positioning/alignment systems [18–20,22,27,29,35,41]. The schematic
diagram of a typical 3RRR-CPM is illustrated in Figure 5a, while Figure 6a illustrates
the three-legged Prismatic–Revolute–Revolute (3PRR) configuration, which is a variant
of 3RRR, where a revolute joint is replaced by a prismatic joint. The 3PRR-CPMs are
preferred to create 3-DOF planar motions in MEMS devices, due to their advantage in
actuation [59,60]. The prismatic joint can be designed as an active joint and easily driven
by a linear actuator. Note that all flexure elements must be in the XY plane to create three
desired motions. The physical prototypes of positioning systems developed based on 3RRR-
and 3PRR-CPMs are shown in Figures 5b and 6b, respectively.

73
Mathematics 2022, 10, 1414

Figure 5. (a) Schematic diagram of 3RRR-CPM and (b) micro-motion stage developed based
on 3RRR-CPM.

Figure 6. (a) Schematic diagram of 3PRR-CPM and (b) a positioning stage developed based
on 3PRR-CPM.

To determine the theoretical motion characteristics of 3RRR-CPM or 3PRR-CPM, the

compliance matrix of each flexure element is represented by C1e , C2e , and C3e , respectively,
and each compliance matrix has a similar form to the one shown in Equation (10). Here, C1e
is the compliance matrix of either the first revolute joint within a 3RRR-CPM or the first
prismatic joint within a 3PRR-CPM. C2e and C3e are the compliance matrices of the second
and third revolute joints, respectively. As illustrated in Figures 5a and 6a, each joint has
a different orientation about the Z” axis, γ j , and a different distance to the moving end
point, r j , with j = 1, 2, and 3. The projections of r j onto the X and Y axes are r x j and ry j ,
respectively. Based on Equation (9), the formula to calculate the compliance matrix of a leg
is written as follows:
3   −1   T
Cl = ∑ Jej Rej Cej Rej Jej (32)
j =1

Jej
where and Rej
are obtained by using Equations (12) and (11), respectively. The result of
Equation (32) is shown in Equation (33), and the detailed expression of each component

74
Mathematics 2022, 10, 1414

is given in Appendix D. It is observed that the compliance matrix of each leg within the
3RRR-CPMs and 3PRR-CPMs is similar to the form expressed in Equation (13). Hence,
this observation suggests that the 3RRR and 3PRR configurations are able to deliver fully
decoupled motion. This performance indicator also highlighted why both configurations
are popular designs for developing state-of-the-art 3L-CPMs.
⎡ ⎤
l
c11
⎢ cl l ⎥
⎢ 21 c22 SYM ⎥
⎢ l ⎥
⎢ 0 0 c ⎥
Cl = ⎢ 33 ⎥ (33)
⎢ 0 l
0 c43 c44 l ⎥
⎢ ⎥
⎣ 0 0 c53 l l
c54 l
c55 ⎦
l
c61 l
c62 0 0 0 l
c66

5.2. Three-Legged Prismatic–Prismatic–Spherical and Three-Legged Revolute–Prismatic–Spherical CPMs

Based on past works in the literature, the three-legged Prismatic–Prismatic–Spherical
(3PPS) [5,32] and the three-legged Revolute–Prismatic–Spherical (3RPS) [21,61] configura-
tions were adopted to develop 3-DOF 3L-CPMs with (θ X -θ Y -Z) out-of-plane motions, as
shown in Figure 7a,b respectively. For both 3L-CPMs, each leg employed three serially
connected flexure elements along the Z axis to create the desired motions. As a result,
the component rz j in Equation (12) exists and generates the off-axis components in the
e
compliance matrix of a flexure element after transformation (C j ). Here, the components
l
within C can be derived from the following:
3
ej
clab = ∑ cab (34)
j =1
ej e
where c ab is a component on row a and column b in C j calculated by Equation (14).

Figure 7. Structure of (a) 3PPS CPM [62] and (b) 3RPS CPM modeled based on the design presented
in [21].

Based on the results obtained from Equation (16) to Equation (24), the components,
ej ej ej ej ej ej
i.e., c31 , c41 , c51 , c32 , c42 , and c52 , are non-zeros because of the existence of rz j . Thus,
the corresponding components in Cl are also non-zeros. Therefore, the legs’ compliance
matrices within the 3PPS- and 3RPS-CPMs do not satisfy Equation (13). As a result, both
the 3-DOF 3PPS- and 3RPS-CPMs will generate coupled motions. This coupled motion
property can be observed from Equation (35), which is the 6 × 6 stiffness matrix of the 3PPS-
CPM (Figure 7a) taken from Reference [5]. Having those five non-diagonal components
indicates that the developed 3PPS-CPM cannot deliver fully decoupled motion.

75
Mathematics 2022, 10, 1414

⎡ ⎤
2.57 × 105
⎢ 0 2.57 × 105 SYM ⎥
⎢ ⎥
⎢ 0 0 3.99 × 105 ⎥
K=⎢
⎢ 5.47 × 101
⎥
⎥ (35)
⎢ 5.18 × 103 0 9.51 × 102 ⎥
⎣ −5.18 × 103 5.47 × 101 0 0 9.51 × 102 ⎦
0 0 −5.63 × 101 0 0 1.72 × 103

5.3. Three-Legged Prismatic–Revolution–Prismatic–Revolution CPMs

Based on past works in the literature, several 3-DOF 3L-CPMs with spatial motions
capability have been developed by using the three-legged Prismatic–Resolution–Prismatic–
Revolution (3PRPR) configuration. Depending on the orientations of the prismatic and
revolute joints, a 3PRPR-CPM can deliver either (X-Y-Z) [24,25] or (θ X -θ Y -Z) [16,17] spatial
motions, as shown in Figure 8a,b respectively. For the 3PRPR-CPM with (X-Y-Z) spatial
motion, as shown in Figure 8a, all the P-joints and R-joints within each leg have their
respective X”Y” planes. The P-joints operate in their respective X”Y” planes, while the
R-joints rotate out of these planes. To analyze its compliance behavior, the leg parallel to the
Y axis of the global frame was selected. Note that detailed modeling follows the procedures
presented in Section 2. Here, each leg consists of four flexure elements (e1 , e2 , e3 , and e4 ),
representing the P-, R-, P-, and R-joints, respectively. The local frame X Y Z is attached at
the free end of the leg. The compliance matrix of each flexure element, Cej (with j = 1, 2, 3,
4), is given as follows:
⎡ P ⎤ ⎡ R ⎤
c11 c11
⎢ 0 P
c22 SYM ⎥ ⎢ 0 R
c22 SYM ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 0 P ⎥ ⎢ R ⎥
C1e = C3e = ⎢
c33 ⎥, and Ce = Ce = ⎢ 0 0 c33 ⎥ (36)
⎢ 0 0 0 P
c44 ⎥ 2 4 ⎢ 0 0 0 R
c44 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0 P
c53 0 P
c55 ⎦ ⎣ 0 0 R
c53 0 R
c55 ⎦
0 P
c62 0 0 0 P
c66 0 R
c62 0 0 0 R
c66

Figure 8. Prototypes of 3PRPR-CPMs with (a) X-Y-Z motions modeled based on the design presented
in [24] and (b) θ X -θ Y -Z motions modeled based on the design presented in [17], respectively.

Based on Figures 2b, 8a, and A1b, the two P-joints can have arbitrary rotation angles
(γ1 and γ3 ) about their Z” axes, while the two R-joints have two rotations, i.e., about the
X” axes with an angle of 90◦ (α2 = α4 = 90◦ ) and about the Z” axes with an angle of 90◦ (γ2
= γ4 = 90◦ ). All flexure elements have no rotation about their Y” axes (β1 = β2 = β3 = β4
= 0). In addition, as the first three flexure elements (e1 , e2 , and e3 ) are located at specific
distances from the local frame of the leg X Y Z , the distance vectors are r1 = r x1 , ry1 , rz1 ,

76
Mathematics 2022, 10, 1414

r2 = r x2 , ry2 , rz2 , and r3 = r x3 , ry3 , rz3 , respectively, as illustrated in Figure 1 and

Equation (12). Moreover, r4 is a zero vector, because the free end of the last R-joint (e4 )
coincides with the frame X Y Z . By substituting these parameters into Equations (9), (11)
and (12), the compliance matrix of the proposed leg can be expressed as follows:
⎡ ⎤
l
c11
⎢ cl l ⎥
⎢ 21 c22 SYM ⎥
⎢ l l l ⎥
⎢ c31 c32 c33 ⎥
C =⎢ l
l
⎥ (37)
⎢ c41 c42 l l
c43 l
c44 ⎥
⎢ l ⎥
⎣ c51 c52 c53 c54
l l l l
c55 ⎦
l
c61 l
c62 0 0 0 l
c66

The expression of each component in Equation (37) is given in Appendix E. It is

observed that the compliance matrix expressed in Equation (37) does not match the form
given in Equation (13) with six non-diagonal components (c31 l , cl , cl , cl , cl , and cl ) not
41 51 32 42 52
equal to zero. This is due to the non-zero translation components along the Z axis of the
first three flexure elements (rz1 , rz2 , rz3 = 0) making the oriented compliance matrix of each
flexure element different from the condition in Equation (15). Thus, the output motions of
the entire 3PRPR 3L-CPM are coupled.
As for the 3PRPR-CPM with θ X -θ Y -Z spatial motion shown in Figure 8b, each leg has
its respective X”Y” plane. Based on the design of each leg, it is observed that the P-joints
(also illustrated in Figure A1b) would operate in their respective X”Y” plane, while the
R-joints were designed to rotate out of that working plane. In addition, the translation of
each flexure element consists of three projection components onto three axes. Similar to the
3-DOF 3PRPR 3L-CPM presented above, this design also generates coupled motions, since
the flexure elements are not distributed in the XY plane (rz j = 0) and the orientations of
the flexure elements also include the rotation out of the X”Y” plane (β j = 0), making the
non-diagonal components within the compliance matrix of each leg become non-zeros, as
proven in Section 3. Lastly, a similar analysis can be adopted to demonstrate that the 3L-
CPM using three serial Prismatic–Universal–Prismatic–Universal (3-PUPU) flexure chains
presented in Reference [33] will also generate coupled motions.

5.4. Three-Legged CPM with Paired Prismatic–Spherical–Spherical Configuration

A 6-DOF 3L-CPM [28] was developed using a pair of Prismatic–Spherical–Spherical
(PSS) serial chain configuration to create each leg as illustrated in Figure 9. By considering
the leg in the front of the 3L-CPM, it is observed that two PSS serial chains located in the
X Z plane and reflect about the Y Z plane. Based on Figure A1, the P-joint is oriented about
the X” axis (α1 = 90◦ ) and Z” axis (γ1 = 90◦ ) respectively to obtain the desired orientation.
In addition, the two S-joints are oriented 45◦ about their Y” axes (β 2 = β 3 = 45◦ ) and the
distance vector from each joint to the local frame X Y Z at the moving end of the leg only
consists of two components, i.e., r1 = {r x1 , 0, rz1 }, r2 = {r x2 , 0, rz2 } and r3 = {r x3 , 0, rz3 }.
Subsequently, the compliance matrix of each flexure element is given as follows:
⎡ P ⎤ ⎡ S ⎤
c11 c11
⎢ 0 P
c22 SYM ⎥ ⎢ 0 S
c22 SYM ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 0 P
c33 ⎥ ⎢ 0 0 S
c33 ⎥
C1e =⎢
⎢ P
⎥, and C = C = ⎢
⎥
e
2
e
3 ⎢ S
⎥
⎥ (38)
⎢ 0 0 0 c44 ⎥ ⎢ 0 0 0 c44 ⎥
⎣ 0 0 P
c53 0 P
c55 ⎦ ⎣ 0 0 S
c53 0 S
c55 ⎦
0 P
c62 0 0 0 P
c66 S S
0 c62 0 0 0 c66

where cuvP and cS (with u, v = 1, 2, . . . , 6) are the components within the compliance
uv
matrices of the P-joints and S-joints, respectively.

77
Mathematics 2022, 10, 1414

Figure 9. Six-DOF 3L-CPM with two reflecting PSS chain in a leg modeled based on the design
presented in [28].

In this design, the two symmetrical PSS serial chains have no offset distance. Hence,
the offset matrices, J(±Δ) , as shown in Equation (30), become identity matrices. By using
Equations (9), (11) and (12) with the defined parameters, the compliance matrix of a serial
flexure chain in the proposed leg, Csc , is written in Equation (39), and the result of each
component is given in Appendix F.
⎡ sc ⎤
c11
⎢ sc
c21 sc
c22 SYM ⎥
⎢ ⎥
⎢ sc
c31 sc
c32 sc
c33 ⎥
C =⎢
sc
⎢ sc sc sc sc
⎥
⎥ (39)
⎢ c41 c42 c43 c44 ⎥
⎣ sc
c51 sc
c52 sc
c53 sc
c54 sc
c55 ⎦
sc
c61 sc
c62 0 0 0 sc
c66

Based on the obtained Csc , the stiffness matrix of each leg, Kl , can be calculated by
Equation (28). It is realized that the form of Csc in this design is different from the condition
shown in Equation (31), as the flexure elements are not located in the XY plane (rz j = 0)
and the P-joint has a rotation about the Y” axis (β 1 = 0). Therefore, even the offset distance
between two serial flexure chains in each leg is zero (Δ = 0), and the output motions of this
6-DOF 3L-CPM are coupled based on the results obtained from Section 4. Using a similar
analysis, the 6-DOF 3L-CPM presented in Reference [31] also generates coupled motions.

5.5. Six-DOF 3L-CPM Synthesized by Constrained-Based Method

A 6-DOF 3L-CPM using double flexure chains configuration, as shown in Figure 10,
was presented in Reference [42]. The construction of each leg contains two horizontal
beam-type flexure elements at both sides and one vertical beam flexure at the center. This
structure can be considered as a reflecting double chain about the Y Z plane, where each
serial flexure chain consists of one horizontal beam and a half of the vertical beam, as
illustrated in Figure 10. Referring to Figure 2a, it is observed that the vertical beam (e2 )
has a rotation of 90◦ about its Z” axis (γ2 = 90◦ ) while the horizontal beam (e1 ) remains
in its original orientation. In addition, the translations of e1 and e2 are represented by
r1 = r x1 , ry1 , 0 and r2 = {0, 0, 0}, respectively, since the free end of e2 coincides with the
frame X Y Z of the leg. The compliance matrices of two flexure elements in a serial flexure
chain are given as follows:

78
Mathematics 2022, 10, 1414

⎡ e ⎤ ⎡ e2 ⎤
c111 c11
⎢ 0
e
c221 SYM ⎥ ⎢ 0 e2
c22 SYM ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 0
e
c331 ⎥ ⎢ e2 ⎥
C1e = ⎢ ⎥, and Ce = ⎢ 0 0 c33 ⎥ (40)
⎢ 0 0 0
e
c441 ⎥ 2 ⎢ 0 0 0 e2
c44 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0
e
c531 0
e
c551 ⎦ ⎣ 0 0 e2
c53 0 e2
c55 ⎦
e e e2 e2
0 c621 0 0 0 c661 0 c62 0 0 0 c66

Figure 10. Six-DOF 3L-CPM modeled based on the design presented in [42].

Based on Equations (9), (11) and (12) and the defined parameters, the obtained compli-
ance matrix of the serial flexure chain shown in Figure 10, Csc , is written as follows:
⎡ e e e ⎤
c111 + c222 + c661 ry21
$ e1 % $ e %
⎢ − c62 e
+ c661 r x1 ry1
e e e
c221 + c112 + r x1 2c621 + c661 r x1 SYM ⎥
⎢ ⎥
⎢ 0 0
e e e e e
c331 + c332 − 2c531 r x1 + c551 r2x1 + c441 ry21 ⎥
Csc = ⎢
⎢ e2 e e e
⎥
⎥ (41)
⎢ 0 0 −c53 + c441 ry1 c441 + c552 ⎥
⎣ 0 0
e e
c531 − c551 r x1 0
e e
c551 + c442 ⎦
e2 e e e e e
−c62 − c661 ry1 c621 + c661 r x1 0 0 0 c661 + c662
It is observed that the compliance matrix presented in Equation (41) matches the form
expressed in Equation (31), and all flexure elements are distributed in the XY plane (Δ = 0
and J(±Δ) become identity matrices). Note that, while c54 sc is zero, those nine essential

non-diagonal components (c31 sc , csc , csc , csc , csc , csc , csc , csc , and csc ) are zero and, thus,
41 51 32 42 52 63 64 65
satisfy the special case proposed in this work. As a result, it shows that this design is able
to produce fully decoupled motion. This can be demonstrated by the resulting diagonal
matrix obtained by applying Equations (1) and (28) to analyze the stiffness property of the
entire 3L-CPM. Several micro-scale 6-DOF 3L-CPMs which developed based on the same
concept were proposed in Reference [43], and a variant of this design was presented in
Reference [58].

5.6. Six-DOF 3L-CPM Synthesized by Optimization Method

A recent 6-DOF 3L-CPM synthesized by the structural optimization method [56] is
shown in Figure 11. Its leg consists of two reflecting flexures about the Y Z plane, and
each flexure is constructed by a curved beam. Here, this curved beam can be considered
as a series of straight beam-type flexure elements (e1 , e2 , . . . , en ) with various orientations
about their respective Z” axes (γ1 , γ2 , . . . , γn ), as illustrated in Figure 3a. Note that each
beam-type flexure element has the compliance matrix in the form shown in Equation (10).

79
Mathematics 2022, 10, 1414

Figure 11. Six-DOF 3L-CPM synthesized by the structural optimization method.

In Figure 11, the translation of each flexure element can be represented by re j =

& '
r x j , ry j , 0 (where j = 1, 2, . . . , n) because all flexure elements are located in the XY plane.
It is observed that the structure of this CPM matches the case shown in Figure 4, with zero
offset distance between two reflecting serial flexure chains (Δ = 0). Thus, this CPM is able
to achieve full decoupling. Referring to the results in Reference [56], we can see that the
stiffness matrix of this CPM has the form of a diagonal matrix, as shown in Equation (42),
and its motion-decoupling capability was also demonstrated.
⎡ ⎤
3.67 × 10−5
⎢
⎢ 0 3.67 × 10−5 SYM ⎥
⎥
⎢ 0 0 9.70 × 10−5 ⎥
C=⎢
⎢
⎥
⎥ (42)
⎢ 0 0 0 3.06 × 10−2 ⎥
⎣ 0 0 0 0 3.06 × 10−2 ⎦
0 0 0 0 0 3.47 × 10−2

5.7. Three-DOF Planar-Motion 3L-CPM Synthesized by Optimization Method

The three-legged configuration with double serial flexure chains within a leg is pre-
ferred in 3L-CPMs that were synthesized by the topological optimization method. Here, an
optimized 3-DOF (X-Y-θ Z ) planar motion 3L-CPM [53] shown in Figure 12a was analyzed
to study its motion property. It is observed that each leg contains two reflecting flexure
chains about the Y Z plane, and each flexure chain has four beam-type flexure elements
(e1 , e2 , e3 , and e4 ) connected together in a series. The thicker segments with each flexure
chain are considered as rigid-body links in the stiffness analysis. With this 3L-CPM being
a planar structure, the translation of each flexure element to the end effector is located
in the XY plane, and there is no projection component (rz j ) in the Z axis. By referring to
Reference [53], the 6 × 6 stiffness matrix of the entire 3L-CPM is as follows:
⎡ ⎤
2.0 × 104
⎢ 0 2.0 × 104 SYM ⎥
⎢ ⎥
⎢ 0 0 2.6 × 106 ⎥
K=⎢
⎢
⎥
⎥ (43)
⎢ 0 −545 0 1.3 × 103 ⎥
⎣ 545 0 0 0 1.3 × 103 ⎦
0 0 0 0 0 12

80
Mathematics 2022, 10, 1414

Figure 12. Three-DOF (X-Y-θ Z ) 3L-CPM designed by using the optimization method (a) structure of
a leg and (b) prototype built based on the design presented in [53].

From Figure 12a, we can see that the 3L-CPM has two reflecting flexure chains with no
offset distance, so that, depending on the results presented in Section 4, its output motions
are decoupled. However, Equation (43) suggests that this 3L-CPM will generate coupled
motions, since there are two non-diagonal components. This result can be explained by
the stiffness modeling used in Reference [53]; the compliance matrices of the leg were
calculated at the local frame X Y Z in the middle plane, while the stiffness matrix of the
entire 3L-CPM was derived at the global frame XYZ (located at the center of the end effector)
in the top plane, as illustrated in Figure 12b. That generates the translation components Dz j
along the Z axis in the translation matrices of the legs, as demonstrated in Equation (3), and
creates the off-axis stiffness components within the final stiffness matrix. Similar designs
presented in Reference [55] will also generate decoupled motions. If the global frame is
attached at the middle plane of the end effector instead, the non-diagonal stiffness will be
eliminated, and the 3L-CPM will have fully decoupled motion characteristic.

5.8. Three-DOF Spatial-Motion 3L-CPMs Synthesized by Optimization Method

Two designs of 3-DOF (θ X -θ Y -Z) spatial-motion 3L-CPMs, as shown in Figure 13, were
synthesized by the structural optimization method [6,7]. Here, the desired motions were
generated by two reflecting beam-type flexure elements about the Y Z plane in each leg.
The design shown in Figure 13a has a small offset distance along the Z axis (Δ = 0) between
two flexure chains, as compared to the other design, which is shown in Figure 13b, where
all flexure elements are located in the same XY plane (Δ = 0).
⎡ ⎤
3.21 × 10−8
⎢
⎢ 0 3.21 × 10−8 SYM ⎥
⎥
⎢ 0 0 8.98 × 10−5 ⎥
⎢
Ca = ⎢ ⎥ (44)
−6 7.84 × 10−8 3.05 × 10−2 ⎥
⎢ −3.31 × 10 0 ⎥
⎣ −7.84 × 10−8 −3.31 × 10−6 0 0 3.05 × 10−2 ⎦
0 0 4.28 × 10−6 0 0 1.50 × 10−4
⎡ ⎤
5.53 × 10−8
⎢
⎢ 0 5.53 × 10−8 SYM ⎥
⎥
⎢ 0 0 1.99 × 10−4 ⎥
Cb = ⎢⎢
⎥
⎥ (45)
⎢ 0 0 0 1.32 × 10−1 ⎥
⎣ 0 0 0 0 1.32 × 10−1 ⎦
0 0 0 0 0 2.33 × 10−4

81
Mathematics 2022, 10, 1414

Figure 13. Three-DOF (θ X -θ Y -Z) 3L-CPMs designed by structural optimization method (a) with offset
distance and (b) without offset distance between two flexure chains in a leg [63].

For the 3L-CPM with Δ = 0, as shown in Figure 13a, the results presented in Refer-
ence [6] agree with the findings described in Section 4, where its compliance matrix, Ca ,
has five non-diagonal components, as derived theoretically in Equation (44). For the other
3L-CPM with Δ = 0, as shown in Figure 13b [7], the compliance matrix, Cb , only has diago-
nal components, as demonstrated theoretically in Equation (45). In summary, the motion
property of 3L-CPMs having two reflecting flexure chains in a leg was demonstrated. The
3L-CPM having an offset distance in the Z axis between two flexure chains will generate
coupled motions, while the other having both flexure chains located in the same plane will
generate decoupled motions.

6. Discussion
In order to deliver fully decoupled motion, a 3L-CPM must have a 6 × 6 diagonal
stiffness/compliance matrix whereby all non-diagonal components are zero. Due to the
property of the parallel architecture, the stiffness matrix of a 3L-CPM can be calculated
based on the stiffness matrices of its legs. However, the stiffness matrix of each leg cannot
be derived directly, since it can be constructed by one or two serial chains; each is formed by
a number of flexure elements and rigid links, and, thus, its characteristic is defined by the
compliance. Due to the challenges in converting between the stiffness and compliance ma-
trices, existing 3L-CPMs failed to analyze their motion property analytically. To overcome
this limitation and, most importantly, for designing a 3L-CPM to obtain a full-decoupled
motion characteristic, the conditions for the compliance matrix of the single serial flexure
chain, i.e., Equation (13), and for the compliance matrix of the double serial flexure chains,
i.e., Equation (31), are provided in this work. A short review of various state-of-the-art
3L-CPMs presented in Section 5 showed that the conditions of the compliance matrix can be
used to identify the motion property of these 3L-CPMs. In order to satisfy these conditions,
analytical analyses show that every flexure element within each leg must be located in
the global XY plane with only two orientations (0◦ and 90◦ ) about its local X” axis, as
illustrated in Figure 3, and the offset distance along the Z axis between two serial flexure
chains, as shown in Figure 4, must be zero (Δ = 0). In other words, these design criteria
can be used to synthesize a 3L-CPM that aims to achieve fully decoupled motion capability.
Moreover, the findings in this work are applicable to any synthesis method, e.g., traditional
pseudo-rigid-body model, constraint-based and topology/structural optimization meth-
ods, etc. Consequently, these design criteria and conditions for the compliance matrices of
the flexure chains can be used as the fundamental design guidelines for the syntheses of
3L-CPMs to achieve desired motion property that can be either decoupled or coupled.

82
Mathematics 2022, 10, 1414

As the 3L-CPM plays an important role in precise motion systems, the motion de-
coupling capability needs to be clearly defined in the design process to make the control
simpler and the output motions more accurate as well. Based on the literature, several
compliant systems which are capable of producing precise motions with a simple control
method, due to the defined motion characteristics of their compliant structures, have been
developed. In particular, the 3-DOF spatial-motion (θ X -θ Y -Z) manipulator [6,7] and the
flexure-based electromagnetic nano-positioning actuator [3] are able to produce a large
workspace with high resolutions, using simple open-loop control systems. Moreover, 3L-
CPMs with decoupled motions can also be applied to design micro-fabrication systems,
e.g., the flexural spindle head in a micro drilling machine tool [11,12], the motion stage in a
micro milling system [13], and the flexural stage to adjust the angles of mirror in advanced
three-dimensional fabrication methods [14,15]. In addition, the benefits offered by decou-
pled 3L-CPMs have been recently exploited in biomedical applications, such as the flexural
micro-dissection device [10]. Since the application range of decoupled-motion 3L-CPMs
has been increasing, it can be said that the fundamental criteria for synthesizing 3L-CPMs
with fully decoupled motions and the motion property of some existing designs presented
in this paper are an important background for developing advanced flexure-based systems.

7. Conclusions
This paper presented the fundamental design criteria for synthesizing any 3L-CPM
with fully decoupled motion capability regardless of the targeted DOF. The stiffness charac-
teristics of a 3L-CPM were analytically modeled. The derived criteria suggested that the
flexure elements in each leg must be distributed in the same plane with the end effector
of the 3L-CPM in order to fulfill the decoupled motions requirements. In the case where
each leg contains two parallel reflecting flexure chains, such requirements are valid if both
flexure chains are located in the same plane with no offset distance. To demonstrate the
effectiveness of the design criteria and conditions obtained from this work, several state-
of-the-art 3L-CPMs were analyzed for their stiffness characteristics and compared with
these criteria. The presented cases show that the proposed design criteria can be applied to
accurately determine the motion characteristics of any 3L-CPM through the analysis of its
stiffness/compliance matrix; only 3L-CPMs having diagonal stiffness/compliance matrices
are able to achieve fully decoupled motions. Findings from this work can be used to define
the motion property of any form of 3L-CPM during the design process.
In this paper, only some special solutions were considered to make the non-diagonal
components within the stiffness/compliance matrix of a 3L-CPM equal to zero; there could
be other solutions that need to be explored. Future work will focus on investigating more
general design criteria for synthesizing any CPMs with desired DOF and motion property.

Author Contributions: Conceptualization, methodology, software, validation, formal analysis, visual-

ization, and writing—original draft preparation, M.T.P.; writing—review and editing, and supervision,
T.J.T. and S.H.Y. All authors have read and agreed to the published version of the manuscript.
Funding: This research is funded by Vietnam National University HoChiMinh City (VNU-HCM),
under grant number C2020-20-01.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A. Stiffness Characteristics of Some Common Types of Flexure Element

Referring to References [22,57], some popular flexure elements have the same compli-
ance matrix form as expressed in Equation (10) and are shown in Figure A1. They can be
revolute hinge; thin beam, as illustrated in Figure 2; and also can be some other forms, such

83
Mathematics 2022, 10, 1414

as spherical joint (Figure A1a) or prismatic joint (linear spring), as shown in Figure A1b.
The notch of spherical joint can have a circular or square cross-sectional area, while the
linear spring can be constructed by four notch hinges or a pair of cantilever beams.

Figure A1. Flexure elements: (a) spherical joint and (b) linear spring.

Appendix B. Conditions of the Leg’s Compliance Matrix for Decoupled Motions

The results of the inversion of Equation (8) are written as follows:
⎡ ⎤ −1 ⎡ ⎤
kl11 l
c11
⎢ kl21 kl22 ⎥ ⎢ l l ⎥
⎢ SYM ⎥ ⎢ c21 c22 SYM ⎥
⎢ ⎥ ⎢ ⎥
⎢ 0 0 kl33 ⎥ ⎢ l
c31 l
c32 l
c33 ⎥
⎢ ⎥ =⎢ ⎥ (A1)
⎢ 0 0 kl43 kl44 ⎥ ⎢ l
c41 l
c42 l
c43 l
c44 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 0 kl53 kl54 kl55 ⎦ ⎣ l
c51 l
c52 l
c53 l
c54 l
c55 ⎦
kl61 kl62 0 kl64 kl65 kl66 l
c61 l
c62 l
c63 l
c64 l
c65 l
c66
} }
−1
(Kl ) Cl

The expression of each component in Cl is given as follows:

1 l l     !
l
c31 = k k − kl21 kl62 kl64 kl53 kl54 − kl43 kl55 + kl65 −kl44 kl53 + kl43 kl54 (A2)
ξ 22 61

1     2   
c41l
= − kl22 kl61 − kl21 kl62 kl64 kl53 − kl33 kl55 + kl65 −kl43 kl53 + kl33 kl54 (A3)
ξ
1      
2

c51l
= − kl22 kl61 − kl21 kl62 kl64 −kl43 kl53 + kl33 kl54 + kl65 kl43 − kl33 kl44 (A4)
ξ
1     !
l
c32 = − kl21 kl61 − kl11 kl62 kl64 kl53 kl54 − kl43 kl55 + kl65 −kl44 kl53 + kl43 kl54 (A5)
ξ

1 l l     2   
l
c42 = k21 k61 − kl11 kl62 kl64 kl53 − kl33 kl55 + kl65 −kl43 kl53 + kl33 kl54 (A6)
ξ
1 l l     
2

l
c52 = k21 k61 − kl11 kl62 kl64 −kl43 kl53 + kl33 kl54 + kl65 kl43 − kl33 kl44 (A7)
ξ
 
1  l 2    !
l
c63 = k21 − kl11 kl22 kl64 kl53 kl54 − kl43 kl55 + kl65 −kl44 kl53 + kl43 kl54 (A8)
ξ
     
1  l 2 2 
l
c64 =− k21 − kl11 kl22 kl64 kl53 − kl33 kl55 + kl65 −kl43 kl53 + kl33 kl54 (A9)
ξ
     
1  l 2  2
l
c65 =− k21 − kl11 kl22 kl64 −kl43 kl53 + kl33 kl54 + kl65 kl43 − kl33 kl44 (A10)
ξ

84
Mathematics 2022, 10, 1414

(  2  2  
l
c21 = 1
ξ −kl21 kl53 kl64 + 2kl43 kl53 kl54 kl61 kl62 + kl21 kl64 kl65 − kl21 kl54 kl66 −
 2   2 
kl43 kl55 kl61 kl62 + kl21 kl65 − kl21 kl55 kl66 +
  2   (A11)
kl33 kl21 kl55 kl64 + kl54 −kl54 kl61 kl62 − kl21 kl64 kl65 + kl21 kl54 kl66 +
      2 )
2
kl44 kl53 −kl61 kl62 + kl21 kl66 + kl33 kl55 kl61 kl62 + kl21 kl65 − kl21 kl55 kl66
(  2  2    )
1 2
l
c61 = −2kl43 kl53 kl54 + kl33 kl54 + kl43 kl55 + kl44 kl53 − kl33 kl55 kl22 kl61 − kl21 kl62 (A12)
ξ
(  2  2    )
1 2
l
c62 =− −2kl43 kl53 kl54 + kl33 kl54 + kl43 kl55 + kl44 kl53 − kl33 kl55 kl21 kl61 − kl11 kl62 (A13)
ξ
(     2
l
c43 = 1
ξ 2kl21 kl53 kl54 − kl43 kl55 kl61 kl62 + kl11 −kl53 kl54 − kl43 kl55 kl62 +
 2   2 
kl21 kl53 kl64 kl65 − kl43 kl65 − kl53 kl54 kl66 + kl43 kl55 kl66 + (A14)
         2 )
2 2
kl22 −kl53 kl54 kl61 + kl11 kl64 kl65 − kl11 kl54 kl66 + kl43 kl55 kl61 + kl11 kl65 − kl11 kl55 kl66
(     2
l
c53 = 1
ξ 2kl21 kl43 kl53 − kl33 kl54 kl61 kl62 + kl11 −kl43 kl53 + kl33 kl54 kl62 −
 2  
kl21 kl33 kl64 kl65 + kl43 kl53 kl66 − kl33 kl54 kl66 + (A15)
        )
2 2
kl22 kl43 kl53 − kl61 + kl11 kl66 + kl33 kl54 kl61 + kl11 kl64 kl65 − kl11 kl54 kl66

where we have the following:

(  2  2   )
2
ξ = −2kl21 −2kl43 kl53 kl54 + kl33 kl54 + kl43 kl55 + kl44 kl53 − kl33 kl55 kl61 kl62 +
(  2  2   ) 
2 2
kl11 −2kl43 kl53 kl54 + kl33 kl54 + kl43 kl55 + kl44 kl53 − kl33 kl55 kl62 +
 2 (  2   2     2   2 
kl21 kl53 − kl64 + kl44 kl66 + 2kl43 kl53 kl64 kl65 − kl54 kl66 + kl43 − kl65 + kl55 kl66 +
    2  2 )
2
kl33 kl55 kl64 − 2kl54 kl64 kl65 + kl44 kl65 + kl54 kl66 − kl44 kl55 kl66 +
(         (A16)
2 2 2
kl22 kl11 kl53 kl64 − 2kl43 kl53 kl54 kl61 + kl11 kl64 kl65 − kl11 kl54 kl66 +
 2   2  2 
kl43 kl55 kl61 + kl11 kl65 − kl11 kl55 kl66 +
  2    
2
kl33 −kl11 kl55 kl64 + kl54 kl54 kl61 + 2kl11 kl64 kl65 − kl11 kl54 kl66 +
         2 )
2 2 2
kl44 kl53 kl61 − kl11 kl66 − kl33 kl55 kl61 + kl11 kl65 − kl11 kl55 kl66

The form of Cl needs to be specified as a standard for the design process of decoupled-
motion 3L-CPMs. It is observed that the expressions of seven compliance components
l , cl , cl , cl , cl , cl , and cl ) corresponding to the seven zero-components in the
(c31 32 41 42 51 52 63
stiffness matrix have similar forms as shown in Equations (A2) to (A8). In this paper,
these seven compliance components are required to be zeros, so that the form of the

85
Mathematics 2022, 10, 1414

leg’s compliance matrix will be the same with its stiffness matrix. This special form
offers simplicity during the design process and can be used as the standard to define the
decoupled-motion capability of various 3L-CPMs. The requirements to make the seven
compliance components equal to zeros are written in Equation (A17).
⎧ ⎧    
⎪
⎪ ⎪
⎪ kl64 kl53 kl54 − kl43 kl55 + kl65 −kl44 kl53 + kl43 kl54 = 0
⎪
⎪ k22 k61 − k21 k62 = 0
l l l l ⎪
⎪
⎪
⎪ ⎪
⎪  
⎨ l l ⎨ l  l 2  
k21 k61 − kl11 kl62 = 0 or k64 k53 − kl33 kl55 + kl65 −kl43 kl53 + kl33 kl54 = 0 (A17)
⎪
⎪  2 ⎪
⎪
⎪
⎪ ⎪
⎪     
⎪
⎪ kl21 − kl11 kl22 = 0 ⎪
⎪ 2
⎩ ⎩ kl64 −kl43 kl53 + kl33 kl44 + kl65 kl43 − kl33 kl44 = 0

As the diagonal components in the stiffness matrix are always non-zeros, while the
non-diagonal components can be zeros or non-zeros, the non-diagonal components are
considered as unknowns, and the diagonal ones are parameters. In the first set of equations,
one of the first two equations can be redundant. The answers to the first set of equations
are as follows: ⎧ #
⎪
⎪
⎨ k21 = ± k11 k22
l l l
" (A18)
⎪
⎪ kl11 l
⎩ kl61 = ± l k 62 k22

The second set of equations contain three equations with five unknowns, so that there
could be many solutions. Here, two simple solutions are proposed, and their answers are
given as follows:
⎧ l l
⎪
⎪ k53 k54 − kl43 kl55 = 0
⎪
⎪ ⎧ #
⎪
⎪ −kl44 kl53 + kl43 kl54 = 0 ⎪
⎪
⎪ ⎪
⎪ kl43 = ± kl33 kl44
* ⎪
⎪  2 * ⎪
⎪
kl64 =0 ⎨ kl64 =0 ⎨ #
or kl53 − kl33 kl55 = 0 ⇔ or kl53 = ± kl33 kl55 (A19)
kl65 =0 ⎪
⎪ kl65 =0 ⎪
⎪
⎪
⎪ ⎪
⎪ #
⎪
⎪ −kl43 kl53 + kl33 kl54 = 0 ⎪
⎩ kl = ± kl kl
⎪
⎪
⎪
⎪   54 44 55
⎩ kl 2 − kl kl = 0
43 33 44

Appendix C. Inversion of the Compliance Matrix of a Leg in a Typical

Decoupled-Motion 3L-CPM
 2
l
c62 − c22
l cl
66
k11 =
l
$ % $ % $ % (A20)
l cl 2 − 2cl cl cl + cl cl 2 + cl 2 cl − cl cl cl
c22 61 21 61 62 11 62 21 66 11 22 66

−c61
l cl + cl cl
kl21 = $ %
62 21 66
$ l %2 l $ l %2 ! (A21)
−2c21
l cl cl + c11
l l 2
c62 + c21 c66 + c22
l c61 − c11
l cl
61 62 66

l cl − cl cl
c22
kl61 = $ %2
61
$ %
21 62
$ % (A22)
l cl
c22 l cl cl + cl cl 2 + cl 2 cl − cl cl cl
− 2c21
61 61 62 11 62 21 66 11 22 66

 2
− c11
l
l cl
c61 66
kl22 = $ %2 $ %2 $ l %2 l (A23)
l cl
c22 61 − 2c21 c61 c62 + c11 c62 + c21 c66 − c11
l l l l l l cl cl
22 66

−c21 l cl + cl cl
kl62 = $ % 2
61 11 62
$ l %2 l $ l %2 ! (A24)
−2c21
l cl cl + cl
61 62 11
l
c62 + c21 c66 + c22
l c61 − c11
l cl
66

86
Mathematics 2022, 10, 1414

 2
− c44
l cl
55
l
c54
kl33 = $ % $ % $ % (A25)
l cl 2 − 2cl cl cl + cl cl 2 + cl 2 cl − cl cl cl
c44 53 43 53 54 33 54 43 55 33 44 55

−c53
l cl + cl cl
kl43 = $ %
54 43 55
$ l %2 l $ l %2 ! (A26)
−2c43
l cl cl + c33
l l 2
c54 + c43 c55 + c44
l c53 − c33
l cl
53 54 55

l cl − cl cl
c44
kl53 = $ %2
53
$ %
43 54
$ % (A27)
l cl
c44 l cl cl + cl cl 2 + cl 2 cl − cl cl cl
− 2c43
53 53 54 33 54 43 55 33 44 55

 2
− c33 l
l cl
c53
55
kl44 = $ %2 $ %2 $ l %2 l (A28)
l cl
c44 53 − 2c43 c53 c54 + c33 c54 + c43 c55 − c33
l l l l l l cl cl
44 55

−c43 l cl + cl cl
kl54 = $ % 2
53 33 54
$ l %2 l $ l %2 ! (A29)
−2c43
l cl cl + cl
53 54 33
l
c54 + c43 c55 + c44
l c53 − c33
l cl
55

 2
− c33
l cl
44
l
c43
kl55 = $ %2 $ % $ % (A30)
l cl
c44 l cl cl + cl cl 2 + cl 2 cl − cl cl cl
− 2c43
53 53 54 33 54 43 55 33 44 55

 2
− c11 l
l cl
c21
22
kl66 = $ %2 $ %2 $ l %2 l (A31)
l cl
c22 61 − 2c21 c61 c62 + c11 c62 + c21 c66 − c11
l l l l l l cl cl
22 66

Appendix D. Compliance Matrix of a Leg in 3RRR- and 3PRR-3L-CPMs

1 3 ej ej ej
 e ej
 ej
!
l
= ∑ c11 + c22 + 2c66 ry2 j + c11 − c22 cos 2γ j + 4c62 ry j sin γ j
j
c11 (A32)
2 j =1

3 & e e e
 e e
 !'
l
c21 = ∑ −c66j r x j ry j − c62j r x j sin γ j + cos γ j −c62j ry j + c11j − c22j sin γ j (A33)
j =1
3  e e

l
c61 = ∑ −c66j ry j − c62j sin γ j (A34)
j =1

1 3 ej ej ej ej
 e ej
 !
2 j∑
l
c22 = c11 + c22 + 2c66 r2x j + 4c62 r x j cos γ j + −c11 + c22 cos 2γ j
j
(A35)
=1
3  ej ej

l
c62 = ∑ c66 r x j + c62 cos γ j (A36)
j =1
3 & e  e ej
   e ej
  
l
c33 = 1
2 ∑ 2c33 + c44 + c55 r2x j + ry2 j − c44 − c55 r x j − ry j r x j + ry j cos 2γ j −
j j j

j =1 (A37)
ej ej
 e ej
 !'
sin γ j − 4r x j cos γ j c53 + c44 − c55 ry j sin γ j
j
4c53 ry j

1 3  ej ej
 ej
 e ej
 !
l
= ∑ c44 + c55 ry j − 2c53 sin γ j + c44 − c55 ry j cos 2γ j − r x j sin 2γ j
j
c43 (A38)
2 j =1

87
Mathematics 2022, 10, 1414

1 3  e e
 e
 e e
 !
2 j∑
l
c53 = − c44j + c55j r x j + 2c53j cos γ j + c44j − c55j r x j cos 2γ j + ry j sin 2γ j (A39)
=1

1 3 ej ej
 e ej
 !
2 j∑
l
c44 = c44 + c55 + c44 − c55 cos 2γ j
j
(A40)
=1

3  ej ej
 !
l
c54 = ∑ c44 − c55 cos γ j sin γ j (A41)
j =1

1 3 ej ej
 e ej
 !
l
= ∑ c44 + c55 + −c44 + c55 cos 2γ j
j
c55 (A42)
2 j =1

3  ej

l
c66 = ∑ c66 (A43)
j =1

Appendix E. Components in the Compliance Matrix of a Leg in the 3-DOF (X-Y-Z)

3PRPR-3L-CPM
  $ P % $ P %
l
c11 = 2c33
R − 2c R r + c R r2 + c P r2 + r2
53 y2 55 y2 66 y1 y3 + c44 rz2 + c11 + c55 rz1 cos γ1 + c11 + c55 rz3 cos γ3 +
R 2 P 2 2 P 2 2
 P $ P %  $ % (A44)
sin γ1 2c62 ry1 + c22 + c44 P r2 sin γ + 2c P r sin γ + c P + c P r2 sin2 γ
z1 1 62 y3 3 22 44 z3 3

$ R % $ %  P $ P % 2
l
c21 = r x2 c53 − c55
Rr
y2 − c66 r x1 ry1 + r x3 ry3 + cos γ1 − c62 ry1 + c11 − c22 + − c44 + c55 rz1 sin γ1 −
P P P P
 $
P (r sin γ + r sin γ ) + cos γ − c P r + c P − c P + − c P + c P r2 sin γ
%  (A45)
c62 x1 1 x3 3 3 62 y3 11 22 44 55 z3 3

 $ % 
l
c31 = −c44R r r − c P r r cos2 γ − c P r r cos2 γ + r cos γ c P + c P − c P r sin γ +
x2 z2 55 x1 z1 1 55 x3 z3 3 z1 1 53 44 55 y1 1
 P $ P %  $ % (A46)
rz3 cos γ3 c53 + c44 − c55
P r sin γ − c P r r sin2 γ + r r sin2 γ
y3 3 44 x1 z1 1 x3 z3 3
 
l
c41 = c44
P
− c55
P
(rz1 cos γ1 sin γ1 + rz3 cos γ3 sin γ3 ) (A47)

l
c51 = c44
R
rz2 + c55
P
rz1 cos2 γ1 + c55
P
rz3 cos2 γ3 + c44
P
rz1 sin2 γ1 + c44
P
rz3 sin2 γ3 (A48)

$ %
l
c61 = 2c53
R
− c55
R
ry2 − c66
P
ry1 + ry3 − c62
P
(sin γ1 + sin γ3 ) (A49)

$ % $ %
l
c22 = 2c11
R + c R r2 + c P r2 + r2 + c R r2 + 2c P r cos γ + c P + c P r2 cos2 γ + 2c P r cos γ +
55 x2 66 x1 x3 66 z2 62 x1 1 22 44 z1 1 62 x3 3
$ P % $ P % 2 $ P % 2 (A50)
c22 + c44 rz3 cos γ3 + c11 + c55 rz1 sin γ1 + c11 + c55 rz3 sin γ3
P 2 2 P 2 P 2

$ R % $ P % $ P %
l
c32 = − c62 + c66Rr
y2 rz2 − c44 ry1 rz1 cos γ1 + c44 − c55 r x1 rz1 cos γ1 sin γ1 + rz1 sin γ1 c53 − c55 ry1 sin γ1 +
P 2 P P
 P $ P % $ % (A51)
rz3 −c44 ry3 cos2 γ3 + c44 − c55P r cos γ sin γ + sin γ c P − c P r sin γ
x3 3 3 3 53 55 y3 3

 
l
c42 = −c66
R
rz2 − c44
P
rz1 cos2 γ1 − c44
P
rz3 cos2 γ3 − c55
P
rz1 sin2 γ1 + rz3 sin2 γ3 (A52)
 
l
c52 = − c44
P
− c55
P
(rz1 cos γ1 sin γ1 + rz3 cos γ3 sin γ3 ) (A53)

l
c62 = c55
R
r x2 + c66
P
(r x1 + r x3 ) + c62
P
(cos γ1 + cos γ3 ) (A54)

88
Mathematics 2022, 10, 1414

$ R % $ %2 $ %
l
c33 = 2c33 P + 2c R + c R r2 + r
22 44 x2 y2 c62 + c66 ry2 + c44 ry1 cos γ1 − r x1 sin γ1
R P − c53
P r cos γ + r sin γ +
x1 1 y1 1
$ %$ P % $
P r cos γ + c P r sin γ + c P r cos γ − r sin γ 2 −
%
r x1 cos γ1 + ry1 sin γ1 −c53 + c55 x1 1 55 y1 1 44 y3 3 x3 3 (A55)
$ % $
P r cos γ + r sin γ + r cos γ + r sin γ
%$ %
3 − c53 + c55 r x3 cos γ3 + c55 ry3 sin γ3
c53 P P P
x3 3 y3 3 x3 3 y3

$ %
l
c43 = 2c62
R + c R r + c P r cos2 γ + c P r cos2 γ + − c P + c P r cos γ sin γ + c P r sin2 γ +
66 y2 44 y1 1 44 y3 3 44 55 x1 1 1 55 y1 1
$ P % (A56)
−c44 + c55 r x3 cos γ3 sin γ3 + c55 ry3 sin γ3 − c53 (sin γ1 + sin γ3 )
P P 2 P

 $ % 
l
c53 = −c44R r − c P r cos2 γ − c P r cos2 γ + cos γ c P + c P − c P r sin γ +
x2 55 x1 1 55 x3 3 1 53 44 55 y1 1
 P $ P %  $ % (A57)
cos γ3 c53 + c44 − c55
P r sin γ − c P r sin2 γ + r sin2 γ
y3 3 44 x1 1 x3 3
   
l
c44 = 2c66
R
+ c44
P
cos2 γ1 + cos2 γ3 + c55 P
sin2 γ1 + sin2 γ3 (A58)
 
l
c54 = c44
P
− c55
P
(cos γ1 sin γ1 + cos γ3 sin γ3 ) (A59)
   
l
c55 = 2c44
R
+ c55
P
cos2 γ1 + cos2 γ3 + c44
P
sin2 γ1 + sin2 γ3 (A60)
 
l
c66 = 2 c66
P
+ c55
R
(A61)

Appendix F. Components in the Compliance Matrix of a Leg in the 6-DOF

6PSS-3L-CPM
1   !
sc
c11 = 2c33
P
+ 2c11
S
+ 2c22
S
+ 2c44
P 2
rz1 + c44
S
+ c55
S
rz22 + rz23 (A62)
2

1 √ S   !
sc
c21 = S
2c11 − 2c22
S
+ 2c53
P
r x1 − 2c62 (r x2 + r x3 ) − c44
S
− c55
S
rz22 + rz23 (A63)
2

1 √ S   !
sc
c31 = −2c44
P
r x1 rz1 + 2c53 (rz2 + rz3 ) − c44 S
+ c55
S
( r x2 r z2 + r x3 r z3 ) (A64)
2
1 S 
sc
c41 = c − c55 S
( r z2 + r z3 ) (A65)
2 44

1   !
sc
c51 = P
2c44 rz1 + c44
S
+ c55
S
( r z2 + r z3 ) (A66)
2
√
sc
c61 = c53
P
− S
2c62 (A67)

1 √ S     !
sc
c22 = c11
P
+ c11
S
+ S
2c22 + 2c55
P 2
r x1 + 2 2c62 (r x2 + r x3 ) + 2c66
S
r2x2 + r2x3 + 2c66 rz1 + c44
P 2 S
+ c55 S
rz22 + rz23 (A68)
2
1 √ S   !
sc
c32 = −c62 P
rz1 + 2c53 (rz2 + rz3 ) + c44
S
− c55
S
( r x2 r z2 + r x3 r z3 ) (A69)
2

1   !
sc
c42 = −2c66
P
rz1 − c44 S
+ c55S
( r z2 + r z3 ) (A70)
2
1 S 
sc
c52 = − c44 − c55
S
( r z2 + r z3 ) (A71)
2
√
sc
c62 = 2c62 S
+ c55
P
r x1 + c66
S
( r x2 + r x3 ) (A72)

89
Mathematics 2022, 10, 1414

1 √ S   !
sc
c33 = P
2c22 + 4c33
S
+ 2c44 r x1 − 2 2c53
P 2
(r x2 + r x3 ) + c44 S
+ c55
S
r2x2 + r2x3 (A73)
2
1 √ S   !
sc
c43 = 2c62 P
− 2 2c53 − c44 S
− c55S
( r x2 + r x3 ) (A74)
2
1 √ S   !
sc
c53 = 2 2c53 − 2c44
P
r x1 − c44 S
+ c55S
( r x2 + r x3 ) (A75)
2
sc
c44 = c66
P
+ c44
S
+ c55
S
(A76)

sc
c54 = c44
S
− c55
S
(A77)

sc
c55 = c44
P
+ c44
S
+ c55
S
(A78)

sc
c66 = c55
P
+ 2c66
S
(A79)

References
1. Wang, D.H.; Yang, Q.; Dong, H.M. A Monolithic Compliant Piezoelectric-Driven Microgripper: Design, Modeling, and Testing.
IEEE/ASME Trans. Mechatron. 2013, 18, 138–147. [CrossRef]
2. Wang, F.; Liang, C.; Tian, Y.; Zhao, X.; Zhang, D. Design and Control of a Compliant Microgripper with a Large Amplification
Ratio for High-Speed Micro Manipulation. IEEE/ASME Trans. Mechatron. 2016, 21, 1262–1271. [CrossRef]
3. Teo, T.J.; Yang, G.; Chen, I.M. A flexure-based electromagnetic nanopositioning actuator with predictable and re-configurable
open-loop positioning resolution. Precis. Eng. 2015, 40, 249–260. [CrossRef]
4. Teo, T.J.; Bui, V.P.; Yang, G.; Chen, I.M. Millimeters-Stroke Nanopositioning Actuator with High Positioning and Thermal Stability.
IEEE/ASME Trans. Mechatron. 2015, 20, 2813–2823. [CrossRef]
5. Teo, T.J.; Chen, I.M.; Yang, G. A large deflection and high payload flexure-based parallel manipulator for UV nanoimprint
lithography: Part II. Stiffness modeling and performance evaluation. Precis. Eng. 2014, 38, 872–884. [CrossRef]
6. Pham, M.T.; Teo, T.J.; Yeo, S.H.; Wang, P.; Nai, M.L.S. A 3D-printed Ti-6Al-4V 3-DOF compliant parallel mechanism for high
precision manipulation. IEEE/ASME Trans. Mechatron. 2017, 22, 2359–2368. [CrossRef]
7. Pham, M.T.; Yeo, S.H.; Teo, T.J.; Wang, P.; Nai, M.L.S. Design and Optimization of a Three Degrees-of-Freedom Spatial Motion
Compliant Parallel Mechanism with Fully Decoupled Motion Characteristics. J. Mech. Robot. 2019, 11, 051010. [CrossRef]
8. Teo, T.J.; Yang, G.; Chen, I.M. Compliant Manipulators. In Handbook of Manufacturing Engineering and Technology; Springer:
London, UK, 2014; pp. 2229–2300.
9. Thomas, T.L.; Kalpathy Venkiteswaran, V.; Ananthasuresh, G.K.; Misra, S. Surgical Applications of Compliant Mechanisms: A
Review. J. Mech. Robot. 2021, 13, 020801. [CrossRef]
10. Huang, H.; Pan, Y.; Pang, Y.; Shen, H.; Gao, X.; Zhu, Y.; Chen, L.; Sun, L. Piezoelectric Ultrasonic Biological Microdissection
Device Based on a Novel Flexure Mechanism for Suppressing Vibration. Micromachines 2021, 12, 196. [CrossRef]
11. Shinde, S.M.; Lekurwale, R.R. Synthesising of flexural spindle head micro drilling machine tool in PLM environment. Int. J.
Virtual Technol. Multimed. 2021, 1, 246–264. [CrossRef]
12. Shinde, S.M.; Lekurwale, R.R. Radial stiffness computation of single Archimedes spiral plane supporting spring loaded in flexural
mechanism mounted in spindle head of micro drilling machine tool. Mech. Based Des. Struct. Mach. 2022, 1–21. [CrossRef]
13. Lv, B.; Lin, B.; Cao, Z.; Li, B.; Wang, G. A parallel 3-DOF micro-nano motion stage for vibration-assisted milling. Mech. Mach.
Theory 2022, 173, 104854. [CrossRef]
14. Gandhi, P.; Bhole, K. 3D Microfabrication Using Bulk Lithography. In Proceedings of the ASME 2011 International Mechanical
Engineering Congress and Exposition, Denver, CO, USA, 11–17 November 2011; pp. 393–399.
15. Gandhi, P.; Deshmukh, S.; Ramtekkar, R.; Bhole, K.; Baraki, A. “On-Axis” Linear Focused Spot Scanning Microstereolithography
System: Optomechatronic Design, Analysis and Development. J. Adv. Manuf. Syst. 2013, 12, 43–68. [CrossRef]
16. Tanikawa, T.; Arai, T.; Koyachi, N. Development of small-sized 3 DOF finger module in micro hand for micro manipulation. In
Proceedings of the 1999 IEEE/RSJ International Conference on Intelligent Robots and Systems, Kyongju, Korea, 17–21 October
1999; pp. 876–881.
17. Tanikawa, T.; Ukiana, M.; Morita, K.; Koseki, Y.; Ohba, K.; Fujii, K.; Arai, T. Design of 3-DOF parallel mechanism with thin plate
for micro finger module in micro manipulation. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots
and Systems, Lausanne, Switzerland, 30 September–4 October 2002; pp. 1778–1783.
18. Yi, B.-J.; Chung, G.B.; Na, H.Y.; Kim, W.K.; Suh, I.H. Design and experiment of a 3-DOF parallel micromechanism utilizing flexure
hinges. IEEE Trans. Robot. Autom. 2003, 19, 604–612. [CrossRef]

90
Mathematics 2022, 10, 1414

19. Pham, H.-H.; Chen, I.-M.; Yeh, H.-C. Micro-motion selective-actuation XYZ flexure parallel mechanism: Design and modeling. J.
Micromechatronics 2003, 3, 51–73. [CrossRef]
20. Lu, T.-F.; Handley, D.C.; Yong, Y.K.; Eales, C. A three-DOF compliant micromotion stage with flexure hinges. Ind. Robot. Int. J.
2004, 31, 355–361. [CrossRef]
21. Chao, D.; Zong, G.; Liu, R. Design of a 6-DOF compliant manipulator based on serial-parallel architecture. In Proceedings of the
2005 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, Monterey, CA, USA, 24–28 July 2005; pp.
765–770.
22. Pham, H.-H.; Chen, I.M. Stiffness modeling of flexure parallel mechanism. Precis. Eng. 2005, 29, 467–478. [CrossRef]
23. Choi, Y.-J.; Sreenivasan, S.V.; Choi, B.J. Kinematic design of large displacement precision XY positioning stage by using cross strip
flexure joints and over-constrained mechanism. Mech. Mach. Theory 2008, 43, 724–737. [CrossRef]
24. Yao, Q.; Dong, J.; Ferreira, P.M. A novel parallel-kinematics mechanisms for integrated, multi-axis nanopositioning: Part 1.
Kinematics and design for fabrication. Precis. Eng. 2008, 32, 7–19. [CrossRef]
25. Dong, J.; Yao, Q.; Ferreira, P.M. A novel parallel-kinematics mechanism for integrated, multi-axis nanopositioning: Part 2:
Dynamics, control and performance analysis. Precis. Eng. 2008, 32, 20–33. [CrossRef]
26. Dong, W.; Sun, L.; Du, Z. Stiffness research on a high-precision, large-workspace parallel mechanism with compliant joints. Precis.
Eng. 2008, 32, 222–231. [CrossRef]
27. Wang, H.; Zhang, X. Input coupling analysis and optimal design of a 3-DOF compliant micro-positioning stage. Mech. Mach.
Theory 2008, 43, 400–410. [CrossRef]
28. Wu, T.L.; Chen, J.H.; Chang, S.H. A six-DOF prismatic-spherical-spherical parallel compliant nanopositioner. IEEE Trans. Ultrason.
Ferroelectr. Freq. Control. 2008, 55, 2544–2551. [CrossRef] [PubMed]
29. Yong, Y.K.; Lu, T.-F. Kinetostatic modeling of 3-RRR compliant micro-motion stages with flexure hinges. Mech. Mach. Theory 2009,
44, 1156–1175. [CrossRef]
30. Li, Y.; Xu, Q. A Totally Decoupled Piezo-Driven XYZ Flexure Parallel Micropositioning Stage for Micro/Nanomanipulation. IEEE
Trans. Autom. Sci. Eng. 2011, 8, 265–279. [CrossRef]
31. Liang, Q.; Zhang, D.; Chi, Z.; Song, Q.; Ge, Y.; Ge, Y. Six-DOF micro-manipulator based on compliant parallel mechanism with
integrated force sensor. Robot. Comput.-Integr. Manuf. 2011, 27, 124–134. [CrossRef]
32. Yang, G.; Teo, T.J.; Chen, I.M.; Lin, W. Analysis and design of a 3-DOF flexure-based zero-torsion parallel manipulator for
nano-alignment applications. In Proceedings of the 2011 IEEE International Conference on Robotics and Automation (ICRA),
Shanghai, China, 9–13 May 2011; pp. 2751–2756.
33. Yun, Y.; Li, Y. Optimal design of a 3-PUPU parallel robot with compliant hinges for micromanipulation in a cubic workspace.
Robot. Comput.-Integr. Manuf. 2011, 27, 977–985. [CrossRef]
34. Li, Y.; Huang, J.; Tang, H. A Compliant Parallel XY Micromotion Stage with Complete Kinematic Decoupling. IEEE Trans. Autom.
Sci. Eng. 2012, 9, 538–553. [CrossRef]
35. Kim, H.-Y.; Ahn, D.-H.; Gweon, D.-G. Development of a novel 3-degrees of freedom flexure based positioning system. Rev. Sci.
Instrum. 2012, 83, 055114. [CrossRef]
36. Xiao, S.; Li, Y. Optimal Design, Fabrication, and Control of an XY Micropositioning Stage Driven by Electromagnetic Actuators.
IEEE Trans. Ind. Electron. 2013, 60, 4613–4626. [CrossRef]
37. Bhagat, U.; Shirinzadeh, B.; Clark, L.; Chea, P.; Qin, Y.; Tian, Y.; Zhang, D. Design and analysis of a novel flexure-based 3-DOF
mechanism. Mech. Mach. Theory 2014, 74, 173–187. [CrossRef]
38. Clark, L.; Shirinzadeh, B.; Zhong, Y.; Tian, Y.; Zhang, D. Design and analysis of a compact flexure-based precision pure rotation
stage without actuator redundancy. Mech. Mach. Theory 2016, 105, 129–144. [CrossRef]
39. Hao, G.; Yu, J. Design, modelling and analysis of a completely-decoupled XY compliant parallel manipulator. Mech. Mach. Theory
2016, 102, 179–195. [CrossRef]
40. Li, Y.; Wu, Z. Design, analysis and simulation of a novel 3-DOF translational micromanipulator based on the PRB model. Mech.
Mach. Theory 2016, 100, 235–258. [CrossRef]
41. Wang, R.; Zhang, X. A planar 3-DOF nanopositioning platform with large magnification. Precis. Eng. 2016, 46, 221–231. [CrossRef]
42. Culpepper, M.L.; Anderson, G. Design of a low-cost nano-manipulator which utilizes a monolithic, spatial compliant mechanism.
Precis. Eng. 2004, 28, 469–482. [CrossRef]
43. Chen, S.-C.; Culpepper, M.L. Design of a six-axis micro-scale nanopositioner—μHexFlex. Precis. Eng. 2006, 30, 314–324. [CrossRef]
44. Awtar, S.; Slocum, A.H. Constraint-Based Design of Parallel Kinematic XY Flexure Mechanisms. J. Mech. Des. 2006, 129, 816–830.
[CrossRef]
45. Hopkins, J.B.; Culpepper, M.L. Synthesis of multi-degree of freedom, parallel flexure system concepts via Freedom and Constraint
Topology (FACT)—Part I: Principles. Precis. Eng. 2010, 34, 259–270. [CrossRef]
46. Hopkins, J.B.; Culpepper, M.L. Synthesis of multi-degree of freedom, parallel flexure system concepts via freedom and constraint
topology (FACT). Part II: Practice. Precis. Eng. 2010, 34, 271–278. [CrossRef]
47. Hopkins, J.B.; Culpepper, M.L. Synthesis of precision serial flexure systems using freedom and constraint topologies (FACT).
Precis. Eng. 2011, 35, 638–649. [CrossRef]
48. Awtar, S.; Ustick, J.; Sen, S. An XYZ Parallel-Kinematic Flexure Mechanism with Geometrically Decoupled Degrees of Freedom. J.
Mech. Robot. 2012, 5, 015001. [CrossRef]

91
Mathematics 2022, 10, 1414

49. Hopkins, J.B.; Panas, R.M. Design of flexure-based precision transmission mechanisms using screw theory. Precis. Eng. 2013, 37,
299–307. [CrossRef]
50. Li, H.; Hao, G. A constraint and position identification (CPI) approach for the synthesis of decoupled spatial translational
compliant parallel manipulators. Mech. Mach. Theory 2015, 90, 59–83. [CrossRef]
51. Li, H.; Hao, G.; Kavanagh, R. A New XYZ Compliant Parallel Mechanism for Micro-/Nano-Manipulation: Design and Analysis.
Micromachines 2016, 7, 23. [CrossRef]
52. Hao, G. Design and analysis of symmetric and compact 2R1T (in-plane 3-DOC) flexure parallel mechanisms. Mech. Sci. 2017, 8,
1–9. [CrossRef]
53. Lum, G.Z.; Teo, T.J.; Yeo, S.H.; Yang, G.; Sitti, M. Structural optimization for flexure-based parallel mechanisms—Towards
achieving optimal dynamic and stiffness properties. Precis. Eng. 2015, 42, 195–207. [CrossRef]
54. Jin, M.; Zhang, X. A new topology optimization method for planar compliant parallel mechanisms. Mech. Mach. Theory 2016, 95,
42–58. [CrossRef]
55. Lum, G.Z.; Pham, M.T.; Teo, T.J.; Yang, G.; Yeo, S.H.; Sitti, M. An XY & thetaZ flexure mechanism with optimal stiffness properties.
In Proceedings of the 2017 IEEE International Conference on Advanced Intelligent Mechatronics (AIM), Munich, Germany, 3–7
July 2017; pp. 1103–1110.
56. Pham, M.T.; Yeo, S.H.; Teo, T.J.; Wang, P.; Nai, M.L.S. A Decoupled 6-DOF Compliant Parallel Mechanism with Optimized
Dynamic Characteristics Using Cellular Structure. Machines 2021, 9, 5. [CrossRef]
57. Lobontiu, N. Compliant Mechanisms: Design of Flexure Hinges; CRC Press: Boca Raton, FL, USA, 2010.
58. Akbari, S.; Pirbodaghi, T. Precision positioning using a novel six axes compliant nano-manipulator. Microsyst. Technol. 2017, 23,
2499–2507. [CrossRef]
59. de Jong, B.R.; Brouwer, D.M.; de Boer, M.J.; Jansen, H.V.; Soemers, H.M.; Krijnen, G.J. Design and Fabrication of a Planar
Three-DOFs MEMS-Based Manipulator. J. Microelectromech. Syst. 2010, 19, 1116–1130. [CrossRef]
60. Mukhopadhyay, D.; Dong, J.; Pengwang, E.; Ferreira, P. A SOI-MEMS-based 3-DOF planar parallel-kinematics nanopositioning
stage. Sens. Actuators A Phys. 2008, 147, 340–351. [CrossRef]
61. Lee, K.M.; Arjunan, S. A three-degrees-of-freedom micromotion in-parallel actuated manipulator. IEEE Trans. Robot. Autom. 1991,
7, 634–641. [CrossRef]
62. Teo, T.J. Flexure-Based Electromagnetic Parallel-Kinematics Manipulator System. Ph.D. Thesis, Nanyang Technological University,
Singapore, 2009.
63. Pham, M.T. Design and 3D Printing of Compliant Mechanisms. Ph.D. Thesis, Nanyang Technological University, Singapore, 2019.

92
 mathematics

Article
Predictive Suspension Algorithm for Land Vehicles over
Deterministic Topography
Alejandro Bustos *, Jesus Meneses, Higinio Rubio and Enrique Soriano-Heras

Department of Mechanical Engineering, Universidad Carlos III de Madrid, 28911 Leganes, Spain;
meneses@ing.uc3m.es (J.M.); hrubio@ing.uc3m.es (H.R.); esoriano@ing.uc3m.es (E.S.-H.)
* Correspondence: albustos@ing.uc3m.es

Abstract: A good suspension system is mandatory for ensuring stability, comfort and safety in land
vehicles; therefore, advanced semi and fully active suspension systems have been developed along
with their associated management strategies to overcome the limitations of passive suspensions.
This paper presents a suspension algorithm for land vehicles traveling through a deterministic
topography. The kinematics of a half-vehicle model and the algorithm are implemented in Simulink.
The algorithm’s inputs are the measurements provided by a position scanner located on the front
wheel of the vehicle. Based on this input, the algorithm reconstructs the topography in real-time and
sends the corresponding command to an actuator located on the rear wheel to compensate for the
irregularities of the terrain. The actuation is governed by the parameter “D”, which represents the
distance over which the algorithm averages the height of the terrain. Two ground profiles were tested
and sensitivity analysis of the parameter “D” was performed. Results show that larger values of “D”
usually yield less vibration on the actuated mass, but this value also depends on the irregularities of
the terrain.

Keywords: suspension algorithm; predictive suspension; land vehicle; sprung mass

Citation: Bustos, A.; Meneses, J.; MSC: 70B15

Rubio, H.; Soriano-Heras, E.
Predictive Suspension Algorithm for
Land Vehicles over Deterministic
Topography. Mathematics 2022, 10, 1. Introduction
1467. https://doi.org/10.3390/
Suspension is an essential system for land vehicles to ensure their stability, comfort and
math10091467
safety; however, increasing the vehicle stability usually results in a worsening of comfort
Academic Editor: Andrey Jivkov and vice versa [1,2]. In an attempt to improve both stability and comfort, semi-active and
Received: 28 March 2022
active suspensions and their corresponding management strategies have been developed
Accepted: 25 April 2022
in recent years.
Published: 27 April 2022 Suspension systems can be classified into three types according to their operation:
passive, semi-active and active.
Publisher’s Note: MDPI stays neutral
The fundamental elements of passive suspensions are springs, dampers and the
with regard to jurisdictional claims in
required rigid links to connect the sprung and unsprung masses. In this kind of suspension,
published maps and institutional affil-
the parameters of springs and dampers are fixed; therefore, the dynamic response of the
iations.
system is fixed too. The design of these suspensions must achieve a trade-off solution for
different conditions due to the different surfaces the vehicle travels through; however, non-
linear springs [3,4], non-linear dampers [5] and hydro–gas rotary suspension systems [6]
Copyright: © 2022 by the authors.
have been used in land vehicles to obtain better response over the widest possible range
Licensee MDPI, Basel, Switzerland. of conditions.
This article is an open access article Semi-active suspensions can adjust the parameters of passive suspensions, but the
distributed under the terms and generated forces will always be dependent on the speed of the damper [7]. Hence, semi-
conditions of the Creative Commons active suspensions are not able to control ride height, roll or pitch angles. Nonetheless,
Attribution (CC BY) license (https:// their simplicity requires less power consumption than fully active suspensions and can be
creativecommons.org/licenses/by/ implemented easily:
4.0/).

Mathematics 2022, 10, 1467. https://doi.org/10.3390/math10091467 93 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 1467

• The stiffness of semi-active suspensions is modified by using magnetorheological

dampers within a specific mechanism [8] or by governing smart air springs [9].
• Variable damping is the most common strategy in the development of semi-active
suspensions, having different solutions:
 Hydraulic dampers with electrically controlled valves for adjusting the fluid flow
rate from one chamber to the other [10].
 Magnetorheological [11] and electrorheological [12,13] dampers are similar to
hydraulic dampers but they contain a magnetorheological or electrorheological
fluid. This fluid, usually non-Newtonian, reacts to the application of magnetic
or electrical fields and changes the properties of the damper. The use of these
devices is widely spread in road and rail vehicles [14–16].
 Electromagnetic dampers take advantage of the known interaction between
a moving coil and the magnetic field generated by a permanent magnet or
electromagnet to generate a damping effect [17]. Karnopp studies the use of
permanent magnet linear motors as variable mechanical dampers for vehicle
suspensions [18].
• The third component of suspension systems, inertia, is modified through a variable
inerter. The inerter is a mechanical device in which the force applied is proportional to
the relative accelerations between the two terminals of the device [19]. These systems
have been proposed to improve the comfort of rail vehicles [20].
Active suspensions allow full management of the suspension motion, as the actuator
force is independent of the speed direction. On the other hand, the main disadvantage of
these suspension types is the high energy consumption that they require. The devices used
in these suspension systems are, typically:
• Electromechanical actuators can use an electric motor attached to a mechanism that
converts rotary motion into linear motion [21] or can use linear motors [22].
• Hydraulic actuators require the installation of pumps, valves and all the devices
needed for the proper operation of the system. These actuators have been applied to
improve the quality of the ride on both road [23] and railway vehicles [24].
Regarding railway vehicles, their own nature allows the development of tilting tech-
nology, which rolls the car body in curves with the aim of reducing the lateral acceleration
perceived by the passengers. Usually, the movement of the car body is achieved by using
hydraulic or electromechanical actuators [25]; however, Talgo introduced in the 1980s a
passive natural tilting system in its trains [26].
The management of active and semi-active suspension systems is performed through
different strategies. Several concepts are found in the literature: classic control [27,28],
Model Predictive Control [29], fuzzy logic [30], skyhook control [31] and H2 and H∞
control [32]. All of them have proved an ability to achieve a reasonable trade-off between
stability and comfort with the appropriate tuning conditions.
Another approach for managing the suspension system of light vehicles is proposed.
This new approach is based on mechanical principles and relies on the measurements
acquired by a position scanner installed at the front end of the vehicle to actuate on the
following suspension systems; therefore, the unevenness of the terrain can be counteracted.
This paper presents a predictive suspension algorithm for land vehicles traveling through
a deterministic topography. The kinematics of a half vehicle model is posed and then
implemented in Simulink, including the suspension algorithm. Our algorithm “reads” the
surface the vehicle is traveling through thanks to a sensor installed in the front wheel of the
vehicle and acts over a device installed in the rear wheel to compensate for the irregularities
of the terrain. The actuation is governed by the parameter “D”, which represents the
distance over which the algorithm averages the height of the terrain. Two ground profiles
were tested and sensitivity analysis of the parameter “D” was performed.
The structure of the paper is as follows: Section 2 describes the equations that define
the vehicle model and the suspension algorithm. Section 3 shows the results obtained after

94
Mathematics 2022, 10, 1467

simulating the vehicle and the response of the algorithm in two types of terrain. The fourth
and last section presents the conclusions of this work.

2. Materials and Methods

The predictive suspension algorithm is implemented on a two-axle vehicle, whose
kinematics must be solved first. For simplicity, only 1/2 vehicle is modeled and symmetry
action on both wheels of the same axle is assumed, so the problem becomes bi-dimensional.
The vehicle will travel through a deterministic topography. Hence, the vertical position of
any wheel can be established from the horizontal position by the function that relates both
variables.
The vehicle model also assumes that the mechanical energy remains constant, only
wheels have mass and pure rolling (no sliding) is considered as it is a condition for the good
performance of the suspension algorithm. This model and the algorithm were implemented
in Simulink.

2.1. Vehicle Model

The mechanical energy of the vehicle is given by Equation (1)

1 1 1 1
E= m v2 + I ω 2 + m1 · g · y1 + m2 v22 + I2 ω22 + m2 · g · y2 (1)
2 1 1 2 1 1 2 2
where E is the mechanical energy, m1 and m2 are the masses of both wheels, I1 and I2 are
the inertias of both wheels, v1 and v2 are the linear velocities of the wheels, ω 1 and ω 2 are
the angular speeds of the wheels, y1 and y2 are the vertical positions of the wheels and g is
the gravity.
By assuming pure rolling motion and the wheels as solid disks, Equation (1) can be
simplified; therefore, the initial energy of the system is obtained from Equation (2), where
the subscript 0 means “initial conditions”.

1 1 3
E0 = 2·m·v20 + 2·m·v20 + 2·m· g · y0 = ·m·v20 + 2·m· g · y0 (2)
2 4 2
The distance between the front and rear wheels (that is, the wheelbase) L is known
and can be described mathematically as a circumference centered on the front wheel and
radius L. The position of the rear wheel is obtained by computing the intersection of this
circumference with the trajectory described by the ground function y = f(x). Hence, the
solution of the equation system given by Equation (3) yields the position of the rear wheel.
(
( x2 − x1 )2 + ( y2 − y1 )2 = L2
(3)
y = f (x)

This system of equations is solved by iteration, as the ground function can be any type
of function, even not analytical. The flowchart describing the iteration process is shown
in Figure 1.
The velocity of both wheels is computed from the conservation of energy, the time
derivative of the constraint equation and the ground function, obtaining the equation
system built up by Equations (4)–(6).

3 3
E0 − (m1 · g · y1 + m2 · g · y2 ) = m v2 + m2 v22 (4)
4 1 1 4
$. . % $. . %
2 x 2 − x 1 ( x2 − x1 ) + 2 y2 − y1 ( y2 − y1 ) = 0 (5)
. d f (x)
y= (6)
dt

95
Mathematics 2022, 10, 1467

ȱ
Figure 1. Flowchart for computing the position of the rear wheel.

In order to solve this system more efficiently, polar coordinates are used. This implies
defining the pitch angle θ as Equation (7).
 
y − y2
θ = π + arctg 1 (7)
x1 − x2

Therefore, the constraint Equation (3) is rewritten as Equation (8).

(
x2 = x1 + L cos θ
(8)
y2 = y1 + Lsinθ

The combination of the conservation of energy with the time derivatives of the ground
function and the constraint equation in polar coordinates yields Equations (9)–(12), which
allows computing all the components of the speed of both wheels.
As previously, this system of equations is solved by iteration following the flowchart
in Figure 2.

3 3 3 .2 .2 3 .2 .2
E0 − (m1 · g · y1 + m2 · g · y2 ) = m v2 + m2 v22 = m1 ( x1 + y1 ) + m2 ( x2 + y2 ) (9)
4 1 1 4 4 4
. d f ( x1 )
y1 = (10)
dt
. . .
x2 = x1 − L·θ ·senθ (11)
. . .
y2 = y1 + L·θ · cos θ (12)

96
Mathematics 2022, 10, 1467

Figure 2. Flowchart for computing the angular speeds.

Two different ground topographies and several values of the parameter D were
proposed for testing the algorithm. The first topography (profile 1) simulates a wiggly
surface and is defined by Equation (13)—where A1 is the amplitude of the wave and λ1 is
the wavelength, both in meters.
 
2π
y = A1 sin x (13)
λ1

The second profile is a combination of two sinusoidal irregularities according to

Equation (14):    
2π 2π
y = A0 + A1 sen ( x − a) + A2 sen x (14)
λ1 λ2
A1 and A2 are the amplitudes of both sines, λ1 and λ2 are the wavelengths, A0 is
the reference height and a is a parameter to guarantee continuity in the first derivative;
therefore, A0 and a are such that:
   
1 1 A2 λ1
a = λ1 1 − arccos (15)
2 π A1 λ2
 
2πa
A0 = A1 sen (16)
λ1

97
Mathematics 2022, 10, 1467

The model parameters and initial conditions used in the simulations of both topogra-
phy profiles are summarized in Table 1. The input values of both topography profiles are
depicted in Table 2, as well as the computed values of variables A0 and a.

Table 1. Model parameters.

Parameter Value
Simulation time 10 s
Horizontal velocity at t = 0 3 m/s
Vertical velocity at t = 0 0 m/s
Wheel diameter 1m
Wheel mass 60 kg
Wheelbase 3m

Table 2. Ground function parameters.

Profile 1 Profile 2
Parameter Value Parameter Value
A1 0.05 m A1 0.05 m
λ1 4m λ1 2m
A2 −0.25 m
λ2 10 m
A0 0m
a 0m

2.2. Suspension Algorithm

The suspension algorithm is applied to the 1/2 vehicle model described in the previous
subsection. The frontwheel acts as a probe wheel, whereas the rear wheel is the actuated
wheel. The probe wheel is instrumented with a vertical accelerometer and both wheels
should be equipped with an accurate system for measuring their angular positions.
In Figure 3, two kinds of terrain defects (peak and corrugated) are sketched, as well
as the trajectories described by the wheel centers and the trajectory that the sprung mass
should follow (a straight line).
.. ..
Let Δt be the sampling time interval of the accelerometer and yn = y(nΔt) the vertical
acceleration registered at time t = nΔt. Assuming a constant vertical jerk along the interval
[(n-1)Δt, nΔt] yields Equation (17).
.. ..
... y − y n −1
yn = n = cte (17)
Δt
Then the expressions for the vertical velocity and position of the wheel center are,
respectively:

. . .. 1 ... . 1 $ .. .. %
yn = yn−1 + yn−1 Δt + y n Δt2 = yn−1 + yn−1 + yn Δt (18)
2 2
. 1 .. 1 ... . 1 $ .. .. %
yn = yn−1 + yn−1 Δt + yn−1 Δt2 + y n Δt3 = yn−1 + yn−1 Δt + 2yn−1 + yn Δt2 (19)
2 6 6
Although the corresponding velocity and position errors grow with time, nΔt, and
time squared, (nΔt)2 , respectively, (easy to show); this is not too much of a problem since
the algorithm does not include the absolute height, but its variation over a certain distance,
“D”. In other words, the database of values obtained in Equations (18) and (19), will be
dynamically refreshed, so that n, and therefore the temporal interval nΔt, does not increase
indefinitely.

98
Mathematics 2022, 10, 1467

Figure 3. Two kinds of terrain defects: (a) peak defects and (b) smooth or corrugated terrain.

The vertical actuator (between the actuated or rear wheel and the sprung mass) has
to maintain the latter at the objective height. In this approach, the objective height is
the result of the moving average of the calculated height over the rolling distance 2D.
Actually, the larger this distance, the softer the objective function, but also the bigger error
is accumulated in calculating the height from the accelerometer reading. Thus, parameter
“D” should be adjusted according to the kind of roughness the topography has (with the
sole requirement not to exceed the distance between wheels).
For the proposed algorithm, it is important to know exactly the position of both wheels
(which are supposed to roll without slipping) on the ground. This is achieved by recording
pro
the corresponding angles, ϕn , ϕnact (angle turned by the probe wheel, and actuated wheel,
respectively), having set both to zero at the initial position (n = 0). The model is sketched
in Figure 4 at the initial position and at the time t = nΔt. The actual trajectories followed
by the center of the wheels and the trajectory the end of the actuator should follow are
also represented.

Figure 4. Model of the vehicle sketched at the initial position (t = 0) and at the time t = nΔt, where the
angles are indicated. Dotted and dashed curves represent respectively the actual trajectory followed
by the center of the wheels and that the actuated mass should follow.

99
Mathematics 2022, 10, 1467

..
At time t = nΔt, the vertical acceleration of the probe wheel,yn , and the angles turned
.
by the wheels, are measured. Then, the velocity, yn , and vertical position, yn , are calculated
by using Equations (18) and (19), respectively. All these values are recorded in a database
in the way indicated in Table 3.

Table 3. Database of recorded and computed values.

Probe Wheel Actuated Wheel Vertical Vertical Vertical

# Record Time
Angle Turned Angle Turned Acceleration Velocity Position
pro .. .
1 ϕ1 ϕ1act y1 y1 y1 t1
.. .. .. .. .. .. ..
. . . . . . .
pro .. .
n ϕn ϕnact yn yn yn tn

Once enough records have been acquired, the algorithm is ready to provide the current
length to be taken by the actuator. The objective height is calculated as the moving average
centered on the record corresponding to the actuated wheel and over a sufficient number
of records; speaking in terms of rolling distances, the moving average is calculated over a
rolling distance of 2D. This is achieved by identifying the records m, i and f that, respectively,
define the current position for the actuated wheel and the first and last of the records to
average over. In other words, at t = mΔt the probe wheel was just on the same point of the
trajectory that the actuated wheel is at t = nΔt (i.e., at present). In Figure 5, the model at
moments corresponding to records m, n, i and f is sketched.

ȱ
Figure 5. Sketch of the model at the key moments for the algorithm: at the initial position (dashed
line in (up) and (middle) diagrams); when the record m is being registered (up); when the record n
is being registered (middle); when the record i is being registered (dotted line, (down)); when the
record f is being registered (dash line, (down)).

100
Mathematics 2022, 10, 1467

On seeing Figure 5, if we consider both wheels have the same radius, r, the following
rules for identifying the numbers m, i and f are deduced:
m is such that:
pro L
ϕm ≈ ϕnact − (20)
r
i is such that:
pro pro D L+D
ϕi ≈ ϕ m − ≈ ϕnact − (21)
r r
f is such that:
pro pro D L+D
ϕ f ≈ ϕm + ≈ ϕnact + (22)
r r
& '
pro
With “≈” we mean “is the nearest value to” (note that ϕ j is a discrete set of values,
so in general, none of the last equalities could be exactly satisfied). We have also considered
that the first section of the trajectory is flat and horizontal, with a minimum distance that is
equal to the wheelbase, L.
Once the records i, m and f have been identified, the algorithm calculates the objective
height (see Figure 6) as:
f
1
f − i + 1∑
yobj = yj + h (23)
j =i

and the actuated distance results:

f
1
f − i + 1∑
y act = yobj − ym = y j + h − ym (24)
j =i

where h is the initial objective height (i.e., the actuator initial length, see Figures 4 and 6).

Figure 6. Heights that are involved in the calculus of the actuated distance.

Finally, the database is refreshed by deleting the records with subscripts less than i,
and decreasing the remainder subscripts in i units:

if j < i:, delete record

if j ≥ i, j = j − i
In this way, the database does not grow indefinitely, and only the records to be used
by the algorithm at the current time of calculation remain.
In summary, the algorithm can be considered as consisting of three tasks:
1. Reading (angles and vertical acceleration) and computing vertical velocity and position.
2. Identifying the indexes i, m and f.
3. Obtaining the actuated distance and refreshing the database.

101
Mathematics 2022, 10, 1467

To understand how the algorithm performs the second task, suppose the database
at time nΔt is as shown in Table 4. For a given angle turned by the actuated wheel ϕnact ,
the algorithm computes the theoretical angle turned by the probe wheel at the three key
moments iΔt, mΔt and f Δt (ϕnact − L+r D , ϕnact − Lr and ϕnact + L+r D , respectively), using
Equations (20)–(22).Then, it searches the closest actual angles recorded in the database to
those computed and determines the indexes i, m and f. As the probe wheel angle turned
does not decrease with n (hence i ≤ m ≤ f ), the search can be performed sequentially.

Table 4. Time “nΔt” database.

# Record Angle Turned by Probe Wheel Angle Turned by Actuated Wheel

1 M 
SUR
MDFW

i MLSUR MLDFW

m MPSUR MPDFW

f M ISUR M DFW
I

n MQSUR MQDFW

The whole algorithm flowchart is represented in Figure 7.

2.3. Implementation in Simulink

The models described above to define the kinematics of the 1/2 vehicle and the
suspension algorithm are implemented in Simulink, and more specifically, in Simscape.
The main diagram is shown in Figure 8 and is composed of six big blocks. The input block
is the “Ground function” block, which defines the parameters of the trajectory that the
vehicle must follow. The “Energy” block performs all the calculations (Equations (1)–(13))
to compute the kinematics of the half vehicle model. Front and rear wheel blocks contain
the relationships between the wheel bodies and the main reference frame of Simscape. The
rear wheel block also has the actuator to move the sprung mass according to the commands
given by the suspension algorithm. The computation of the rotation angles turned by both
wheels is carried out within the rotation block. Finally, the block “Suspension algorithm”
performs all the calculations (Equations (14)–(21)) to actuate the sprung mass.
The ”Energy” block is composed of another block called “Calculations”, which is the
block that actually computes the kinematics of the model. The inputs of this block are the
ground profile, the wheel properties and the position of the front wheel. The outputs of this
block are the kinematics of both wheels. Within this block, an S-function is responsible for
executing the algorithms that compute the kinematics. To perform this action, this function
is feedbacked with the positions and velocities computed in the previous integration step.

102
Mathematics 2022, 10, 1467

Figure 7. Suspension algorithm flowchart.

103
Mathematics 2022, 10, 1467

Figure 8. Diagram of the Simulink implementation of the whole model.

3. Results
This section presents the results obtained after testing the predictive suspension
algorithm over the two ground topographies defined in Section 2.1. The first topography
(profile 1) simulates a wiggly surface; the second topography is a combination of two
sinusoidal irregularities. The parameters of both topographies are summarized in Table 2.
As is described in Section 2.2, the algorithm’s key parameter to define the motion of
the sprung or actuated mass is the distance “D”. This parameter is set from D = 0 m (no
actuation over the sprung mass) to D = 3 m (the vehicle’s wheelbase) in steps of 0.5 m.
Hence, seven values of “D” were tested.
Additionally, the parameter “D” is normalized to the wheelbase of the vehicle by
defining a new variable d = D/L, where L is the wheelbase. As the value of the “D” is
limited by the wheelbase (it is not possible to obtain information on positions not yet
traveled by the front wheel), the value of “d” will always be between 0 and 1.

3.1. Profile Type 1

The results of the simulation of the first ground profile are shown in Figures 9–12.
Figure 9 illustrates the kinematics of both wheels of the 1/2 vehicle. The trajectory of
both wheels follows the input ground profile as expected. The velocity in both axes varies
according to the topography of the ground profile and the energy conservation. Similar
behavior is noticed in the accelerations. The significant acceleration peak at the horizontal
position equal to 0 is due to the sudden transition from flat terrain to an irregular surface.

104
Mathematics 2022, 10, 1467

Figure 9. Kinematics of the vehicle when traveling through the ground profile 1.

Figure 10. The motion of the actuator in the first ground profile.

105
Mathematics 2022, 10, 1467

Figure 11. Displacement of the actuated mass in the first ground profile.

Figure 12. Acceleration spectra of the actuated mass in the first ground profile tested.

The displacement of the actuator placed between the rear wheel and the sprung
mass is plotted in Figure 10. As it can be seen, the actuator does not move when the
value of the normalized parameter “d” is 0. When increasing the value of this parameter,
the actuator moves to compensate for the irregularities of the ground, which even leads
to a wave out phased 180º to the ground profile. Lower values of parameter “d ”yield
smaller displacements of the actuator, though it also generates some discontinuities. Larger
values of parameter “d” depict better an out phased trajectory from the terrain, but the
amplitude of the movement is larger too. The actuator always works in extension (positive

106
Mathematics 2022, 10, 1467

displacement) at the beginning of these simulations due to the initial elevation of the
front wheel.
Figure 11 shows the path followed by the center of the rear wheel and the position of
the actuated mass for a better comparison of the followed trajectory and the actuation of
the suspension system. It is observed that the movement of the sprung mass is the opposite
of the rear wheel when the parameter “d” is larger than 2/3, that is, when the distance “D”
over which the moving average is computed is larger than 2 m. In this situation, when the
rear wheel moves upwards, the sprung mass is moved downwards, and vice versa. For
values of parameter “D” below 2 m (d < 2/3), the sprung mass moves vertically in the same
direction as the rear wheel, but this movement is limited.
In order to establish what is the best value of the normalized parameter “d” among
all the tested and for ground profile 1, the spectra of the vertical accelerations experienced
by the sprung mass are computed. The results obtained are plotted in Figure 12. The
analysis of the graph shows that the largest values of the parameter “d” yield the lowest
accelerations at low frequencies; however, it appears to be an undesirable vibration at high
frequencies; therefore, according to this criterion, the best value will be one that obtains the
best arrangement between low and high-frequency vibrations.
To make easier the search for the optimum value of parameter “d”, Table 5 summarizes
the maximum amplitude and the RMS of the spectra shown in Figure 12. Combining the
spectra and the data in the table, it can be concluded that the best value for parameter “d”
in the conditions of this simulation is d = 2/3, which means the distance D = 2 m.

Table 5. Characterization of the acceleration spectra in the first ground profile tested.

d=0 d = 1/6 d = 1/3 d = 1/2 d = 2/3 d = 5/6 d=1

Maximum 2.8035 2.0397 1.2130 1.0728 1.0361 1.2695 2.1981
RMS 0.6057 0.9422 0.7551 0.6772 0.6728 0.7363 1.1141

3.2. Profile Type 2

The results obtained from the simulation of the second ground profile are shown in
Figures 13–16. Figure 13 displays the kinematics of both wheels of the 1/2 vehicle. As
in the previous case, the trajectory of both wheels follows the input ground profile as
expected. The speed varies according to the topography of the ground profile and the
energy conservation. Velocity variations are slower on the horizontal axis than on the
vertical axis, where the velocity is more sensitive to profile variations. Similar behavior is
noticed in the accelerations, being the vertical accelerations larger than the horizontal ones.
The displacement of the actuator is plotted in Figure 14. As it can be seen, the actuator
does not move when the value of the parameter “d” is zero. Once the simulation begins,
the actuator starts to move in compression (negative displacement) to compensate initial
descent of the vehicle. As it happens in the first set of simulations, the larger the value of
the parameter “d”, the larger the displacement of the actuator. The use of the largest values
results in an actuation that clearly highlights the two harmonics that compose the second
ground profile.
Figure 15 shows the path followed by the center of the rear wheel and the position of
the actuated mass for a better comparison of the followed trajectory and the actuation of
the suspension system. The movement of the mass follows the movement of the rear wheel,
but the larger the value of “d”, the smoother the displacement of the actuated mass. In fact,
the algorithm moves the mass in advance to the irregularity of the surface the vehicle will
encounter to try to achieve a smoother ride.

107
Mathematics 2022, 10, 1467

Figure 13. Kinematics of the vehicle when traveling through the second ground profile.

Figure 14. The motion of the actuator in the first ground profile.

108
Mathematics 2022, 10, 1467

Figure 15. Displacement of the actuated mass in the second ground profile tested.

ȱ
Figure 16. Acceleration spectra of the actuated mass in the second ground profile tested.

The PSD of the vertical accelerations experienced by the sprung mass through the
second ground profile is computed to establish what is the best value of the parameter “d”
among all the tested values. The results obtained are plotted in Figure 16. As was seen in
the previous subsection, there is a big peak at low frequencies and also some undesirable
accelerations at high frequencies; however, in this set of simulations, a larger value of the
parameter “d” does not mean lower acceleration amplitudes at low frequencies. In fact,
larger acceleration amplitudes are achieved with d = 1 than with d = 5/6, which indicates
that moving the sprung mass in advance to the terrain elevation is not as good as it can
be expected.

109
Mathematics 2022, 10, 1467

Once again, to make easier the search for the optimum value of parameter “d”, the
maximum amplitude and the RMS of the spectra are summarized in Table 6. Combining the
spectra and the numbers in the table, it can be concluded that the best value for parameter
“d” in the conditions of this simulation is d = 5/6, as it achieves the lowest maximum
amplitude and RMS values. The normalized parameter d = 5/6 means that the distance
“D” is equal to 2.5 m.

Table 6. Characterization of the acceleration spectra in the second ground profile tested.

d=0 d = 1/6 d = 1/3 d = 1/2 d = 2/3 d = 5/6 d=1

Maximum 37.9703 22.1259 19.0021 24.2496 18.3643 17.6156 19.4286
RMS 8.4317 5.6769 5.0129 6.7450 5.9922 5.7087 5.7104

4. Conclusions
This paper presents a predictive suspension algorithm for land vehicles traveling
through a deterministic topography. Prior to the virtual implementation of the algorithm,
the kinematics of a two-wheel 2D (1/2) vehicle were proposed and solved. The algorithm
inputs are the measurements given by a position scanner installed in the first wheel of the
vehicle. From these inputs, the algorithm reconstructs the terrain and averages its height
over a distance “D”. Then, it generates the command for moving the actuator installed in
the rear wheel and connected to the sprung mass. Both the algorithm and the kinematic
model of the 1/2 vehicle were implemented in Simulink.
The algorithm was tested on two terrains with different profiles. Furthermore, sen-
sitivity analyses were carried out in order to establish the best value of the parameter
“D” in each profile. The performance of the algorithm is quantified for several values
of the parameter “D”, using the spectral analysis of the accelerations experienced by the
sprung mass.
The results show that, in general, increasing the parameter “D” reduces the motion of
the sprung/actuated mass and, hence, improves comfort; however, there is an optimum
value of the parameter “D” for each ground profile, and this value does not coincide
necessarily with the highest value of “D”. In fact, the optimum “D” value will depend on
the terrain irregularities and their characteristic length. For the first ground profile, the
optimum value of the parameter “D” is 2 m (d = 2/3), which is half the wavelength of the
unevenness of the terrain. The optimum value for the second terrain is D = 2.5 m (d = 5/6).
The value of the “D” parameter is limited by the wheelbase since it is not possible to
obtain information on positions not yet traveled by the front wheel. In fact, the maximum
value of the “D” parameter must be slightly less than the wheelbase, as the data processing
and calculation time of the algorithm must be taken into account. This is not of major sig-
nificance in the current development status of the research as the Simulink implementation
computes the algorithm in real-time, but it should be taken into account when developing
the physical device to install in a light vehicle for real testing.
Finally, it can be concluded that the algorithm works correctly and decreases the
movement of the sprung/actuated mass when the vehicle travels through a deterministic
topography. The next steps in the research would be the development of a physical device
that executes the proposed predictive algorithm and its implementation in a light vehicle.

Author Contributions: Individual contributions are as follows: original concept: J.M.; mathematical
development: J.M. and A.B.; implementation, simulation and draft preparation: A.B.; review, editing,
validation, and formal analysis: E.S.-H., H.R. and J.M. All authors have read and agreed to the
published version of the manuscript.
Funding: The research work described in this paper is part of the R&D and Innovation projects MC4.0
PID2020-116984RB-C21 and MC4.0 PID2020-116984RB-C22 supported by the MCIN/AEI/10.13039/
501100011033.

110
Mathematics 2022, 10, 1467

Institutional Review Board Statement: Not applicable.

Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Reimpell, J.; Stoll, H.; Betzler, J.W. The Automotive Chassis: Engineering Principles, 2nd ed.; Butterworth Heinemann: Oxford, UK,
2001; ISBN 978-0-7506-5054-0.
2. Theunissen, J.; Tota, A.; Gruber, P.; Dhaens, M.; Sorniotti, A. Preview-Based Techniques for Vehicle Suspension Control: A
State-of-the-Art Review. Annu. Rev. Control 2021, 51, 206–235. [CrossRef]
3. Luo, R.; Shi, H.; Guo, J.; Huang, L.; Wang, J. A Nonlinear Rubber Spring Model for the Dynamics Simulation of a High-Speed
Train. Veh. Syst. Dyn. 2020, 58, 1367–1384. [CrossRef]
4. Anubi, O.M.; Patel, D.R.; Crane, C.D., III. A New Variable Stiffness Suspension System: Passive Case. Mech. Sci. 2013, 4, 139–151.
[CrossRef]
5. Kalyan Raj, A.H.; Padmanabhan, C. A New Passive Non-Linear Damper for Automobiles. Proc. Inst. Mech. Eng. Part J.
Automob. Eng. 2009, 223, 1435–1443. [CrossRef]
6. Solomon, U.; Padmanabhan, C. Hydro-Gas Suspension System for a Tracked Vehicle: Modeling and Analysis. J. Terramech. 2011,
48, 125–137. [CrossRef]
7. Fu, B.; Giossi, R.L.; Persson, R.; Stichel, S.; Bruni, S.; Goodall, R. Active Suspension in Railway Vehicles: A Literature Survey.
Railw. Eng. Sci. 2020, 28, 3–35. [CrossRef]
8. Anubi, O.M.; Crane, C.A. New Semiactive Variable Stiffness Suspension System Using Combined Skyhook and Nonlinear Energy
Sink-Based Controllers. IEEE Trans. Control Syst. Technol. 2015, 23, 937–947. [CrossRef]
9. Sun, S.; Deng, H.; Li, W. Variable Stiffness and Damping Suspension System for Train. In Proceedings of the SPIE Volume 9057,
Active and Passive Smart Structures and Integrated Systems 2014, San Diego, CA, USA, 9–13 March 2014; p. 90570P.
10. Faraj, R.; Graczykowski, C.; Holnicki-Szulc, J. Adaptable Pneumatic Shock Absorber. J. Vib. Control 2019, 25, 711–721. [CrossRef]
11. Spencer, B.F.; Dyke, S.J.; Sain, M.K.; Carlson, J.D. Phenomenological Model for Magnetorheological Dampers. J. Eng. Mech. 1997,
123, 230–238. [CrossRef]
12. Gavin, H.P.; Hanson, R.D.; Filisko, F.E. Electrorheological Dampers, Part I: Analysis and Design. J. Appl. Mech. 1996, 63, 669–675.
[CrossRef]
13. Dixon, J. The Shock Absorber Handbook; John Wiley & Sons, Ltd.: Hoboken, NJ, USA, 2008; ISBN 978-0-470-51642-3.
14. Mei, T.X.; Zaeim, A.; Li, H. Control of Railway Wheelsets—A Semi-Active Approach. In Advances in Dynamics of Vehicles on
Roads and Tracks; Klomp, M., Bruzelius, F., Nielsen, J., Hillemyr, A., Eds.; Lecture Notes in Mechanical Engineering; Springer
International Publishing: Cham, Swizerland, 2020; pp. 16–23. ISBN 978-3-030-38076-2.
15. Yang, J.; Ning, D.; Sun, S.S.; Zheng, J.; Lu, H.; Nakano, M.; Zhang, S.; Du, H.; Li, W.H. A Semi-Active Suspension Using a
Magnetorheological Damper with Nonlinear Negative-Stiffness Component. Mech. Syst. Signal Process. 2021, 147, 107071.
[CrossRef]
16. Tudon-Martinez, J.C.; Hernandez-Alcantara, D.; Amezquita-Brooks, L.; Morales-Menendez, R.; Lozoya-Santos, J.D.J.; Aquines, O.
Magneto-Rheological Dampers—Model Influence on the Semi-Active Suspension Performance. Smart Mater. Struct. 2019, 28,
105030. [CrossRef]
17. Soliman, A.; Kaldas, M. Semi-Active Suspension Systems from Research to Mass-Market—A Review. J. Low Freq. Noise Vib. Act.
Control 2021, 40, 1005–1023. [CrossRef]
18. Karnopp, D. Permanent Magnet Linear Motors Used as Variable Mechanical Dampers for Vehicle Suspensions. Veh. Syst. Dyn.
1989, 18, 187–200. [CrossRef]
19. Smith, M.C. Synthesis of Mechanical Networks: The Inerter. IEEE Trans. Autom. Control 2002, 47, 1648–1662. [CrossRef]
20. Lewis, T.D.; Jiang, J.Z.; Neild, S.A.; Gong, C.; Iwnicki, S.D. Using an Inerter-Based Suspension to Improve Both Passenger Comfort
and Track Wear in Railway Vehicles. Veh. Syst. Dyn. 2019, 58, 472–493. [CrossRef]
21. Kawamoto, Y.; Suda, Y.; Inoue, H.; Kondo, T. Electro-Mechanical Suspension System Considering Energy Consumption and
Vehicle Manoeuvre. Veh. Syst. Dyn. 2008, 46, 1053–1063. [CrossRef]
22. Klimenko, Y.I.; Batishchev, D.V.; Pavlenko, A.; Grinchenkov, V.P. Design of a Linear Electromechanical Actuator with an Active
Vehicle Suspension System. Russ. Electr. Eng. 2015, 86, 588–593. [CrossRef]
23. Kilicaslan, S. Control of Active Suspension System Considering Nonlinear Actuator Dynamics. Nonlinear Dyn. 2018, 91, 1383–1394.
[CrossRef]
24. Goodall, R.; Freudenthaler, G.; Dixon, R. Hydraulic Actuation Technology for Full- and Semi-Active Railway Suspensions. Veh.
Syst. Dyn. 2014, 52, 1642–1657. [CrossRef]
25. Goodall, R. Tilting Trains and Beyond-the Future for Active Railway Suspensions. 1. Improving Passenger Comfort. Comput.
Control Eng. J. 1999, 10, 153–160. [CrossRef]

111
Mathematics 2022, 10, 1467

26. Carballeira, J.; Baeza, L.; Rovira, A.; García, E. Technical Characteristics and Dynamic Modelling of Talgo Trains. Veh. Syst. Dyn.
2008, 46, 301–316. [CrossRef]
27. Jamil, I.A.A.; Moghavvemi, M. Optimization of PID Controller Tuning Method Using Evolutionary Algorithms. In Proceedings
of the 2021 Innovations in Power and Advanced Computing Technologies (i-PACT), Kuala Lumpur, Malaysia, 27–29 November
2021; pp. 1–7.
28. Di Gialleonardo, E.; Facchinetti, A.; Bruni, S. Control of an Integrated Lateral and Roll Suspension for a High-Speed Railway
Vehicle. Veh. Syst. Dyn. 2022, 1–27. [CrossRef]
29. Rodriguez-Guevara, D.; Favela-Contreras, A.; Beltran-Carbajal, F.; Sotelo, D.; Sotelo, C. Active Suspension Control Using an
MPC-LQR-LPV Controller with Attraction Sets and Quadratic Stability Conditions. Mathematics 2021, 9, 2533. [CrossRef]
30. Palanisamy, S.; Karuppan, S. Fuzzy Control of Active Suspension System. J. Vibroeng. 2016, 18, 3197–3204. [CrossRef]
31. Díaz-Choque, C.S.; Félix-Herrán, L.C.; Ramírez-Mendoza, R.A. Optimal Skyhook and Groundhook Control for Semiactive
Suspension: A Comprehensive Methodology. Shock Vib. 2021, 2021, 8084343. [CrossRef]
32. Boada, B.L.; Boada, M.J.L.; Vargas-Melendez, L.; Diaz, V. A Robust Observer Based on H ∞ Filtering with Parameter Uncertainties
Combined with Neural Networks for Estimation of Vehicle Roll Angle. Mech. Syst. Signal Process. 2018, 99, 611–623. [CrossRef]

112
 mathematics

Article
A Novel Denoising Method for Retaining Data Characteristics
Brought from Washing Aeroengines
Zhiqi Yan 1 , Ming Zu 2 , Zhiquan Cui 3, * and Shisheng Zhong 4

1 School of Mechatronics Engineering, Harbin Institute of Technology, Harbin 150001, China;

cumtmaris@163.com
2 School of Electronic Information & Media, Dongying Vocational Institute, Dongying 257091, China;
dyzm2006@126.com
3 School of Automotive Engineering, Harbin Institute of Technology, Weihai 264209, China
4 School of Ocean Engineering, Harbin Institute of Technology, Weihai 264209, China; zhongss@hit.edu.cn
* Correspondence: xiaocui2002yan@163.com

Abstract: Airlines evaluate the energy-saving and emission reduction effect of washing aeroengines
by analyzing the exhaust gas temperature margin (EGTM) data of aeroengines so as to formulate
a reasonable washing schedule. The noise in EGTM data must be reduced because they interfere
with the analysis. EGTM data will show several step changes after cleaning the aeroengine. These
step changes increase the difficulty of denoising because they will be smoothed in the denoising.
A denoising method for aeroengine data based on a hybrid model is proposed to meet the needs
of accurately evaluating the washing effect. Specifically, the aeroengine data is first decomposed
into several components by time and frequency. The amplitude of the component containing the
most noise is amplified, and Gaussian noise is added to generate noise-amplified data. Second, a
Gated Recurrent Unit Autoencoder (GAE) model is proposed to capture engine data features. The
GAE is trained to reconstruct the original data from the amplified noise data to develop its noise
Citation: Yan, Z.; Zu, M.; Cui, Z.;
Zhong, S. A Novel Denoising Method reduction ability. The experimental results show that, compared with the current popular algorithms,
for Retaining Data Characteristics the proposed denoising method can achieve a better denoising effect, retaining the key characteristics
Brought from Washing Aeroengines. of the aeroengine data.
Mathematics 2022, 10, 1485. https://
doi.org/10.3390/math10091485 Keywords: autoencoder; aeroengine; denoising; GRU

Academic Editors: Higinio Rubio

MSC: 68T09
Alonso, Alejandro Bustos Caballero,
Jesus Meneses Alonso and Enrique
Soriano-Heras

Received: 14 March 2022 1. Introduction

Accepted: 26 April 2022 An aeroengine will be polluted after long-term operation. Pollution of aeroengines
Published: 29 April 2022
leads to mechanical failure, degraded engine performance, and pollutant gas emission.
Publisher’s Note: MDPI stays neutral Washing is a main approach to controlling aeroengine pollution specified in aircraft mainte-
with regard to jurisdictional claims in nance manuals. Engine washing constitutes spraying cleaning fluid into the engine through
published maps and institutional affil- multiple spray guns to remove fouling in the Engine. Washing an aeroengine is expensive.
iations. In order to minimize the washing cost while maintaining the normal performance of the
engine, airlines must formulate a reasonable aircraft engine washing schedule.
A reasonable washing schedule can reduce 2735 tons of fuel consumption and 8626 tons
of harmful gas emissions per 100 engines per year, saving USD 10.08 million [1]. However,
Copyright: © 2022 by the authors.
the mechanical failure of the polluted engine causes high-frequency noise in the aeroengine
Licensee MDPI, Basel, Switzerland.
data. The noise of aeroengine data will interfere with the reasonable formulation of an
This article is an open access article
aircraft engine cleaning schedule, so it is necessary for airlines to denoise aeroengine data.
distributed under the terms and
There are two difficulties in denoising aeroengine data: the noise of the engine itself is
conditions of the Creative Commons
difficult to eliminate, and the step change of data generated by washing will be smoothed.
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
Aeroengines are man-made systems. Different from noise in nature, the probability distri-
4.0/).
bution of the engine’s own noise does not obey the normal distribution, which increases

Mathematics 2022, 10, 1485. https://doi.org/10.3390/math10091485 113 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 1485

the difficulty of denoising. After washing the aircraft engine, the data will experience a step
change. Traditional denoising methods often treat the step changes in data as noise and
smooth them out, making it difficult for the data to accurately reveal the washing effect,
adding difficulties for subsequent evaluation. Therefore, a new denoising model is needed
to eliminate the noise of the engine itself and retain the impact of cleaning on the data, as
shown in Figure 1.


 
(a) (b)

(c)

Figure 1. Demand for denoising aeroengine data. (a) Raw engine data; (b) Denoising effect of
traditional methods; (c) Denoising required for aircraft engine

Many noise reduction methods have been proposed, including empirical mode decom-
position (EMD), wavelet threshold denoising, and filtering. Xue Feng et al. [2] decomposed
the data with EMD after adding Gaussian noise to the data and successfully separated the
noise from the signal by searching the dominant component of noise through the continuous
mean square error criterion. Lu et al. [3] randomly shuffled the high-frequency noise part of
the data and then decomposed the data with EMD to achieve noise reduction. Kai et al. [4]
denoised complex images with a non-sampling wavelet transform method. Mohammad
Saleh Sadooghi et al. [5] denoised the compressor vibration signal of aeroengines with the
wavelet threshold denoising method. Because the effect of wavelet threshold denoising is
related to the threshold and threshold function, the author evaluates 84 matching effects of
the threshold and threshold function to search for the most suitable method for denoising
engine compressor vibration signals. P. Maragos et al. [6] proposed morphological filters,
which are filters composed of basic operations of mathematical morphology. Morphological
filters can selectively suppress image noise. M. H. Sedaaghi et al. [7] proposed mediated
morphological filters by combining morphological filters with media filtering and classical
gray scale morphological operators. Mediated morphological filters have a good effect on
eliminating salt and pepper noise in images. Yang Chao et al. [8] denoised the temperature
data with a two-way Kalman filter method. They improved the computational efficiency
by simplifying the Kalman filter algorithm by precalculating the filter coefficients.
The above methods show some disadvantages. The EMD algorithm must calculate
the dividing point between noise components and key characters; the Wavelet threshold
denoising algorithm must find the threshold and threshold function manually. These
two algorithms are inefficient, and the calculation results can only be approximate. The
Kalman filter is suitable for linear systems rather than highly nonlinear systems such as
aeroengines. The above three methods are not adaptive methods. Researchers must select
or set model parameters according to data characteristics [9]. Therefore, these methods are

114
Mathematics 2022, 10, 1485

not suitable for aeroengine data with non-uniform data distribution, hardly meeting the
needs of accurately evaluating the washing effect.
Autoencoders are adaptive artificial intelligence algorithms that are frequently used as
noise reduction models. Denoising Autoencoders (DAE) were first proposed by Bourlard
H. [10] to extract features from raw data. The idea of the denoising autoencoder is to make
the autoencoder reconstruct the original data from the noise-added data to obtain its noise
reduction ability. Vincent et al. [11] enhanced the robustness of the model by setting the
input of the autoencoder to zero at a specific scale. Song Hui et al. [12] used seismic data to
verify the above algorithm and found that the algorithm can filter out random noise with
strong intensity, but the algorithm’s efficiency is low. According to the characteristics of
underwater heterogeneous information data, Wang et al. [13] combined a three-layer sparse
autoencoder with two-layer convolution to build a model with strong noise reduction ability.
Peng et al. [14] added convolutional layers to the autoencoder, which made the autoencoder
more robust with a smaller reconstruction error. However, the signal amplitude is reduced
by this. Hui et al. [15] constructed a convolutional autoencoder to filter seismic data with
a low signal-to-noise ratio and achieved desirable results. Alexander Kensert et al. [16]
denoised the chromatogram to a completely or almost completely noise-free state with a
deep convolutional autoencoder.
However, the DAE also smooths the edge features of the data while denoising. Al-
though the literature [11,13] preserves the characteristics of the data to the greatest extent,
the premise is that the noise of underwater heterogeneous information data and seismic
data comes from nature, conforming to the Gaussian noise distribution. The noise of
aeroengine data comes from the engine system rather than nature. There is no reference
documenting that the engine noise conforms to the Gaussian noise distribution. Therefore,
the above-mentioned algorithms are not very applicable to engine data. Therefore, the ex-
isting denoising methods based on autoencoders can only eliminate the conventional noise,
but it is difficult to eliminate the specific noise of the aeroengine. These methods are not
suitable for engine data and hardly meet the need to accurately evaluate the washing effect.
To meet the needs of accurately evaluating the washing effect, a denoising method
for aeroengine data based on a hybrid model is proposed. This method can filter out the
noise of the aeroengine data, retaining the edge features of the data. The method first
splits the data by washing time, and then decomposes the data into several components by
frequency. Second, the method finds the component that contains the most noise, amplifies
its magnitude, and adds Gaussian noise to compose the noise-amplified data. Finally, the
method inputs the noise-amplified data together with the original data into the proposed
autoencoders, training the model’s ability to recover the detailed information of the engine
data from the noise-amplified data. Figure 2 shows the principle of the aeroengine data
denoising method based on the hybrid model.
The rest of the paper is as follows: The second part presents the decomposition method
of the aeroengine data and the identification method of the data frequency band containing
the most noise. A GAE model is proposed to denoise aeroengine data. The third part first
introduces the source of the engine data and secondly gives the identification results of
the data frequency bands containing the most noise and the determination results of the
hyperparameters of the GAE model. Finally, this part tests the noise reduction effect of
the EMD model, the EAD model, and the GAE model on aeroengine data. The fourth part
summarizes the superiority of the proposed aeroengine data denoising method through
the analysis of the above noise reduction testing results.

115
Mathematics 2022, 10, 1485

$HURHQJLQHGDWD

6SOLWE\ZDVKLQJ
UHFRUG

'HFRPSRVHLQWRVHYHUDOFRPSRQHQWV
DFFRUGLQJWRIUHTXHQF\

,GHQWLI\WKHFRPSRQHQWWKDW
FRQWDLQVWKHPRVWQRLVH

$PSOLI\WKHFRPSRQHQWWKDW
FRQWDLQVWKHPRVWQRLVH

$GG*DXVVLDQQRLVH

1RLVHDPSOLILHGGDWD 6SOLWHGGDWD

,QSXWWRWKHDXWRHQFRGHUWRWUDLQ

'HQRLVLQJPRGHO

Figure 2. Denoising model principle.

2. Denoising Method for Aeroengine Data Based on a Hybrid Model

The paper involves more abbreviations, and their meanings are given in Table 1.

Table 1. Meaning of abbreviations.

Abbreviations Full Title

DAE Denoising Autoencoders
EMD Empirical Mode Decomposition
EGTM Exhaust Gas Temperature Margin
DTW Dynamic Time Warping
IMF Intrinsic Mode Functions
GRU Gated Recurrent Unit Autoencoder
OEM Original Equipment Manufacturer
MAE Mean Absolute Error

2.1. Analysis of Noise Reduction Problems and Corresponding Solutions

The problems in eliminating engine noise and remaining data step sizes from washing
are analyzed from the following three aspects, and methods are proposed.
Problem 1: Difficulty in retaining the data step size generated by washing the engine.
Aeroengines are complex nonlinear systems. Engine data is the output of a complex
system that characterizes the performance and condition of the engine. Cleaning can be
viewed as a stimulus applied to the engine system. The engine system has changed state
due to being purged. As the system changes, the distribution of data output by the system
changes. This change will be removed as noise, thus smoothing out the data’s features.
The solution to this problem is to divide aeroengine data into several data segments
according to the washing time. The data distribution within each data segment does not
change, avoiding the smoothing of data features.

116
Mathematics 2022, 10, 1485

Problem 2: Difficulty in identifying engine-specific noise.

Each data segment contains an engine characteristic part and an engine-specific noise
part. The process of denoising is performed to maximize the separation of engine-specific
noise components from data segments. Traditional methods can identify noise with a
Gaussian distribution, but it is difficult to identify engine-specific noise.
The solution to this problem is to decompose the data segment into several components
with different frequencies. The similarity of each component to the noise data specific to
the aeroengine is calculated to identify the component that contains the most noise.
Problem 3: Difficulty in reducing engine-specific noise.
Because the components containing the most noise still contain engine characteristic
information, it is unscientific to directly delete the data components containing the most
noise. Removing these components results in a loss of the feature components in the
aeroengine data, smoothing the edge features of the data.
To reduce the loss of feature components, a new denoising method is proposed. First,
the amplitudes of the data components that contain the most noise are amplified. Artificial
noise is added to the data. Secondly, the autoencoder model is built. Finally, let the model
reconstruct the original data from the noise-amplified data, training its ability to perform
denoising.
Based on the above analysis, a denoising method for aeroengine data based on a mixed
model is proposed. The noise reduction process of this method is shown in Figure 3. After
the EGTM data were collected, they were split into several sets of data according to the
water wash records. The data components containing the most noise as well as artificial
Gaussian noise are superimposed on the original data. The original data are input into the
autoencoder for training to generate several sets of denoised EGTM data. Finally, several
sets of data are assembled into denoised complete EGTM data.
As shown in Figure 3, the denoising method for aeroengine data based on mixed
models includes a proposed method for identifying data components that contain noise
and a proposed denoising method.
Method 1: The proposed method for identifying data components that contain noise.
The paper first decomposes the aeroengine data according to the washing records;
secondly, the paper decomposes the data into several components by frequency with the
EMD algorithm; lastly, the noise content in the data components is evaluated with the
distance calculation method based on the DTW algorithm to find the data components that
contain the most noise.
Method 2: The proposed denoising method.
First, the amplitudes of the data components that contain the most noise are amplified.
Artificial noise is added to the data. Secondly, the autoencoder model is built. Finally, let
the model reconstruct the original data from the noise-amplified data, training its ability to
perform denoising.

2.2. Identification Method of Aeroengine Data Components Containing Noise Based on EMD
and DTW
This section is about splitting the data, decomposing the data, and identifying noise.
The washing time is used as the basis for splitting the engine data. The engine data
in this study come from OEM factories, and they are recorded with the flight cycle as the
observation time, so the aeroengine data is time series data. The engine washing records
are tables that record the time, location, engine type, and aircraft type when the engine was
washed. What can be used in this study is wash time. Data is denoted as X = {x1 , x2 , x3 , . . . }.
If the number of washings is n – 1, the washing records can be defined as: Twashing = {t1 , t2 ,
. . . , tn −1 }. Twashing can split data X into n pieces of split data. Let X(i) denote split data,
which can be given by Equation (1).
& '
X (i ) = xt(i−1) +1 , xt(i−1) +2 , · · · , xti , i = 1, 2, · · · , n (1)

117
Mathematics 2022, 10, 1485

FROOHFW

'DWDVSOLWWLQJ

X
6SOLWGDWD X
DFFRUGLQJ
WRZDVKLQJ
X
UHFRUG X &RQWDLQVWKHPRVWQRLV\
ήήή FRPSRQHQW

*DXVVLDQ +
'DWDGHFRPSRVLWLRQEDVHG QRLVH
RQ(0'
*$(DXWRHQFRGHU

1RLVHLGHQWLILFDWLRQEDVHG
RQ'7:
'HQRLVHGGDWDVHJPHQW

&RPELQDWLRQ

 'HQRLVHGGDWD
7KHSURSRVHGPHWKRGWRLGHQWLI\
WKHGDWDFRPSRQHQWVWKDWFRQWDLQ
WKHPRVWQRLVH 7KHSURSRVHGGHQRLVLQJPHWKRG

Figure 3. Denoising method for aeroengine data based on a hybrid model.

Then X(i) is decomposed in frequency with the EMD algorithm. The EMD algorithm
is a signal analysis algorithm, which decomposes the signal according to the timescale
of the aeroengine data without setting any basis functions in advance. Aeroengine data
can be divided into several “intrinsic mode functions” (IMF) by the EMD method without
deviating from the time domain. The aeroengine data can be expressed as the trend
component and the sum of several IMF functions.
The steps of EMD decomposing the data segment X(i) obtained by dividing the
aeroengine data are as follows:
Step 1: Find all the extreme points of the aeroengine data, use the spline curve to
connect all the maximum points into the upper envelope, and connect all the minimum
points to the lower envelope.
Step 2: Calculate the average value m(i) of the upper and lower envelopes and the
engine data IMF component c(i) according to Equation (2).

c(i) = X(i) - m(i) (2)

Nordne E. Hunag [17] put forward the concept of IMF and defined two conditions
of IMF: 1. In the whole time range, the number of local extreme points and zero-crossing
points of IMF differs by at most one; 2. The mean value of the upper and lower envelopes

118
Mathematics 2022, 10, 1485

is zero. If c(i) does not satisfy these two conditions, repeat steps 1 and 2 until they are
satisfied, and then go to step 3.
Step 3: Separate c(i) from the aeroengine data X(i): X(i) - c(i). Repeat step 1, step 2, and
step 3 until X(i) - c(i) becomes a monotonic sequence. Given r(i), defined by Equation (3):

r(i) = X(i) - ∑c(i) (3)

where r(i) is the residual component of engine data X(i).

The pseudocode of the EMD algorithm is given by Algorithm 1.

Algorithm 1. EMD algorithm

IMFs = []
While haspeaks(X(i)):
maximum_points = search(X(i), maximum)
minimum_points = search(X(i), minimum)
max_envelope = fitting(maximum_points)
min_envelope = fitting(minimum_points)
c(i) = (max_envelope + min_envelope)/2
r(i) = X(i) − c(i)
IMFs.append(data_imf)

Real engine data can be seen as a superposition of ideal data and noise data caused by
minor faults. To identify the characteristic noise components of aeroengines, four kinds
of fault data were prepared in this study, including compressor fault data, fan fault data,
high-pressure turbine fault data, and low-pressure turbine fault data. The distance of the
fault data from each data component is calculated to find the data component with the
most noise.
However, each component of the engine data has unequal data lengths and non-
existing linear correspondences with the four types of fault data. There are phenomena
such as amplitude scaling and linear drift between data [18]. The Euclidean distance is
directly used to represent the similarity between each component of the engine data and
the four types of fault data, which may encounter the problem of information flooding
to a certain extent, resulting in an unrealistic distance between each component of the
engine data and the four types of fault data. Therefore, a distance calculation method that
is compatible with various data lengths is required.
The DTW algorithm is proposed for measuring the distance between two time series
of different lengths. DTW is widely used in the field of speech recognition, and it is also
suitable for recognizing two similar aeroengine data sets. In this study, DTW is used to
calculate the minimum distance between the four types of fault data and c(i) or r(i).
The process of the DTW algorithm is shown in Figure 4. First, a matrix grid is created,
and the length and width are respectively the lengths of the engine data component and
the length of the fault data. The elements of the matrix represent the Euclidean distance
d of the data at the corresponding position of the engine data component and the fault
data. Then take the lower-left corner (1,1) point of the matrix grid as the starting point and
move to the adjacent grid corresponding to the smallest element of the right and upper
elements until it reaches the upper-right corner of the matrix grid to form a path. Finally,
calculate the cumulative value of the matrix elements that the path passes through. This
accumulated value is the distance between the engine data component and the fault data.

119
Mathematics 2022, 10, 1485


Figure 4. Principle of DTW algorithm.

The pseudocode of the DTW algorithm is given by Algorithm 2.

Algorithm 2. DTW algorithmM

n = length(engine_data)
m = length(fault_data)
dtw_matrix = [∞](n+1)×(m+1)
dtw_matrix(0,0) = 0
for i = 2:n + 1
for j = 2:m + 1
dtw_matrix(i,j) = ∑[engine_data(i − 1) − fault_data(j − 1)]2 + min[dtw_matrix(i − 1,j),
dtw_matrix(i − 1,j − 1), dtw_matrix(i,j − 1)]
end
end
return dtw_matrix(n + 1, m + 1)

Define the distance as dis. The distance of c(i) or r(i) and the four kinds of fault data
are given in Table 2, as are symbols and explanations.

Table 2. Distance table.

dis Interpretation of dis

disc(i) Distance sum between c(i) and all fault data
disr(i) Distance sum between r(i) and all fault data
disc(i), FAN Distance between c(i) and fan fault data
disc(i), COMP Distance between c(i) and compressor fault data
disc(i), HPT Distance between c(i) and high-pressure turbine fault data
disr(i), FAN Distance between r(i) and fan failure data
disr(i), COMP Distance between r(i) and compressor failure data
disr(i), HPT Distance between r(i)and high-pressure turbine failure data
disr(i), LPT Distance between r(i) and low-pressure turbine failure data

In Table 2, the distance between c(i) and all fault data can be given by Equation (4).

disc(i) = ∑(disc(i), FAN , disc(i), COMP , disc(i), HPT , disc(i), LPT ) (4)

The distance between r(i) and all fault data can be given by Equation (5).

disr(i) = ∑(disr(i), FAN , disr(i), COMP , disr(i), HPT , disr(i), LPT ) (5)

The smaller the distance, the more noise it contains. The engine data component
corresponding to the smallest distance is selected as the noise component of the aeroengine.

120
Mathematics 2022, 10, 1485

2.3. Gated Recurrent Unit Autoencoder(GAE): A Proposed Denoising Autoencoder Model for
Aeroengine Data
DAE is a promising method for data denoising, which can be used for aeroengine data
denoising. It is a feature extractor with a denoising function whose purpose is to convert
noisy aeroengine data into clean aeroengine data.
In addition, the aeroengine data is a time series, and each element in the sequence
is affected by all the previous elements, which requires the model to be able to learn the
influence from the historical accumulation of the aeroengine data. Real engine data is
difficult to obtain, and the limited amount of data has difficulty supporting large-scale
models. This requires the model to have a simple structure, and the GRU (Gated Recurrent
Unit) model meets the requirements. The GRU was proposed by Cho et al. [19]. GRUs have
fewer parameters, making training faster and requiring less data [20].
Therefore, in this paper, the Gated Recurrent Unit Autoencoder (GAE) model is
proposed as a denoising module for aeroengine data by combining the autoencoder with
the GRU. The structure of the GAE model is shown in Figure 5. The model includes
encoders and decoders with special structures. The input end of the encoder and the output
end of the decoder are both GRU modules, and a three-layer autoencoder is used as the
connection in the middle. The three layers of the autoencoder are marked as h1 , h2 , and
h3 . In the coding stage, the aeroengine data is continuously input to the GRU and the
characteristic data is output through h1 and h2 . In the decoding stage, the feature data is
input into the h3 layer, and the denoised data is output through the GRU module.

*DXVVLDQ &RPSRQHQWFRQWDLQLQJ
QRLVH WKHPRVWQRLVH

+
(QFRGHU 'HFRGHU

*58 *58
+

y

*58 *58
+

y
Xi

*58 *58
+

h ...
h h
*58 *58
+

Figure 5. GAE Model.

In Figure 5, the input data is the aeroengine data that amplifies its own noise and adds
Gaussian noise, which can be expressed as X(i)noise ; the label data is the raw aeroengine
data, expressed as X(i). The GRU involved in Figure 5 has two doors: update door and
reset door. the basic structure of the GRU is shown in Figure 6.
In Figure 6, the definitions are as follows: z is the update gate, r is the reset gate, h is
the current state, and h—1 is the previous state. The update gate z is used to control the
amount of previous state information which is brought into the current state. The reset
gate r controls how much information from the previous state is written to the current
candidate set. The smaller the reset gate, the less information from the previous state is

121
Mathematics 2022, 10, 1485

written. Define the current candidate set as +

h. For the input data X(i)noise , the forward
propagation formula of the network can be constructed, which is given by Equation (6):
⎧
⎪ r = σ(Wr · [ h−1 , X (i )noise ])
⎪
⎪
⎪
⎨ z = σ(Wz · [ h−1 , X (i )noise ])
+
h = tanh(Wh · [r · h−1 , X (i )noise ]) (6)
⎪
⎪ +
⎪
⎩ h = (1 − z ) · h −1 + z · h
⎪
y = σ(Wy · h)

where r is the weight from the input and the hidden layer at the previous moment to the
reset gate r; Wz is the weight from the input and the hidden layer at the previous moment
to the update gate z; Wh is the input and the hidden layer of the previous moment to +
h; Wy
is the weights from the hidden layer to the output layer.

hü × +

× ü
r z ×
~
ı ı h
7DQK

Figure 6. The structure of GRU.

The h2 layer’s output of the denoising autoencoder model is the feature code of the
aeroengine data. The encoding process is that the encoding function maps high-dimensional
aeroengine data vectors to low-dimensional feature vectors. The encoder function is the
activation function of the h2 layer of the autoencoder. The activation function is defined as
the sigmoid function. The feature vector output by the h2 layer is co. Then the aeroengine
data encoding process of the h2 layer is given by Equation (7).

1
co = (7)
1 + e−whid ·y−bhid
where whid is the h2 layer weight matrix; bhid is the h2 layer bias vector.
The decoding process maps the feature vector co to the reconstructed aeroengine data.
The activation function of the h3 layer in the decoder is defined as a linear function, and the
feature vector output by the h3 layer is deco, which is given by Equation (8).

deco = wout · h + bout (8)

where wout is the h3 layer weight matrix; bout is the h3 layer bias vector.
Finally, deco is output by the GRU module as denoised data, which is defined as
X(i)denoise . The loss function in this model is the mean absolute error (MAE). Define the
number of output nodes of the GAE model as nout , then MAE is given by Equation (9).
nout
1
nout j∑
MAE( X (i )denoise , X (i )noise ) = ( X (i )denoise − X (i )noise )2 (9)
=1

3. Experiment
All functions are written on the python 3.5 platform of the Windows system with
Tensorflow as the framework. Tensorflow is an open source machine learning platform. The

122
Mathematics 2022, 10, 1485

hardware platforms involved in the calculation are CPU and GPU; the models are Intel(R)
Core (TM) i3-9100F CPU_3.60GHz and NVIDIA GeForce GTX TITAN X, respectively.
The data involved in this paper is aeroengine EGTM data. Airlines take EGTM as a
reference for aeroengine performance. EGTM refers to the difference between the red line
value of engine exhaust temperature and the exhaust temperature when the engine takes
off at full thrust. When the engine is washed, there are step changes in the EGTM data.
During denoising, these step changes may be recognized as noise and smoothed. There-
fore, selecting EGTM data as the noise reduction object can better verify the performance of
the proposed method.
Since exhaust gas temperature margin (EGTM) data is an important indicator for
evaluating the efficiency of engine washing, the data form of the paper is EGTM data that
includes the number of cycles recorded in the water wash. The EGTM data is collected at
the outlet of the low-pressure turbine of the engine, and Figure 7 is the measurement point
of the exhaust temperature of the aeroengine.

Figure 7. Aeroengine data sources.

EGTM data is real data provided by the OEM. In this study, the flight cycle was used
as a time unit to record the aeroengine data from takeoff to landing. Therefore, aeroengine
data is time series data. In practice, the airline provided three materials for an aeroengine,
including OEM data in Table 3, fault records in Table 4, and washing records in Table 5.

Table 3. OEM data.

ID ESN Time EGTM

B-5793 657208 16 September 2013 1:52 90.845
B-5793 657208 16 September 2013 5:09 84.52
B-5793 657208 16 September 2013 8:50 87.208
B-5793 657208 16 September 2013 12:31 89.306
B-5793 657208 17 September 2013 8:00 82.397
B-5793 657208 17 September 2013 11:45 85.755
B-5793 657208 18 September 2013 9:44 85.973
B-5793 657208 21 September 2013 0:06 66.281
B-5793 657208 21 September 2013 4:08 86.617
B-5793 657208 21 September 2013 8:36 75.281

Table 4. Fault records.

ID Full Title Noise Source Noise Type

B2530 1 August 2009 Abnormal left engine Compressor fault noise
B2588 2 January 2014 Fan seal broken Fan fault noise
B6076 26 February 2002 Turbine oil leakage Turbine fault noise
B6070 2 January 2014 Abnormal turbine blade Turbine fault noise

123
Mathematics 2022, 10, 1485

Table 5. Washing records.

Date ID Base CAMP

8 September 2014 0:00 B-1816 Beijing A320 720000-CCA-C-02
8 September 2014 0:00 B-1816 Beijing A320 720000-CCA-C-02
28 September 2014 0:00 B-2210 Hangzhou A320 720000-CCA-C-02
16 January 2015 0:00 B-2210 Hangzhou A320 720000-CCA-C-02
9 March 2014 0:00 B-2210 Hangzhou A320 720000-CCA-C-03
19 April 2014 0:00 B-2210 Hangzhou A320 720000-CCA-C-03
16 April 2014 0:00 B-2364 Chengdu A320 720000-CCA-C-02
16 April 2014 0:00 B-2364 Chengdu A320 720000-CCA-C-02

The data required for the study are spread out across three tables. The collection
process of the data to be denoised is: According to the aircraft registration number (ID) and
water washing date in Table 5, the IDs of the OEM data in Table 3 are located to collect all
the corresponding EGTM data; the collection process of the fault data is: According to the
ID in the fault record in Table 4, the ID of the OEM data in Table 3 is located to collect all
the corresponding EGTM data. Figure 8 shows the aeroengine EGTM data. Figure 9 shows
all types of data.


 (*70
:DVKLQJWLPH

(*70ႏ





 
          
)OLJKW&\FOHV
Figure 8. The engine data involved in this paper.

During the wing period, the engine was washed a total of nine times. Taking the
washing time as the time point, the aeroengine data is split into eight segments. The first
six segments are used as the training set, and the last two segments are used as the testing
set. Number the eight segments sequentially to obtain Table 6.

Table 6. Split aeroengine data grouping table.

Data Number Flight Cycles

Training data 1 1~329
2 330~481
3 482~708
4 709~1142
5 1143~1484
6 1485~1711
Testing Data 7 1712~1981
8 1982~2222

124
Mathematics 2022, 10, 1485

 




(*70ႏ
(*70ႏ

 

 



 
                   
)OLJKW&\FOHV )OLJKW&\FOHV
(a) (b)
 í
 í
í

(*70ႏ
(*70ႏ

 í
í

í
 í
í

                   
)OLJKW&\FOHV )OLJKW&\FOHV
(c) (d)

Figure 9. The fault data involved in this paper. (a) EGTM data for compressor failure; (b) EGTM
data for fan failure; (c) EGTM data for high pressure turbine failure; (d) EGTM data for low pressure
turbine failure.

3.1. Noise Identification Results for Data Components

To identify the data components that contain the most noise, the aeroengine data
components are first decomposed based on EMD after being collected. Secondly, based on
DTW technology, the noise identification of each component is carried out. The aeroengine
EGTM data in Figure 8 are decomposed into residual components and IMF components.
Figure 10 shows the two components derived from the decomposition of the six-segment
aeroengine training data.
Based on the DTW algorithm, the distances were calculated between the two com-
ponents of aeroengine EGTM data and four kinds of fault data, such as compressor fault
data, fan fault data, high-pressure turbine fault data and low-pressure turbine fault data.
In the calculation results, a very large number (>1 × 10308 ) appeared in the calculation of
the distances, marked as “∞”. Table 7 records the distances between the IMF component of
the six-segment aeroengine training data and the four kinds of fault data. Table 8 records
the distances between the residual component of the six-segment aeroengine training data
and the four types of fault data.

Table 7. Distances of IMF components.

Data Fan Fault Compressor High Pressure Low Pressure

Number Data Fault Data Turbine Fault Data Turbine Fault Data
1 5.236 5.227 6.493 6.199
2 4442.453 5582.838 ∞ ∞
3 10,463.488 12,162.886 ∞ ∞
4 2.907 4.205 2.521 5.467
5 3.167 3.981 3.0158 6.385
6 6209.860 7909.258 ∞ ∞

125
Mathematics 2022, 10, 1485


 1

(*70ႏ

(*70ႏ
 
0
 ƺ1

     
)OLJKW&\FOHV )OLJKW&\FOHV
(a) (b)

1
(*70ႏ

(*70ႏ

 0
 
ƺ1
     
)OLJKW&\FOHV )OLJKW&\FOHV
(c) (d)

(*70ႏ

(*70ႏ
0.5

 0

ƺ0.5
       
)OLJKW&\FOHV )OLJKW&\FOHV
(e) (f)
2
(*70ႏ

(*70ႏ

 1
 0

ƺ1

       
)OLJKW&\FOHV )OLJKW&\FOHV
(g) (h)


(*70ႏ

(*70ႏ

1
 0
 
ƺ1
 
ƺ2
     
)OLJKW&\FOHV )OLJKW&\FOHV
(i) (j)

 1
(*70ႏ

(*70ႏ

 0

 
ƺ1
       
)OLJKW&\FOHV )OLJKW&\FOHV
(k) (l)

Figure 10. Components of the six-segment aeroengine training data. (a) No. 1 data residual com-
ponent; (b) No. 1 data IMF component; (c) No. 2 data residual component; (d) No. 2 data IMF
component; (e) No. 3 data residual component; (f) No. 3 data IMF component; (g) No. 4 data residual
component; (h) No. 4 data IMF component; (i) No. 5 data residual component; (j) No. 5 data IMF
component; (k) No. 6 data residual component; (l) No. 6 data IMF component.

126
Mathematics 2022, 10, 1485

Table 8. Distances of residual components.

Data Fan Fault Compressor High Pressure Low Pressure

Number Data Fault Data Turbine Fault Data Turbine Fault Data
1 10.703 13.716 9.991 13.171
2 ∞ ∞ ∞ ∞
3 ∞ ∞ ∞ ∞
4 8.680 11.051 8.301 11.654
5 9.745 11.480 9.929 14.420
6 ∞ ∞ ∞ ∞

In Tables 7 and 8, the distances between the residual components of No. 2, 3, and 6 in
the training data and the four types of fault data are all “∞”, indicating that these three sets
of residual component data hardly contain noise components. The distances between the
residual components of No. 1, 4, and 5 and the four kinds of fault data are all within 15,
while the distances between the IMF components of these three groups and the four kinds
of fault data are smaller than those of the residual components, which shows that more
noise is included in the IMF component of the aeroengine EGTM data. Therefore, the IMF
component is selected as the data component containing the most noise. Separate the data
components that contain the most noise in the testing data (No. 7 and No. 8), as shown
in Figure 11.

2
1
(*70ႏ

(*70ႏ

1
0 0
 ƺ1
ƺ1 

        

)OLJKW&\FOHV )OLJKW&\FOHV
(a) (b)

Figure 11. Components that contain the most noise in the testing data. (a) Data No. 7 contains the
most noise data component; (b) Data No. 8 contains the most noise data component.

3.2. Hyperparameter Settings for GAE

In this part, the number of nodes in the h1 layer and the number of nodes in the h2
layer of the GAE model are determined. Define the number of nodes in the h1 layer as nin ,
the number of nodes in the h2 layer as nhid , and the number of nodes in the h3 layer as nout .
Since the number of nodes in the h1 layer is same as that in the h3 layer, nin = nout .
There is currently no formula defining the nin . To obtain reliable results, the enumera-
tion method is used to determine the most suitable nin . The enumeration range is 15–25.
The mean absolute error (MAE) is used to describe the reconstruction error of the denoising
autoencoder for the null engine data.
The model uses the Adam algorithm as the descent algorithm. The number of iterations
is 3000 and the training batch is 100. The learning rate is 0.001, and the component
containing the most noise is amplified by a factor of 1.1. The model was run 10 times, and
the nin and the reconstruction error were plotted as shown in Figure 12.
The vertical line in Figure 12 marks the nin when the reconstruction error is the smallest.
In order to analyze the results in more detail, the mean, standard deviation, and minimum
and maximum values of 10 experiments are given in Table 9. The data in the Table 9 are
reserved for three significant figures.

127
Mathematics 2022, 10, 1485

0.06

Reconstruction error
0.055
0.05
0.045
0.04
0.035
15 20 25
nin
Figure 12. Relationship between the number of nodes in the h1 layer and reconstruction error.

Table 9. Table of the nodes number in the h1 layer and reconstruction error data.

nin Mean Reconstruction Error Standard Deviation Minimum Maximum

15 0.0442 0.0476 0.0174 0.152
16 0.0420 0.0237 0.0208 0.0863
17 0.0486 0.0317 0.0236 0.111
18 0.0420 0.0286 0.0190 0.114
19 0.0398 0.0272 0.0197 0.117
20 0.0376 0.0397 0.0170 0.140
21 0.0481 0.0349 0.0261 0.150
22 0.0484 0.0441 0.0173 0.168
23 0.0548 0.0356 0.0254 0.114
24 0.0528 0.0496 0.0194 0.117
25 0.0497 0.0507 0.0223 0.186

From the data given in Table 9, the average reconstruction error is the lowest when
nin is 20. The average reconstruction error is slightly higher when nin is 19. However, the
standard deviation when nin is 20 is larger than that when nin is 19, being 0.0397 and 0.0272,
respectively. The minimum error value when nin is 20 is smaller than that when nin is 19.
The maximum error when nin is 20 is greater than that when nin is 19. Table 5 shows that
the output is unstable when nin is 20, although the average reconstruction error is small.
Although the average reconstruction error is slightly larger, the output is stable when nin is
19. Therefore, nin is determined to be 19 in this paper.
Similarly, to minimize reconstruction error, nhid must be determined. At present,
researchers have developed empirical formulas [21] for nhid , which help narrow the search
for optimal nodes. The empirical formula is given by Equation (10).
√
nhid = nin × nout (10)

According to Equation (10), it can be roughly determined that nhid is around 19. The
optimal nhid is searched by the enumeration method in the interval of 15–25. After the
experiment is carried out 10 times, nhid and the reconstruction error are plotted in Figure 13.
As can be seen from Figure 13, when nhid is 21, the reconstruction error is the lowest.
To analyze the results in more detail, in Table 10, the mean reconstruction error, standard
deviation, minimum error, and maximum error of 10 experiments are given. The above
four types of data in Table 10 all retain three significant figures.

128
Mathematics 2022, 10, 1485



5HFRQVWUXFWLRQHUURU






     
nKLG

Figure 13. Relationship between nhid and reconstruction error.

Table 10. Table of nhid and reconstruction error.

nhid Mean Reconstruction Error Standard Deviation Minimum Maximum

15 0.0172 0.00522 0.00900 0.0260
16 0.0151 0.00460 0.00900 0.0220
17 0.0157 0.00748 0.00900 0.0350
18 0.0143 0.00487 0.0100 0.0270
19 0.0129 0.00375 0.00900 0.0190
20 0.0153 0.00503 0.00800 0.0260
21 0.0123 0.00343 0.00800 0.0180
22 0.0144 0.00469 0.00900 0.0220
23 0.0128 0.00418 0.00700 0.0190
24 0.0132 0.00486 0.00900 0.0230
25 0.0133 0.00421 0.00800 0.0230

From the data given in Table 10, the mean reconstruction error is the lowest when nhid
is 21. The standard deviation is the smallest when nhid is 21. This means that the model’s
output is currently stable. Therefore, the most suitable nhid is determined to be 21.

3.3. Validation of Aeroengine Data Denoising Method

3.3.1. Reconstruction Accuracy Verification of GAE
First, the reconstruction accuracy of the GAE model is verified. Figure 14 visually
shows the reconstruction errors of the three denoising models. Data No. 1–6 are the recon-
struction error of the training data in the figure, and data No. 7–8 are the reconstruction
error of the Testing data. The figure shows that the reconstruction error of the denoised
data output by the EMD model is much larger than that of the GAE model and the EAD
model. Except for data No. 4 and No. 8, the reconstruction error of the GAE model is
smaller than that of the EAD model.
The reconstruction errors of the DAE, EMD, and GAE models on the training data are
given in Table 11.

Table 11. Reconstruction errors of training data.

Model Reconstruction Error of Training Data Average

GAE 0.0810 0.0909 0.0637 0.0681 0.0585 0.0861 0.0747
DAE 0.0853 0.0995 0.0991 0.0628 0.0901 0.0817 0.0864
EMD 0.185 0.3578 0.320 0.288 0.248 0.705 0.351

129
Mathematics 2022, 10, 1485


 *$( '$( (0'

5HFRQVWUXFWLRQHUURU







       
'DWDQXPEHU
Figure 14. Comparison of reconstruction accuracy of three noise reduction models.

Tables 11 and 12 show that the MAE error of GAE training data is (0.0864 − 0.0747)/
0.0864 × 100% = 13.54% lower than that of DAE and 78.72% lower than that of EMD; the
MAE error of GAE test data is 3.11% lower than that of DAE and 76.80% lower than that
of EMD.

Table 12. Reconstruction errors of testing data.

Model Reconstruction Error of Testing Data Average

GAE 0.0807 0.0807 0.0965
DAE 0.161 0.161 0.0996
EMD 0.365 0.467 0.416

The MAE error of GAE is relatively smaller than that of DAE, which reflects that
the GRU in GAE can learn the time series relationship of aeroengine data so as to better
reconstruct pure data. The combination of the GRU module and the DAE model effec-
tively improves the noise reduction effect, making the MAE error smaller than that of the
DAE model.
The MAE error of GAE is much smaller than that of EMD. The reason for this is that
the algorithm of EMD is susceptible to noise interference. The algorithm of EMD separates
the noise signal by calculating the envelopes of the extrema of the data. Noise makes the
extremum of the data change, and the envelopes follow. The changes of the envelope also
make the EMD algorithm unstable, resulting in a large error.
To summarize, Tables 8 and 11 reflect that the noise reduction effect of the DAE model
is slightly better than that of the EMD model. The mean reconstruction error of the GAE
model is smaller than that of the EAD model. The data reconstruction ability of the GAE
model is stronger than that of the DAE model and the EMD model. It is proved that the
GAE model can preserve the characteristics of the aeroengine better.

3.3.2. The Effectiveness of the Proposed Noise Reduction Method Based on Hybrid Model
Is Verified
Figure 15 shows the reconstruction curves of DAE, EMD, and the proposed model on
the EGTM Testing data of an aeroengine. In the figure, the original data are marked with
black curves, the denoised data of the proposed model are marked with green dotted lines,
the denoised data of the EAD model are marked with light blue, and the noise reduction of
the EMD model are marked with dark blue.

130
Mathematics 2022, 10, 1485




(*70ႏ



2ULJLQDO'DWD
(0'
7KHSURSRVHGPRGHO
($'
 :DVKLQJWLPH

       
)OLJKW&\FOHV

Figure 15. Comparison of noise reduction effects of three noise reduction models.

The red vertical line in Figure 15 represents the time of washing the aeroengine. Due
to the washing, there is a step change in the original EGTM data. EAD, EMD, and the
proposed method show different performances under the influence of cleaning. The EAD
method significantly smoothed the step change of EGTM. It shows that the EAD model has
an underfitting problem in training. The EMD method separates the step change as noise
from the EGTM, thus smoothing the step change of the EGTM.
In Figure 15, the proposed model can reconstruct the EGTM sequence and express
data mutation well. The denoised data of the proposed model in the figure is closer to the
original data than DAE and EMD, which shows that the denoised data of the proposed
model is more suitable for the subsequent analysis of washing effects. In order to more
accurately express the noise reduction effect of DAE, EMD, and the proposed model on
aeroengine data, the EGTM step value after aeroengine washing is investigated.
The influence of the denoising model on the subsequent evaluation of the washing
effect can be known by calculating step size after washing. The step size after washing is a
direct expression of the washing effect. Therefore, calculating the post-washing step size of
the denoised data is an important indicator to investigate whether the denoising model
will affect the evaluation of the washing effect. Since the data after washing has a gradual
increase rather than a sudden increase, the step size calculated in the paper is the difference
between the average values of 10 data before and after washing. The step size is calculated
from the No. 2 data.
Figure 16 visually shows the steps sizes of No. 2–8 raw data and the step sizes of
denoised data of DAE, EMD and the proposed model. Table 13 gives the step sizes of the
training data and testing data. The data in Table 13 are the steps sizes of each piece of data.
Since No. 1 is unwashed engine data, the step size is recorded from No. 2 data.

Table 13. Step of each segment of data.

Step Size of the EGTM/◦ C

Model
Training Data Testing Data
Original data 11.326 5.212 7.199 12.206 12.078 5.613 6.971
The proposed model 10.604 4.841 6.832 11.292 11.555 5.584 5.108
DAE 9.649 4.525 5.435 9.979 10.958 5.871 2.709
EMD 10.220 2.813 5.067 8.137 9.469 10.288 0.738

131
Mathematics 2022, 10, 1485


2ULJLQDOGDWD 7KHSURSRVHGPRGHO '$( (0'


7KHVWHSVVL]HVRIWKH(*70ႏ



      
'DWDQXPGHU
Figure 16. The step sizes from data of all models.

Table 13 is the variance of the steps of the original data and the steps of the denoised
data of DAE, EMD, and the proposed model.
It can be seen from Table 14 that EMD has a large variance in both the testing set and
the training set. The variances of the DAE in the training set and testing set are much
smaller than those of the EMD, but they are still large. The variances of the proposed model
are extremely small in both the testing set and the training set. The step sizes calculated
from the denoised data of the proposed model are the closest to the original data. Therefore,
the proposed model has the least impact on the analysis of subsequent washing effects. It
is more suitable for aeroengine data noise reduction than other models.

Table 14. Step value variance between noise reduction data and original data.

Model Training Data Testing Data

The proposed model 0.381 1.736
DAE 2.522 9.116
EMD 6.978 30.358

4. Conclusions
To improve prediction accuracy, a denoising method for aeroengine data based on
a hybrid model is proposed in this paper. The method first amplifies the noise part of
the data, and then adds Gaussian noise to the data as the input of the autoencoder. Let
the autoencoder reconstruct the original data from the amplified noise data so that the
autoencoder can perform targeted noise reduction. In the paper, the proposed model is
compared with EMD and DAE, which reflects that the proposed model can effectively
denoise the data and retain mutation characteristics after aeroengine washing.
The autoencoder involved in the hybrid-model-based aeroengine data denoising
method is the GAE model proposed in this paper. The GAE model is composed of three
fully connected layers connecting two GRU modules. The model is good at working with
time series data. After testing with real aeroengine data, compared with EMD and DAE,
the reconstruction error of the GAE model is the smallest, preserving the data features to
the greatest extent.
The model proposed in this paper has an ideal effect on the denoising of EGTM data
after aeroengine washing. This model is applicable for denoising the various data with
sudden changes, such as the gas path data of aeroengines or gas turbines after maintenance.
In the future, we will plan to collect more real data to improve our methods.

132
Mathematics 2022, 10, 1485

Author Contributions: Methodology, M.Z.; writing—original draft preparation, Z.Y., Z.C., and S.Z.;
writing—review and editing, Z.C.; funding acquisition, M.Z. and Z.C. All authors have read and
agreed to the published version of the manuscript.
Funding: This research was funded by the National Natural Science Foundation of China, grant
number U2133202; and the National Natural Science Foundation of China, grant number 51975157.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Chen, D.; Sun, J. Fuel and emission reduction assessment for civil aircraft engine fleet on-wing washing. Transp. Res. Part D
Transp. Environ. 2018, 65, 324–331. [CrossRef]
2. Xue, F.; Sun, X.; Dong, Z.; Yang, H.; Wang, H. Research on Data Denoising Algorithm Based on EEMD. Mech. Eng. Autom. 2021, 5,
9–11.
3. Lu, T.; Qian, W.; He, X.; Le, Y.; Huang, J. An improved EMD noise reduction method based on noise statistical characteristics, Bull.
Surv. Mapp. 2020, 11, 71–75.
4. Hu, K.; Cheng, Q.; Li, B.; Gao, X. The complex data denoising in MR images based on the directional extension for the undecimated
wavelet transform. Biomed. Signal Process. Control 2018, 39, 336–350. [CrossRef]
5. Sadooghi, M.S.; Khadem, S.E. A new performance evaluation scheme for jet engine vibration signal denoising. Mech. Syst. Signal
Process. 2016, 76, 201–212. [CrossRef]
6. Maragos, P.; Schafer, R.W. Morphological Filters. Part 1. Their Set-Theoretic Analysis and Relations to Linear Shift-Invariant
Filters. IEEE Trans. Acoust. Speech Signal Process. 1987, 35, 1153–1169.
7. Sedaaghi, M.H.; Daj, R.; Khosravi, M. Mediated morphological filters. In Proceedings of the 2001 International Conference on
Image Processing (Cat. No.01CH37205), Thessaloniki, Greece, 7–10 October 2001; pp. 692–695.
8. Yang, C.; Li, J.; Yang, W.; Yang, W. Denoising Method for Temperature Log Data Based on A Kalman Filter. Well Logging Technol.
2020, 2, 168–171.
9. Li, Y.; Wang, C.; Tian, Y.; Wang, S. Parameter-shared variational auto-encoding adversarial network for desert seismic data
denoising in Northwest China. J. Appl. Geophys. 2021, 11, 104428. [CrossRef]
10. Bourlard, H.; Kamp, Y. Auto-association by multilayer perceptrons and singular value decomposition. Biol. Cybern. 1988, 4,
291–294. [CrossRef] [PubMed]
11. Vincent, P.; Larochelle, H.; Bengio, Y.; Manzagol, P.A. Extracting and Composing Robust Features with Denoising Autoencoders.
In Proceedings of the 25th international conference on Machine Learning, Helsinki, Finland, 5–9 June 2008.
12. Song, H.; Gao, Y.; Chen, W.; Zhang, X. Seismic noise suppression based on convolutional denoising autoencoder. Oil Geophys.
Prospect. 2020, 6, 1210–1219.
13. Wang, X.; Zhao, Y.; Teng, X.; Sun, W. A stacked convolutional sparse denoising autoencoder model for underwater heterogeneous
information data. Appl. Acoust. 2020, 167, 107391. [CrossRef]
14. Peng, F.; Gao, Y. BPSK Signal Denoise Based on Convolution Auto-Encoder Network. Inf. Commun. 2020, 8, 41–44.
15. Song, H.; Gao, Y.; Chen, W.; Xue, Y.J.; Zhang, H.; Zhang, X. Seismic random noise suppression using deep convolutional
autoencoder neural network. J. Appl. Geophys. 2020, 178, 104071. [CrossRef]
16. Kensert, A.; Collaerts, G.; Efthymiadis, K.; Van Broeck, P.; Desmet, G.; Cabooter, D. Deep convolutional autoencoder for the
simultaneous removal of baseline noise and baseline drift in chromatograms. J. Chromatogr. A 2021, 1, 462093. [CrossRef]
[PubMed]
17. Wu, Z.; Huang, N.E. A study of the characteristics of white noise using the empirical mode decomposition method. Proc. R. Soc.
London. Ser. A Math. Phys. Eng. Sci. 2004, 460, 1597–1611. [CrossRef]
18. Park, S.; Chu, W.W.; Yoon, J.; Won, J. Similarity Search of Time-Warped Subsequences via a Suffix Tree. Inf. Syst. 2003, 7, 867–883.
[CrossRef]
19. Chung, J.; Gulcehre, C.; Cho, K.H.; Bengio, Y. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling.
arXiv 2014, arXiv:1412.3555.
20. Olah, C. Understanding Lstm Networks. 2015. Available online: http://colah.github.io/posts/2015-08-Understanding-LSTMs
(accessed on 9 March 2021).
21. Sequin, C.H.; Clay, R.D. Fault tolerance in artificial neural networks. In Proceedings of the 1990 IJCNN International Joint
Conference on Neural Networks, San Diego, CA, USA, 17–21 June 1990; pp. 703–708.

133
 mathematics

Article
Adaptive Rejection of a Sinusoidal Disturbance with Unknown
Frequency in a Flexible Rotor with Lubricated Journal Bearings
Gerardo Amato 1 , Roberto D’Amato 2,3, * and Alessandro Ruggiero 4

1 Nonlinear and Adaptive Controls Laboratory, Department of Electronic Engineering, University of Rome
“Tor Vergata”, Via del Politecnico, 00133 Rome, Italy; amato@ing.uniroma2.it
2 Escuela Técnica Superior de Ingeniería y Diseño Industrial, Universidad Politécnica de Madrid,
Ronda de Valencia, 28012 Madrid, Spain
3 Structural Materials Research Center (CIME-UPM), Departamento de Ciencia de Materiales ETS de Ingenieros
de Caminos, Canales y Puertos, C/Profesor Aranguren S/N, 28040 Madrid, Spain
4 Department of Industrial Engineering, University of Salerno, Via Giovanni Paolo II, 84084 Fisciano, Italy;
ruggiero@unisa.it
* Correspondence: r.damato@upm.es

Abstract: Frequency-dependent adaptive noise cancellation-based tracking controller (ANC-TC) is a

known technique for the stabilization of several nonlinear dynamical systems. In recent years, this
control strategy has been introduced and applied for the stabilization of a flexible rotor supported
on full-lubricated journal bearings. This paper proposes a theoretical investigation, based on robust
immersion and invariance (I&I) theory, of a novel ANC-frequency estimation (FE) technique designed
to stabilize a flexible rotor shaft affected by self-generated sinusoidal disturbances, generalized to the
case of unknown frequency. A structural proof, under assumptions on closed-loop output signals,
Citation: Amato, G.; D’Amato, R.; shows that the sinusoidal disturbance rejection is exponential. Numerical simulations are presented
Ruggiero, A. Adaptive Rejection of a to validate the mathematical results in silico. The iterative Inexact Newton method is applied to
Sinusoidal Disturbance with
the disturbance frequency and phase estimation error point series. The data fitting confirms that
Unknown Frequency in a Flexible
the phase estimation succession has an exponential convergence behavior and that the asymptotical
Rotor with Lubricated Journal
frequency estimation is a warm-up phase of the overall close-loop disturbance estimation process.
Bearings. Mathematics 2022, 10, 1703.
In two different operating conditions, the orders of convergence obtained by phase and frequency
https://doi.org/10.3390/
math10101703
estimate timeseries are p ϕ = 1, pω,unc = 0.9983 and pω,cav = 1.005. Rejection of the rotor dynamic
disturbance occurs approximately 76% before in the cavitated than in the uncavitated condition, 2 (s)
Academic Editors: Higinio Rubio
and 8.5 (s), respectively.
Alonso, Alejandro Bustos Caballero,
Jesus Meneses Alonso and
Keywords: rotordynamic; adaptive rejection control; sinusoidal disturbance; flexible rotor;
Enrique Soriano-Heras
hydrodynamic journal bearing
Received: 15 April 2022
Accepted: 10 May 2022 MSC: 37M05; 37N35
Published: 16 May 2022

Publisher’s Note: MDPI stays neutral

with regard to jurisdictional claims in
published maps and institutional affil-
1. Introduction
iations. 1.1. Problem Statement: Flexible Rotor with Lubricated Journal Bearings
In rotating machines, bearings are used to transfer radial and axial forces in a sup-
porting structure ensuring a low value of the coefficient of friction and good system
stability [1,2]. For cases with high loads and high rotational speeds, fluid film lubricated
Copyright: © 2022 by the authors.
bearings are often preferred. The latter can be of two main categories: hydrodynamic
Licensee MDPI, Basel, Switzerland.
bearings and hydrostatic bearings [3]. In the case of hydrodynamic bearings, the relative
This article is an open access article
movement between the two coupled surfaces (journal and bearing) generates a pressurized
distributed under the terms and
fluid film that reduces the friction forces, ensuring the necessary support for the exter-
conditions of the Creative Commons
nal load [3]. In the case of hydrostatic bearings, the pressure of the lubricating fluid is
Attribution (CC BY) license (https://
guaranteed by the action of an external pump, which injects pressurized lubricant into the
creativecommons.org/licenses/by/
4.0/).
bearing [4]. The most common version of hydrodynamic bearings is the full journal bearing,

Mathematics 2022, 10, 1703. https://doi.org/10.3390/math10101703 135 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 1703

which consists of a rotor (journal) that rotates inside a bearing with a diameter slightly
larger than that of the rotor, and the fluid film exists in the small space between the two
(meatus) [3]. In the unsteady dynamical behavior, fluid film bearings can, however, present
oscillating behavior (typical elliptical orbits) due to destabilizing cross coupling forces
caused by the nonlinear fluid dynamic phenomena in the oil film [2,5]. To obtain acceptable
operating conditions, it is necessary to simultaneously analyze the rotor-journal bearing
couple in the case of a flexible rotor. These types of dynamic systems exhibit a particular
type of self-excited vibration due to fluid dynamics phenomena in the oil film [2] known as
oil whirl and oil whip, and are characterized by subsynchronous processional motions [6,7].
These vibrations appear when the subsynchronous vortex frequency reaches the natural
frequency of the system [8] and is typically characterized by high vibration amplitudes.
In recent years, research has shown that oil whirl phenomena are generated also when
the journal bearing runs on micropolar lubricant [9–11]. Many authors have studied the
effect of lubricant contamination [12] and the nonlinear behavior of film-oil main bearings
in rotating machines [13]. Furthermore, the performance of the bearing, in order to avoid
instability phenomena, has been analyzed, studied and simulated considering factors such
as misalignment [14], elasticity of the bearing liner [15], dynamic conditions and coupled
surface roughness [16].

1.2. Literary Review: Active Noise Rejection Control

One of the objectives of recent research has been to propose control strategies for
solving the classical problem of vibration rejection, in the frame of flexible mechanical
structures [17,18], flexible rotor bearings control [19–23] and their active balancing [21].
The aim is to absorb, or reject, any vibrational phenomena on a rotating shaft connected to
a motor, and stabilize it around an arbitrary equilibrium position [24], for example, using
a closed-loop controller with notch filtering actions [24–26]. Good attenuation features
are obtained if the disturbance frequency is well-known. Indeed, the stop bands of these
filtering-based controllers exhibit very steep edges. Usually, in the presence of unmodelled
dynamics, observer-based and sliding mode observer-based controls allow for the tracking
control precision to be improved by adapting to the unmatched uncertainties [18]. Never-
theless, the use of an adaptive observer to solve this category of problems, the vibration
suppression in flexible mechanical system, is not so widespread. For example [26], avoids
the application of an adaptive observer due to the risk of control spillover. Conversely, the
ANC control in D’Amato et al. [24] is implemented as an observer-based control. In [24],
the equations of the elastic contributions in the journal bearings and the relative vibrations
are modelled as a separated exosystem [27], which represents the disturbance acting on the
control. Then, the ANC operation consists of the observation of such disturbance elastic
contributions, so that the control injects, on purpose, a counterphase oscillation, aiming to
delete them. This makes the closed-loop control independent of the operating conditions of
the bearings: cavitated and uncavitated.
First results concerning controllability and observability problems in rotordynamic
systems date back to the 1980s. The flexible rotor and the flexibly-mounted journal bearings
were modelled by Stanway and Burrows, 1981 [22], as a spring-damping-mass system,
with well-known natural vibration modes. They determined the system observability
and controllability conditions in order to assign directly the closed-loop eigenvalues with
a linear controller. A linear control approach was also addressed by Lei and Palazzolo,
2008 [25]. To this aim they used the Finite-Element-Method (FEM) to model the rotor
dynamic system. In their approach, active magnetic bearings (AMB) were employed to
handle the rotordynamic systems by acting coaxially to the inertial axis through a linear
control law. In the framework of control techniques which uses AMB, Kumar and Tiwari,
2020 [28], investigated the rotordynamic system modeling unbalance and misalignment
with the rigid body theory. The possibility of modeling the flexible system with a new
set of nonlinear closed-form equations is a definite novelty introduced by D’Amato et al.,
2022 [24], and retrieved in this presented study. By considering the lubricated bearing

136
Mathematics 2022, 10, 1703

dynamic effects and with lubricant film cavitation conditions, an input–output feedback
linearizing controller can be implemented. The proposed control law is adaptive with
respect to the control and disturbance parameter (i.e., phase and frequency of the sinusoidal
disturbance, corresponding to rotor angular speed of the rotor).

1.3. Proposed Control Strategy: ANC with Frequency Estimation

This paper proposes a theoretical investigation of a novel ANC-frequency estimation
(FE) technique designed to stabilize a flexible rotor shaft affected by self-generated sinu-
soidal disturbances. This work extends the adaptive noise cancellation tracking control
(ANC-TC) algorithm presented by D’Amato et al. in 2022 [24,29], generalized to the case of
unknown frequency. The disturbance frequency corresponds to the rotor operating angular
speed (ω) [20,24,30], which is driven by an external actuator, so that uncertainties may arise
in the frequency actual value due to actuation operating point fluctuations. Other incoming
nonlinear phenomena, such as gyroscopic moments acting on the disk—for example due to
asymmetries of the rotor support [31]-can make the operating frequency vary. The ANC-TC
in [24,29] uses such a frequency value as a known parameter, which is also a parameter
of the dynamics of the system. With respect to the specific uncavitated case study [29], a
theoretical formulation has been provided in [24]: structural proofs for noise suppression
estimation, cancellation and system stabilization are given when the disturbance frequency
is known but not the initial condition. In this paper, inspired by immersion and invariance
(I&I) robust control techniques [18,32,33], the possibility of estimating online the constant
unknown frequency of the disturbance is considered. Following I&I-based approaches, the
studied controlled system is immersed into a target dynamical model so that its trajectories
are an image of the target dynamic system into the immersion map. Then, the aim of this
kind of approach is to make the target system image an attractive manifold, so that the
controlled system is made globally, uniformly, asymptotically stable around its invariant
trajectories [18,32,33].
The approach of [33] was convoluted in the hybrid logic (even if in a more gen-
eral sense with respect to [32]). The general multi-harmonic disturbance case may be
considered [17,33]. An I&I-based strategy, adaptive with respect to perturbation parame-
ters, was recently presented and applied to a reversible cold strip rolling mill to control
the speed and tension system [18], characterized by severe coupling effects, multiple state-
variables, nonlinearities and model uncertainties. In [18], it was shown that the estimation
errors follow an exponential convergence.
A recent research work [17] deals with the problem of active multi-harmonic dis-
turbances cancellation in flexible mechanical structures, under the analysis of passivity
properties of a closed-loop system. The controller operates an inner loop, which performs
the position control of the system, while the frequency estimation is demanded to an outer
control loop, elaborating the position measurements of the plant. A robust linear compen-
sator scheme, non-accurate model-based, was employed and experimentally validated. An
exponential cancellation of the disturbance was obtained.
The ANC-TC in [24,29] already showed an intrinsic robustness since the vibrational
disturbances acting on the rotor were cancelled. Consequently, the introduction of a
frequency identifier for vibrational modes improves the robustness of the method.
The closed-loop sinusoidal noise rejection is conceived in two phases: first, the operat-
ing frequency identification is performed as a combination of state-observer and asymp-
totical parameter estimation [32,33]; second, the asymptotical frequency estimate is fed
back to the closed-loop adaptive noise cancellation control (ANC), which hooks the phase
of the disturbance by injecting on rotor dynamics a counterphase sinusoid acceleration.
The frequency estimation is in practice a warm-up phase of the overall closed-loop con-
trol [34,35]. This study proposed a unified theoretical structure including an I&I-based
frequency estimator as a plug-in control block. The proposed control architecture design
concerns the externalization of the FE process, which appears as an additive standalone
block downstream to the ANC closed-loop, elaborating its output data and generating the

137
Mathematics 2022, 10, 1703

asymptotical frequency estimate. Once such an asymptotical estimate converges at the

steady-state closed-loop, the ANC control exponentially hooks the sinusoidal disturbance,
following [24].
Table 1 itemizes and catalogues the scientific literary references that are investigated
as background for the proposed study.

Table 1. Literary review.

Section Section
Authors Year Source Authors Year Source
Content Content
Poritsky 1953 [8] PS Avramov and Borysiuk 2012 [31] PCS
Stanway and Burrows 1981 [22] LR Harika et al. 2013 [12] PS
Burrows and Sahinkaya 1983 [19] LR Marino and Tomei 2016 [27] LR
Lund 1987 [2] PS Carnevale 2016 [33] PCS
Vance 1988 [30] PCS Salazar and Santos 2017 [4] PS
Muszynska 1988 [6] PS Zheng et al. 2017 [23] LR
D’Agostno et al. 2001 [7] PS Ruggiero et al. 2018 [20] LR, PCS
Zhou and Shi 2001 [21] LR Ballnus et al. 2018 [35] PCS
Hamrock and Schimd 2004 [3] PS Das and Guha 2019 [10] PS
Sukumaran Nair and
2004 [15] PS Bhattacharjee et al. 2019 [11] PS
Prabhakaran
Das et al. 2005 [14] PS D’Amato et al. 2019 [29] PCS
Ypma 2006 [36] PCS Marko et al. 2020 [26] LR
Prabhakaran Nair et al. 2007 [16] PS Kumar and Tiwari 2020 [28] LR
Lei and Palazzolo 2008 [25] LR Chen et al. 2021 [5] PS
Carnevale and Astolfi 2008 [32] PCS Tripathy and Bhattacharyya 2022 [9] PS
Friswell et al. 2010 [1] PS Liu et al. 2022 [18] LR, PCS
Hoad et al. 2010 [34] PCS Marko et al. 2022 [17] LR, PCS
Vania et al. 2012 [13] PS D’Amato et al. 2022 [24] PCS
PS: Problem Statement PCS: Proposed Control Strategy
LR: Literary Review
• flexible rotor with lubricated • adaptive noise
• flexible rotor bearings control;
journal bearings; cancellation-rejection;
• active rejection of vibrational phenomena;
• fluid dynamics; • immersion and invariance;
• notch filtering;
• self-excited vibrations due to • method comparison;
• observer-based control;
fluid dynamics; • numerical analysis methods;

A mathematical proposition for the ANC-FE is formulated and a structural proof

is given under assumptions on closed-loop output signals, showing that the sinusoidal
disturbance rejection is exponentially performed. Numerical simulations are presented
to validate the theoretical results in silico. The simulations are designed to illustrate
the adaptive rejection performance in terms of time and residual vibration amplitude.
Moreover, it is shown that the proposed I&I-based asymptotical frequency estimation may
have a practical application since its duration is acceptable as an initialization process. This
is a training/warm-up phase for the ANC controller. In addition, in order to evaluate the
order of convergence of the disturbance frequency and phase estimation error point series
the iterative Inexact Newton method [36] was applied.
The main novelties of this study are summarized:
i generalization of ANC-TC control devised in [24] to the unknown/uncertain operating
frequency;
ii application of I&I robust technique for frequency estimation (FE), considering as input
an analytical reconstruction of the closed-loop output noise signals;
iii plug-in integration of the FE module with the ANC-TC control;
iv formulation of a mathematical proposition for the novel ANC-FE control;
v structural proof of the proposition under assumptions on closed-loop output signals.
This paper has the following outline: Section 2 introduces the closed formulation of
the rotor dynamics model and ANC-FE control conceptualization and design; in Section 3
the theoretical proposition is formulated and its proof is given; Section 4 reports numerical
simulation results with a detailed description of the data fitting analysis; in Section 5 the
numerical simulation results are discussed and some conclusions are reported.

138
Mathematics 2022, 10, 1703

2. Materials and Methods

The dynamical system under investigation is composed by an unbalanced thin disk,
symmetrically disposed on a flexible shaft, and supported by two hydrodynamic full (short)
journal bearings, as represented in [24]. The mass of the rotor is 2 m and the flexible shaft
is a thin bar of a circular section with a negligible mass with respect to the disk mass.
According to Figure 1, during the motion, the disk remains in plane motion in x–y plane
without gyroscopic effects. The lubricated dynamical system can be studied by applying
the classical approach based on the Reynolds equation [20] in cylindrical coordinates and
under the classical hypothesis of short bearing approximation [30]. The scalar nonlinear
differential motion equations of the system are [20]
⎧ .. $ %
⎪
⎪ m x $+ k xc − % x j = muω
2 sinωt
⎪
⎪ ..
⎨ my + k yc − y j =−muω cosωt − W2
. . $ % (1a)
⎪ Fx x j , x J , y j , y J = −k xc − x j
⎪
⎪   $ %
⎪
⎩ . .
Fy x j , x J , y j , y J = −k yc − y j

where the forces Fx , Fy are defined in [24].

Figure 1. Journal bearing geometry.

2.1. Adaptive Noise Cancellation Tracking Control (ANC-TC)

In the following, for the lubricated dynamical system, the cavitated (π-film) and the
uncavitated short bearing model are considered. A closed-form model formulation for the
controlled model, which has been introduced in [24], is written as follows
⎧  .. 
⎪ .. x . $ % .. →
⎪
⎪ x (t) = .. c = − mk x − z j − dw − de (t, ω ) + u x
⎨ yc
. (    ) (1b)
⎪
⎪ . xj . 3 −1 $ % $ % xj xc
⎪
⎩ j z ( t, ω, i ) = . = ke l t, z j , e, i Σ t, z j , e, ω, i + 1
μ0
yj yj yc

$ %T
where: x = ( xc , yc ) T and z j = x j , y j denote the disc center and journal center coordinates,
→
respectively; dw = [0, W/m] T is a constant load term and u x is the control input. The
vector of sinusoidal disturbances acting on the two coordinate dynamics is denoted by
de (t, ω ) = u[sin(ωt + ϕ01 ), − cos(ωt + ϕ02 )] T (ϕ01,2 unknown initial phase), so that
   
. . 0 −ω u sin(ωt + ϕ01 )
de (t, ω ) = Ω · de (t, ω ) ⇒ de (t, ω ) = · (2)
ω 0 −u cos(ωt + ϕ02 )

139
Mathematics 2022, 10, 1703

The dynamics of acceleration noise in Equation (2) has the canonical parametrization
(Ω) of a disturbance with ellipsoidal trajectories, whose dimensions depend on the values
of the system parameters, as found by the numerical simulations of the analytical model
Equation (1a) presented in [24,29]. The ω is the frequency of the disturbance and corre-
sponds to the operating rotation speed of the flexible rotor. The two matrices l −1 and Σ are
defined as follows
*       , −1
$ % −1
$ % zx j −2ζπ (i−1) ey j −y j ex j ex j ey j
l −1 t, z j , e, i i · Λ t, z j , e, i = i + (1 − i ) · 8λ ·
zy j +2ζπ (i−1) ex j xj ey j −y j xj
    (3)
$ % − μ10 −ω ζk πi (i−1) 2λ t0
Σ t, z j , e, ω, i + (1 − i ) · k ω ke
ω ζk πi (i−1)
t0
− μ10 − ke 2λ
k ω

#
where: e = x2j + y2j denotes the orbit eccentricity; the i ∈ {1, 2} is the summary index,
which allows us to switch the model representation among uncavitated (i = 1) and cavitated
(i = 2) operating conditions, respectively. In [1], it was shown-analytically and by numerical
simulations- that the dynamical time-variant matrix Λ is always nonsingular with a high
determinant.
To reduce Equation (1a) into Equations (1b)–(3), the following analytical positions are
introduced in [24]

2πζ (1+2ε2 )
 3
z= (1− ε2 )
ζ= 1
3/2 μ0 = μR L
c
2(1− ε2 )
σ= e
1/2 λ = σζ (4)
( c2 − e2 )
ρ0
ρ0 = 2RLp0 ε= e
c t0 = μ0

where: k, m, μ, R, L, c, π, p0 , u are dynamical and geometrical parameters of the system

collected in Table 1 [24].
In the case of known operating frequency ω, in [24] the closed-loop ANC-TC algo-
→
rithm u x in Equations (5)–(7) was proposed for the flexible rotor Equations (1a) and (2),
and an exponential tracking of a desired trajectory for the disc center
. ..
→ → → . . .. ..
[ x r ; x r ; x r ] = [ xcr , ycr ; x cr , yc r ; x cr , yc r ] is obtained (the detailed formal proof is reported
in [24]). The ANC-TC includes two main parts: it allows both the sinusoidal disturbance
in closed-loop and the springback force contribution to be adaptively canceled due to the
always nonzero displacement between disc and journal (eccentricity). The overall vectorial
control law is reported in Equations (5)–(7) (the vector symbols are omitted for convenience)
⎧ ..
⎪
⎨ u x  û x0 + dˆe (t, ω )
û x0  mk x + dw + v̂ x (5)
⎪
⎩ .. .
v̂ x  − mk ẑ j + xr − k1 x+ − k2 x+
⎧ ..  
⎪
⎪ −usin(ωt + ϕ̂(t))
⎪
⎪ dˆ(t, ω )  ω 2
⎪
⎪     ∼ ucos(ωt +
  ϕ̂(t))    ..
⎨ -t k ϕ1 0 k ϕ1 0 . ∼

ϕ̂(t, ω )  0 φ(τ, ω ) · ω de (τ, ω ) dτ = φ(τ, ω )
2 · x+ + k1 x+ + k2 x dτ (6)
⎪
⎪ 0 k ϕ2  0 k ϕ2
⎪
⎪
⎪
⎪ cos(ωt + φ0 ) 0
⎩ φ(t, ω )  u
0 sin(ωt + φ0 )
* $ .%
ẑ j  − k η̂ + cη · x
m
. m .
. . ⇒ η̂ (0) =0 , ẑ j  − · c η · x (0) , (7)
η̂  −cη · v̂ x + cη · η̂ + c2η · x k
. . . .. .. ..
where: x+ = ( x − xr ), x+ = x − xr , x+ = x − xr are the tracking errors of position, velocity
and acceleration of disc center, respectively; dˆe (t, ω ) is the sinusoidal disturbance estimate;

140
Mathematics 2022, 10, 1703

ϕ̂(t) = [ ϕ̂1 , ϕ̂2 ] T is the phase estimate vector, which is responsible for matching at steady-
state the initial conditions of the noise (ϕ0 = [ ϕ0 , ϕ0 ] T = [ ϕ01 , ϕ02 ] T ) in (2), provided
that the amplitude (u) and frequency (ω) are known; φ0 = ϕ̂i (0) = rand(−2π, 2π ) is a
random value chosen as the (arbitrary) initial value for the phase estimation. According
.. .
to [24], ω 2 d+e = x+ + k1 x+ + k2 x+ is a known term, which has been modeled in Equation (6) in
terms of residual spurious dynamics, under the assumption that the asymptotic control
û x0 —linked to the estimation of the displacement Equation (7)—is already converged on
target value. Equation (6) is written according to a linear parametrization of the estimation
error d+e (t, ω ) = de − dˆe [37]: it appears as product of regressor Φ(t, ω ) and the parameter
estimation error, namely d+e = Φ T (t, ω ) · ϕ +(t) (see [24]).
The control gains k1 , k2 , k ϕ1 , k ϕ2 and cη are found by trial-and-error procedures re-
ported in Table 2.

Table 2. Simulation parameters.

Dynamical Initial Operating

Values Values Values
Parameters Conditions Parameters
m (kg) 1.5 xc (0) 0 f (Hz) 500
.
K (N/m) 4 × 106 xc (0) 0 ω (rad/s) 3141
u (m) 10−3 yc (0) −10−5 μ0 (m kg/s) 7.06 × 104
.
R (m) 1.6 × 10−2 yc (0) 0 ρ0 (N) 1 × 10−3
L (m) 1.6 × 10−2 xj (0) 0 t0 (s−1 ) 1.45 × 10−8
μ (kg/s) 3.4 × 10−2 yj (0) −10−5 p0 (bar) 2
c (m) 3.16 × 10−5 ϕ0 (rad) [5.2, 5.2]
g (m/s2 ) 9.81
π 180◦
Step 5 × 10−6
Time (s) 12

The phase estimator part of Equation (6) is devised since the main problem in sinu-
soidal noise rejection is the knowledge of the initial disturbance phase [27,38], so it acts as a
dynamic Phase-Locked-Loop (PLL [39]). The displacement estimation  . . part of Equation (7)
represents a reduced-order observer of the dynamical subsystem x c , yc , x j , y j of the over-
all system Equation (1b). It is required due to technical limitation to install a sensor to
measure the center position of the journal. For the same reasons, moreover, even if an
analytical value of z j was considered by Equation (1b), even in this case, the uncertainty on
the initial condition z j0 = z j (0) may have a fundamental role in the vibration attenuation
at steady-state.
. ..
Remark 1. It is noteworthy that only the disc center position (x), speed (x) and acceleration ( x)
measurements are required and fed back for control closed-loop adaptation Equation (7), as well as
for frequency estimation (Figure 2). ♦

2.2. Adaptive Noise Cancellation with Frequency Estimation (ANC-FE) Control

Now, even if the operating frequency was known, which is used as a parameter in the
ANC-TC [24], it can be affected by a percentage uncertainty due to possible miscalibration
of the actuator, which is keeping the rotor in rotation at ω speed. In practice, this makes the
operating frequency, and consequently the disturbance frequency, unknown. Moreover,
other incoming nonlinear phenomena, such as gyroscopic moments acting on the disk, for
example, due to asymmetries of the rotor support [31], can make the operating frequency
vary. For this scope, consider the problem of estimating the operating frequency of the
system. The following control scheme is adopted (Figure 2). The scheme recovers the
one adopted in [24], where the exosystem Equation (2) (hereafter EXO) acts as an exter-
nal dynamics. Downstream of the control loop, a further stage is added: the frequency

141
Mathematics 2022, 10, 1703

estimator (FE). This module elaborates the analytical output y an of the closed-loop system
Equations (1a)–(7) in loop with the frequency online estimate ω̂ (t). Eventually, the FE
module feeds back to the control loop the value of estimated frequency ωest , which is used
in the ANC-TC module as an input.

Figure 2. ANC-FE (adaptive noise cancellation with frequency estimation) [1].

2.2.1. ANC-FE Control Design with Measured Output Noise (y)

In the following, a mathematical formulation of the novel adaptive noise cancellation
control with frequency estimation (ANC-FE) is given. It is inspired by immersion and
invariance (I&I) approach (see [32,33]), which unifies the asymptotic parameter and state
estimation problems as reduced-order-observer-based problems (see Remarks 3–5). Let
some notations be introduced. From Equation (2)
 
. usin(ωt + ϕ0 )
de (t, ω ) = Ω
−ucos(ωt + ϕ0 )
⎡ . ⎤ ⎡ ⎤
.. .. *
  . (8)
⎢. ⎥ ⎢ . ⎥ .
y = dei = ε
⎢d ⎥(t, ω ) = Ω · ⎢dei ⎥ = Ω de1 ⇒ y(t, ω )dei (t, ω ) ⇒
⎣ ei ⎦ ⎣ ⎦ de2
.
ε = −ςy = −ω 2 y
.. ..
. .
First, the signal y can be regarded as a (measured) scalar output of the closed-loop
system Equations (1a)–(7) (Figure 1); ς = ω 2 , u and ϕ0 are related to the initial states
y(0) = y(t = 0) and (0) = (t = 0).

Remark 2. The following computation is performed for the only scalar variable dei first since the
two components’ analyses are analogous and since the two disturbance components are shifted of
π/2, so that the estimation is shared among the two components [24]. For convention, let the first
component y = de1 = u sin(ωt + ϕ0 ) be used. ♦

Consistently with the notation used in Equations (5)–(7) (see [24]), by calling ˆ and
ς̂ the online state and parameter estimates of the disturbance, respectively, the following
injected scalar estimation errors [18] are defined as follows
&   '
. .
z (y) = k  y + ˆ k2 + ς −  = β  (y) + (α ) ˆ −  ;zς (y) = y ˆ + { ς̂ − ς } = β ς (y, ˆ ) + ς+ (9)

where k  > 0 is an arbitrary control parameter reported in Table 2. The two errors [z , zς ] in
 = ˆ −  and ς+ = ς̂ − ς. They contain
Equation (9) differ from the classical error definition +

142
Mathematics 2022, 10, 1703

$ %
two injection terms [β  (y), β ς (y, ˆ )] acting as a dynamical correction; α = k2 + ς > 0
acts a kind of scaling. The Equation (9) is written as follows

 = −z (y) + ˆ (α ) + β  (y); ς = −zς (y) + ς̂ + β ς (y, ˆ ). (10)

By setting at zero the estimation errors z (y), zς (y) in Equations (9) and (10), it follows that
 
est = k2 + ς est ˆ + k  y;ς est = ς̂ + y ,
ˆ (11)
$ %
where, est (0) = k2 + ς est (0) ˆ (0) + k  y(0) and ς est (0) = ς̂(0) + y(0)ˆ (0).
The two asymptotic estimates [est , ς est ] are found as with a PI (proportional integral)
law with respect to the measured output y. Indeed, both (α ) ˆ and ς̂ turn out to be an
integration process, α denotes the gain of the integral part, while both the two injection
terms represent the proportional part. To prove this point, derive in time the errors [z , zς ](y)
in Equation (9). From Equation (8) it follows that
. . . . .
z (y) = (α )ˆ + ςy + k  ;zς (y) = ς̂ +  ˆ + y ,
ˆ (12)

while, replacing Equation (10) in Equation (12), it follows

. $ % . $ %
z = k2 + ς ˆ + ςy + k  −z + k2 ˆ + ˆ ς + k  y &
$ 2 % . ' (13a)
= −k  z + k  + ς y + k  ˆ + ˆ ;
. . .  
& .zς = ς̂ +. y ˆ + ˆ −z + k  ˆ + k  y + !' ˆ (−zς + ς̂ + y ˆ )
2
& . ' (13b)
= −ˆ z − ˆ 2 zς + ς̂ − −y ˆ − k2 ˆ 2 − k  yˆ − ˆ 2 ς̂ − y ˆ 3 = −ˆ z − ˆ 2 zς + ς̂ − Δ(y, ,
ˆ ς̂) .

The term Δ(y, ,ˆ ς̂) is a function of known signals. By setting at zero the quantities
between braces in both Equations (13a) and (13b), the two scalar adaptation laws for
(,
ˆ ς̂)(t) are found as follows
. . . !
ˆ = −y − k  ,ˆ ς̂ = − y ˆ + k2 ˆ 2 + k  yˆ + ˆ 2 ς̂ + y ˆ 3 , (14)

where, from Equations (6) and (8), ˆ (0) = u ω0 cos(ω0 t + ϕ̂i (0)) and ς̂(0) = ω02 .
Consequently, Equations (13a) and (13b) become
 .     
z −k  0 z
. = · . (15)
zς −ˆ −ˆ 2 zς

The first-row of Equation (15) makes that z → 0 exponentially. However, this is not
sufficient to construct an estimation of the state variable  since the overall convergence
of the estimation is linked to the convergence of all the error dynamics [z , zς ]. However,
! . .
ˆ ς̂ = [0, 0],
once the two adaptation laws have reached their steady-state values, namely ,
the asymptotic estimates for [est , ς est ] = [, ω 2 ] are given by Equation (11) (see the proof in
Section 3). Note that ω0 may be the rough initial estimate of the operating rotor frequency,
taken from the rotor actuator datasheet as the steady-state operation value. As mentioned
above, the estimates of Equation (11) are two PI, with integral parts (α ) ˆ and ς̂, as may be
inferred from Equation (14).

2.2.2. ANC-FE Control Design with Analytical Output Noise (y an )

Regarding the assumption about y, which in Equations (8)–(14) is regarded as the
measured output of the closed-loop system Equations (1a)–(7), it requires a sensor in output
able to distinguish and provide only the sinusoidal vibration measurement (de (t, ω )). First,
it is a logic short-circuit since once the disturbance measurement is available a direct
cancellation could be viable. Second, the sensor in turn may be affected by measurement

143
Mathematics 2022, 10, 1703

noise, which is source of uncertainty. Then, the ANC-FE control Equations (8)–(14) with
measured output noise (y(t, ω )) cannot be directly implemented in closed-loop. For this
scope, the value of the output noise y, which is used as a known quantity in the ANC-FE
Equations (8)–(14), is reconstructed analytically from Equation (1b). By using position,
 .. .  T
speed and acceleration measurements of the disc center coordinates ( x, x, x ) and an
analytical chasing of the journal center displacement (z J (ω0 )), the analytical output noise
(y an ) is computed from Equations (1b)–(3) as follows
( )
$ % .. k$ % →
y an t, z j , ω̂ − x (t) + x − z j (t, ω̂ (t)) + dw − u x (16)
m
. t $
(
% $ %
)
1
z j (t, ω̂ (t), i )z j0 + ke3 l −1 τ, z j , e, i Σ τ, z j , e, ω̂ (τ ), i z j (τ, ω̂ (τ )) + x (τ ) dτ (17)
0 μ0
where: z j0 is the arbitrary initial condition of the analytical chasing z j (t, ω̂ (t), i ) of the
journal
$ displacement
% (z j ); z j (t, ω, i ) is a function of the frequency through the matrix
Σ t, z j , e, ω, i in Equation (3); e is the eccentricity computed in the analytical chasing
coordinates (z j ). From Equations (1b) and (16), it follows

⎡ ⎤
..
$ % ⎢ .
⎥ . .. k  .. $ % $ %
y an t, z j , ω̂ = ⎢ ⎥
⎣ y ani ⎦ = de (t, ω ) + m −z j (t, ω ) + z j (t, ω̂ (t)) = de (t, ω ) + δy t, z j0 , ω, ω̂ (t) ≡ y an t, z j , ω, ω̂ (18)
..
.
.. ..
where from (8) dei (ω, t) ∝ dei = y, so that y ani ∝ dei (ω, t). The constant proportional
factor is the square of unknown frequency −ω 2 , which, from (11), is absorbed in the
tuning procedure of k  . Then, y an and the bounded disturbance de (t, ω ) are consistent
since, eventually, the
$ two sinusoidal
% vibrations have the same frequency (ω). Indeed, from
Equation (18), y an t, z j , ω, ω̂ is a function of ω. We now state the main assumption in
this paper.
$ %
Assumption 1. The analytical chasing error δy t, z j0 , ω, ω̂ (t) in Equation (18) is bounded and
$ % .
small enough that y an t, z j , ω̂ is assumed at the same frequency of de (t, ω ) = Ω de (t, ω ). 
From Equations (8) and (16)–(18), according to Remark 2, and under the Assump-
tion 1, the asymptotic adaptation laws Equations (11)–(14) are updated considering the
computation for only one component y ani , as follows
$ %
est = k2 + ς est ˆ + k  y ani ; !
. . .
ς est = ς̂ + y ani ;
ˆ ˆ = −y ani − k  ;
ˆ ς̂ = − y ani ˆ + k2 ˆ 2 + k  y ani ˆ + ˆ 2 ς̂ + y ani ˆ 3 ,
#
ωest = lim ω̂ (t) = lim ς est (t) (19)
t→∞ t→∞
$ %
with est (0) = k2 + ς est (0) ˆ (0) + k  y ani (0); ς est (0) = ς̂(0) + y ani (0)ˆ (0);
ˆ (0) = u ω0 cos(ω0 t + ϕ̂i (0)) ; ς̂(0) = ω02 .
Equations (16), (17), and (19), constitute the frequency estimator (FE) equations.
Equation (15) holds as function of y ani .
By absorbing Equations (16), (17) and (19) in the ANC-TC Equations $ %(5)–(7), the
ANC-FE control, Equations (19)–(20c) with analytical output noise (y an t, z j , ω̂ ), is imple-
mentable in closed-loop as follows
⎧ ..
⎪
⎨ û x  û x0 + dˆe (t, ωest )
û x0  mk x + dw + v̂ x (20a)
⎪
⎩ .. .
v̂ x  − mk ẑ j + xr − k1 x+ − k2 x+

144
Mathematics 2022, 10, 1703

⎧ ..  
⎪
⎪ −u sin(ωest t + ϕ̂(t))
⎪
⎪ dˆe (t, ωest )  ωest 2
⎪
⎪   u cos(ωest t + ϕ̂(t))  
⎨ -t ∼
.. .
∼ ∼
k ϕ1 0
ϕ̂(t, ωest )  0 ϕ(τ, ωest ) · x + k1 x + k2 x dτ (20b)
⎪
⎪  0 k ϕ2 
⎪
⎪
⎪
⎪ cos(ωest t + φ0 ) 0
⎩ ϕ(t, ωest )  u
0 sin(ωest t + φ0 )
* $ .%
ẑ j  − mk η̂ + cη · x . m .
. . ⇒ η̂ (0) = 0, ẑ j  − · cη · x (0) , (20c)
η̂  −cη · v̂ x + cη · η̂ + c2η · x k

where, in Equations (20a)–(20c), the asymptotic estimate ωest = ω̂ (t)t→∞ (Equation (19)) is
considered for the constant ω as the constant frequency parameter in the phase estimator
(compare Equations (5)–(7) with Equations (20a)–(20c).
At steady-state, when all the estimates’ terms are converged (i.e.,
v̂ x → v x , ẑ j → z j , dˆe → de ), the closed-loop system, Equations (1b)–(4) and (19)–(20c),
is regulated by û x → u x around the reference xr (Figure 2).

Remark 3. The idea of frequency estimation, namely Equations (8)–(15), inspired from immersion
and invariance (I&I) approach, are borrowed from the algorithm proposed in [32]. Nevertheless,
the novelty of applying this technique to a rotordynamic model Equation (1a) [24,30] consists of a
further manipulation. Equations (16)–(19) have been introduced with the aim of reconstructing
analytically (y an ) an unavailable signal, the output noise (y)). The signal chasing Equation (17)
is possible by virtue of the closed-form formulation, Equations (1b)–(4), presented in [24], which
allows a reliable numerical extraction of the coordinates of the journal disc center to be performed,
with a very high determinant value of the matrix Λ, by integration of Equation (1b).
In particular, the system Equation (8) of dimension 3 [ς, , y ani ] is immersed in a wider space
ˆ y ani ]. Then, the observer problem is attracted by the surface at z = 0
with 5 dimensions [ς, ς̂, , ,
and zς = 0, where it remains at steady-state, so that the resultant invariant locus is a manifold
(Figure 3). ♦

Remark 4. It is worth highlighting that Equation (20c) is identical to Equation (7), namely, the
reduced-order-observer, which estimates the z j displacement, acts independently from the other
control blocks. The FE controller designed in this paper (Equations (16)–(19)) based on the analytical
output y an results as a plug-in estimation block downstream to the ANC-TC Equations (5)–(7)
proposed in [24]. This property shall be exploited for control convergence proof. ♦

Remark 5. The control ANC-FE Equations (19) and (20a)–(20c) are a general-purpose control with
respect to the actual lubricated rotordynamics’ operating condition,$namely cavitated/uncavitated.
% $ Indeed,
%
according to the notations in Equations (1b)–(4), the matrices l −1 t, z j , e, i and Σ t, z j , e, ω̂ (t), i in
Equation (17) are functions of the parameter i. ♦

145
Mathematics 2022, 10, 1703

Figure 3. Immersion and Invariance application on analytical output noise (y ani ).

3. Closed-Loop ANC-FE Control Convergence Proof

3.1. Proposition Statement
In this section a formal convergence proof of the proposed closed-loop ANC-FE control
Equations (19) and (20a)–(20c) is provided through the following proposition statement.

Proposition 1. Consider the lubricated rotordynamic model Equations (1b)–(4) and y = dei (t, ω )
Equation (8) as the sinusoidal output noise from the exosystem (2), with unknown constant frequency
→
( ω) and initial phase (ϕ0 ). Consider the reference dynamics x r = [ xcr , ycr ] for the disc center
coordinates, with x+ = ( x − xr ). $ %
From Equations (9) and (16)–(20c), ϕ + = ( ϕ0 − ϕ̂(t)), + z j = z j − ẑ j , z (t, y ani ) =
$ $ 2 % % $ %
−est + k  + ς est ˆ + k  y ani and zς (t, y ani ) = −ω̂ 2 (t) + ς̂ + y ani ˆ .
Under the Assumption 1, the û x control in Equations (20a)–(20c) makes that for the closed-loop
system Equations (1b)–(4) and (19)–(20c):
i. the equilibrium point (z (y ani ), ˆ (y ani )zς (y ani )) = (0, 0) is asymptotically stable;
$ %
ii. est = k2 + ς est ˆ + k  y ani . is an asymptotic estimate of the analytical output noise deriva-
. /
tive y ani and ωest = lim ω̂ (t) = ς̂ + y ani ˆ is an asymptotic estimate of the rotor
t→∞
operating
$ % frequency ω; 
iii. ϕ z j , x+ = (0, 0, 0) is exponentially stable (thesis from [24]). 
+, +

3.2. Proof
Let V1 (t, z , zς ) be a suitable Lyapunov function

z2 z2ς
V1 (t, z , zς ) = + ≥ 0. (21)
2 k 2

Deriving in time Equation (21), from Equation (15)

  
.   1z 1
V 1 (t, z , zς ) = −z2 − z
ˆ ς z − ˆ 2 z2ς = − z z
ˆ ς 2 ≤ 0, ∀(z , zˆ ς ). (22)
1
2 1 z
ˆ ς
.
Due to V 1 ≤ 0, V1 (t, z , zς ) is always nonincreasing in t ∈ (0, ∞), so that V1 is bounded
.
and, consequently, (z , zς ) are bounded. By deriving in time V 1 Equation (22) and from
Equations (15) and (19), it follows

.. . . . . $.% .
V 1 (t, z , zς ) = −2z z − zˆ ς z − ˆ zς z − z ˆ ς z − ˆ 2 z2ς − ˆ 2 z2ς
$ % (23)
= 2k  z2 − zς z (−y ani − k  ˆ ) − z ˆ  − $ ˆ  − ˆ 2 zς %+ k  z
z ˆ ς z
−2z
ˆ 2ς (−y ani − k  ˆ ) − 2ˆ 2 zς −z ˆ  − ˆ 2 zς .

146
Mathematics 2022, 10, 1703

..
The V 1 is bounded since under the Assumption 1 the analytical output noise y ani
is bounded and the estimate ˆ (s) = [−y ani (s)/(s + k  )] is a passive filter of a bounded
.
signal, so it is bounded and also its derivative ˆ = (−y ani − k  ˆ ) is bounded in turn.
. .
Then, B.2.1 Barbalat’s Lemma [40] applies on V 1 , so that lim V 1 (t) = 0. Consequently,
t→∞
ˆ ς ) → (0, 0) globally asymptotically.
(z , z $ % Now, according to [32] and from
Equations (10), (13) and (19), est = k2 + ς est ˆ + k  y ani is an
/asymptotic estimate of ana-
.
lytical output noise derivative y ani , while ωest = ω̂ (t)t→∞ = ς̂ + y ani ˆ is an asymptotic
estimate of the rotor operating frequency ω. This proves the theses (i)–(ii).
The rest of the proof is inherited from [24].
Let V2 (t, ϕ
+) and V3 (t, η+) be introduced as two suitable Lyapunov functions

1 T
+) =
V2 (t, ϕ + ·ϕ
ϕ + ≥ 0, (24)
2
1 T
V3 (t, η+) = η+ · η+ ≥ 0. (25)
2
Deriving V2 (t) Equation (24) in time, from Equations (6) and (20b) it follows

.
     !
. k ϕ1 0
V 2 (t) = ϕ + = −ϕ
+T ϕ +T Φ(t, ωest ) ω 2 d+e (t, ωest ) = −d+eT ω 2 k ϕ I(2×2) d+e ≤ 0, (26)
0 k ϕ2
where: k ϕ1 = k ϕ2 = k ϕ > 0; I(2×2) is the 2-by-2 identity matrix; the linear parametrization of
the estimation error in Equation (6) has been considered, i.e., d+e (t, ωest ) = Φ T (t, ωest ) ϕ
+(t, ωest )
(where the ωest has been considered constant at steady-state so that it is replaced for the actual
.
constant ω ). Due to V 2 ≤ 0, V2 (t) does not increase in t ∈ (0, ∞). Then, V2 is bounded
  and .
+ is bounded. By time-deriving V2 again, under the assumption of bounded Φ, Φ , the
so ϕ
. .
Barbalat’s Lemma in [40] applies on V 2 . Then, lim V 2 (t) = 0 and d+e → 0 ; consequently,
t→∞
+ → 0 asymptotically. To prove the exponential convergence of ϕ
ϕ +(t, ωest ) estimation error the
Persistency of Excitation condition [40] is required, provided that ∃ T, k T ∈ R+ such that
. t+ T
Φ(τ, ωest )Φ T (τ, ωest ) dτ ≥ k T I(2×2) > 0 , ∀t ≥ 0 , (27)
t

which is always trivially fulfilled, given the regressor definition in Equations (6) and (20b).
.
Now, consider, from Equations (7) and (20c), the following definitions of η+,η+, v+x , u+x
$ % $ . % $ . %
− mk z j − ẑ j = cη x + η − cη x + η̂ η+
. . .
η − η̂ = (−cv x + cη ) − (−cv̂ x + cη̂ ) = −c+ v x + c+ η η+ ,
(28)
v x − v̂ x = (u x − û x ) + ω 2 d+e = u+x + ω 2 d+e v+x
z j − ω 2 d+e u+x
− mk +

with cη < 0 (see [24] for detailed derivation of Equation (28)). Deriving V3 (t) Equation (25)
in time, from (28) it follows

. .   $ % 2
V 3 (t) = η+T · η+ = c · η+T η+ + (−c) · η+T u+x + ω 2 d+e ≤ 2c · η+2 + − 2c u+x  + ω 2 d+e 
$ % $ c%
= 2c · η+2 + − 2c ω 2 d+e 2 + − − k+ z j − ω 2 d+e 
2 + (− c ) − k + 2 + 2+
m z j − ω de  ω de  
$ % $ % 2 k m (29)
c 2
≤ · η+ + − ω de  −
c 2 + 2 c
− + 2 +
z j  + ω de   − + k
z j  + ω de  + 2ω 2 d+e 
2 +
2 2 2 m m
= −2c ω 2 d+e 2 − 2cη+ω 2 d+e  ≤ cη+2 − c ω 4 d+e 2 .
 
Equation (29) is obtained considering the triangular inequalities η+T u+x + ω 2 d+e ≤
    2
 2
2 η
1 +2 + u+x + ω 2 d+e 2 and  − mk +
z j + (−ω 2 d+e ) ≤  − mk +
z j  +  − ω 2 d+e  ; then,

147
Mathematics 2022, 10, 1703

 
the square of ξ (t)  η+ + ω 2 d+e  has been reconstructed, so that the inequalities
 
η+ω 2 d+e  ≤ − 12 η+2 + ω 2 d+e 2 have been used.
Following this, d+e (t) is bounded on [0, ∞) so that, according to Lemma A.1 in [41], any
globally exponentially convergent observer for the modal disturbance d+e (t) guarantees the
exponential convergence for η+ → 0 and, consequently, for +z j → 0 and u+x = (u x − û x ) → 0
(see Equation (28)). This is enforced by the global exponential zero convergence of d+e (t), as
has been proved in Equations (26) and (27).
Eventually, by replacing in Equation (1a) the control û x Equation (20a) (û x for u x ), it
follows that for u+x → 0
.. . .. . .
k
x+ = −k1 x+ − k2 x+ + z − d+e = −k1 x+ − k2 x+ − u+x = −k1 x+ − k2 x+
+ (30)
m j
is a Hurwitz system dynamics. Hence, also x+ → 0 globally exponentially. This proves the
thesis (iii). 

Remark 6. It has been proved that, under the Assumption 1, the convergence of$ asymptotical %
frequency estimate FE in Equation (19)—with its additional inner loop (ω̂ (t), y an t, z j , ω, ω̂ ) )
in which the FE is elaborated—is standalone with respect to the rest of the control loop: it is a
plug-in estimation block, according to Remark 2. In this case, once ωest → ω , at steady-state ωest
is considered as an internal parameter for the control loop. Then, the frequency estimation phase
is a warm-up process for the ANC-TC Equations (5)–(7) [34,35], which through Equation (19) is
transformed into Equations (20a)–(20c). ♦

Remark 7. From [24,32], as aforementioned at the beginning of Section 2.2.1, the closed-loop system
from the input disturbance de (t) to the output x (t), including the phase estimator
Equation (20b), the journal displacement observer Equation (20c) and the asymptotic frequency
estimator Equation (19), results in a regulation loop unifying the parameter and state estimation
problems as reduced-order observed-based problems, with adaptive notch filtering
characteristics [24]. ♦

4. Results
4.1. Numerical Simulation Setup and Method Description
The proposed control is implemented in the Simulink MATLAB environment with
the parameter set as shown in Table 2 and with the control parameters reported in Table 3.
The results are presented in four simulation cases for cavitated and uncavitated operating
conditions with measured and analytical output noise.
The subscript “i” in the notation of the analytical output noise y ani is omitted for
editing convenience of the graphs. Hence, y an in all figures is referred to as the only one
component y ani , accordingly to Equation (16) and Remark 2.
..
The signal dei (t, ω ) and z j in the figures represent the actual sinusoidal disturbance
and the journal center vector coordinates, respectively, whose values are computed from
..
the simulated model differential equations. These signals, (dei (t, ω ), z j ) would not be
available in practice, so their analytical chasing in simulation is reported only for the sake
of comparison with respect to the analytical signals.

148
Mathematics 2022, 10, 1703

Table 3. Control Parameters.

Control Initial
Values Values
Gains Conditions
cη [s−1 ] −400 x̂j (0) 0
^
k1 , k2 2000 yj (0) 0
^
kϕ1 50.66 ϕ(0) (rad) [4, 4]
kϕ2 50.66 ω0 (rad/s) 2985
k 7.5 zJ0 (m) [0, −2 × 10−5 ]

Bias and initialization issues can mislead the estimated response measure of perfor-
mance obtained from a simulation. Several techniques surveyed in [34] can be used to
estimate online the length of the warm-up phase in the output data collection of a simu-
lation model, to understand, automatically, when this preliminary phase ends and when
the actual numerical simulation starts giving a reliable response. The same issue is tackled
in Bayesian parameter estimation [35] where the parameters of the model under study
are usually derived from the available data, using optimization and sampling method
estimations. It is shown that a warm-up phase occurs (the asymptotical frequency estima-
tion phase) before the closed-loop ANC control can start its exponential hooking of the
exosystem output disturbance signal. In practice, such a length cannot be quantified, but
for the proposed technique, the warm-up phase length estimation is not required.
The simulations are designed in order to validate numerically the theoretical analytical
proof reported in the previous Section 3. In practice, the settling time and trajectory
behavior of the phase estimate can be forecast since its trend is exponential, properly,
while this does not occur for the frequency estimate, which converges asymptotically.
Hence, to highlight the differences in the convergence behaviors of the estimation errors, a
further numerical evidence is provided. The Inexact Newton method [36] is applied to the
frequency estimation error and phase estimation error successions, while the convergence of
the overall disturbance error succession is considered only in the time since it follows from
the previous two errors. The iterative Newton method is applied to the error successions in
both cavitated and uncavitated conditions, and only in the case of analytical output noise,
which is more interesting in practice.
To analyze and evaluate the order of the convergence of the estimation errors, the
following point series are defined as the difference between the estimation dynamics terms
and the constant target values of each series (ω and ϕ0 ). Denoting the successions of
frequency and phase estimation with ωest,yann n and ϕ̂−,yann n , respectively, sampled at
discrete time instants n (1 sample each 5 (ms)), the following

eωn = ωest,yann n
− ω e ϕn = ϕ̂−,yann n
− ϕ0 (31)

represent the successions of the frequency and phase estimation errors, respectively. Consid-
ering the absolute value of the ratio between the {n + 1}-element of each error succession
and the {n}-element power of pω,ϕ , as long as the limits of these ratios are constant for
n → ∞ , namely
   
ωest,yan − ω   ϕ̂−,yan − ϕ0 
n +1 n +1
lim   = constant lim   = constant (32)
n→∞  ωest,yan − ω  pω n→∞  ϕ̂−,yan − ϕ  p ϕ
n 0 n

then, each succession converges with at least a polynomial behavior.

We can rewrite Equation (32) as follows
       
ωest,yan − ω  ≤ k ω · ωest,yan − ω  pω  ϕ̂−,yan − ϕ0 ≤ k ϕ ·  ϕ̂−,yan − ϕ0  p ϕ (33)
n +1 n n +1 n

149
Mathematics 2022, 10, 1703

$ %
parametrized by the coefficient pair k ω,ϕ , pω,ϕ . In Equation (33): pω,ϕ represents the
order of convergence for each error succession, Equations (31) and (32), while k ω,ϕ is a
proportional factor, denoting the exponential convergence time-constant (when pω,ϕ = 1).
The following numerical simulations are proposed to: (1a) evaluate the convergence
performance of disturbance estimation errors in time; (2) verify that the initial frequency
estimation warm-up phase length is acceptable in practice in a real application (namely the
asymptotical convergence occurs at t < ∞). In all four simulation scenarios, 5% of initial
frequency estimation error and about 25% (about 70 deg) of initial phase estimation error
are considered.

4.2. Figure Descriptions

Figures 4 and 5 illustrate the frequency estimation performance in time, in the case of
the cavitated condition (Figure 4) and in the case of the uncavitated condition (Figure 5).
Three curves are plotted: ω denotes the real constant operating frequency; ωest,yan denotes
the asymptotical frequency estimate with the analytical output noise; ωest,y denotes the
asymptotical frequency estimate in the case of measured output noise.

Figure 4. Frequency estimation performance in the case of cavitated condition.

150
Mathematics 2022, 10, 1703

Figure 5. Frequency estimation performance in the case of uncavitated condition.

In Figure 6 the frequency estimation performance in time, considering the analytical

output noise y an , is compared in uncavitated and cavitated conditions.

Figure 6. Comparison of frequency estimation performance in uncavitated and cavitated conditions

with the analytical output noise y an .

151
Mathematics 2022, 10, 1703

Figure 7 shows the comparison (in offset view) of phase estimation performance
in time, in the four simulation cases. The ( ϕ̂unc,yan , ϕ̂unc,y ) represent the phase estimate
for the uncavitated case, with analytical and measured output noise, respectively; the
( ϕ̂cav,yan , ϕ̂cav,y ) represent the phase estimate for the cavitated case, with analytical and
measured output noise, respectively. In the graph, the warm-up phase edges, and time
interval labels, are marked-up in the four cases for illustration purposes.

Figure 7. Offset view of phase estimation in the four simulation cases, with warm-up phase edges.

Figure 8 shows
 the  scatter plot of the (k, p)-parametrized successions of frequency esti-
mation error eωn+1  = k ω |eωn | pω (Figure 8A,B) and phase estimation error
    pϕ
e ϕ  = k ϕ e ϕ  (Figure 8C,D), following Equations (31)–(33). Both cavitated and
n +1 n
uncavitated conditions with analytical output noise y an are considered. The timeseries in
Figure 8A,B are plotted considering the overall simulation time data. To evaluate whether
the phase estimates’ convergent behavior is exponential starting from the warm-up phase
end (according to Proposition 1), the timeseries in Figure 8C,D are plotted on the two
restricted simulation time intervals starting from 8.41 (s) and 1.59 (s) for uncavitated and
cavitated cases, respectively, as shown in Figure 7.
The text boxes in Figure 8 report the fitting data statistics obtained applying the
iterative Inexact Newton method on the frequency and phase estimation error timeseries,
in cavitated and uncavitated conditions with analytical output noise. Such statistics are
used to evaluate the order of the convergence pω,ϕ in the four cases. The arrows indicate
the direction of the succession point distribution from eωn=0 to eωn→∞ , and from e ϕn=0 to
e ϕn→∞ , as defined in Equation (31).
Figure 9 shows the difference between the scalar analytical noise y an and the actual
..
disturbance component dei (t, ω ) for the uncavitated case (Figure 9A) and for the cavitated
case (Figure 9C). For ease of graph interpretation, due to the high frequency of the signals,
a restricted time lapse (0–0.0012) (s) was chosen, arbitrarily, on the overall simulation time
of 12 (s), in order to highlight the detail of the comparison over a few periods.

152
Mathematics 2022, 10, 1703

Figure 8. Convergence of frequency estimation error timeseries (A,B) and phase estimation error
timeseries (C,D) for uncavitated and cavitated condition, with analytical output noise.

..
Figure 9. Scalar analytical output noise y an and actual disturbance component dei (t, ω ) for uncavitated
and cavitated conditions (A,C); “i” component of actual journal displacement z ji and its analytically
reconstruction z J i with the scaled difference δyi for uncavitated and cavitated conditions (B,D).

153
Mathematics 2022, 10, 1703

In Figure 9B,D, the behavior of only one component of the journal displacement
coordinate vector, z ji , is shown in time. It is compared with the same component “i” of its
analytically reconstructed value z J i . Both Figure 9B,D contain a subplot showing the scaled
$ %
difference δyi = − mk z ji − z J i , where
$ δyi is the acceleration
% injected error referred to as
only one component of the vector δy t, z j0 , ω, ω̂ (t) , as defined in Equation (18).
Figure 10 shows the journal displacement observation errors + z j as defined according
 T
to Equations (20c) and (28), in the four simulation cases. The signals z j = x j , y j are
not available in practice, but they are shown only for comparison to their corresponding
 T
observation ẑ j = x̂ j , ŷ j .

Figure 10. Observation errors of the journal displacement coordinate +

z j with analytical output noise
for uncavitated and cavitated conditions (A,C) and with measured output noise for uncavitated and
cavitated conditions (B,D).
..
Figure 11 illustrates the hooking of the sinusoidal disturbance dei (t, ω ) by the estimate
..
dˆei (t, ωest ) in the case of the uncavitated condition with analytical noise. For ease of graph
interpretation, due to the high frequency of the signals, the signal hooking is shown on the
restricted time interval (7.5–8.5) (s).

154
Mathematics 2022, 10, 1703

.. ..
Figure 11. Hooking of the sinusoidal disturbance dei (t, ω ) by its estimate dˆei (t, ωest ) in the case of
uncavitated condition with analytical disturbance (y an ).
..
Figure 12 illustrates the sinusoidal disturbance estimation errors d+ei (t, ω ) in the four
simulation cases.

..
Figure 12. Sinusoidal disturbance estimation errors d+ei (t, ω ) are reported for uncavitated and cav-
itated conditions with analytical output noise (A,C) and for uncavitated and cavitated conditions
with measured output noise (B,D).

In Figure 13, the control vector û x = [û x,1 , û x,2 ] T (Equation (20a)) is plotted in the four
simulation cases. A zoomed plot, referred to by the dashed rectangle, highlights the detail
of the last transient variation in the control before the steady-state behavior.

155
Mathematics 2022, 10, 1703

Figure 13. Control vector û x = [û x,1 , û x,2 ] T for uncavitated and cavitated conditions with analytical
output noise (A,C) and for uncavitated and cavitated conditions with measured output noise (B,D).

Figure 14 illustrates the disturbance attenuation on disc center coordinates [xc , yc ] in

the four simulation cases.

Figure 14. Disturbance attenuation of disc center coordinates (xc , yc ) for uncavitated and cavitated
conditions with analytical output noise (A,C) and for uncavitated and cavitated conditions with
measured output noise (B,D).

156
Mathematics 2022, 10, 1703

5. Discussions and Conclusions

5.1. Discussions
The closed-loop ANC-FE operates on the output noise date to compute the asymptoti-
cal frequency estimate. In the following simulative analysis, both measured and analytical
noise reconstruction is considered (in two cavitated and uncavitated operating conditions).
Due to the reconstruction error δy (Figure 9), the analytical output noise is affected by an
intrinsic additive disturbance, which entails a data quality lowering with respect to the
measured output noise case. The latter is considered an ideal case, and it is used in this
analysis only for the sake of comparison.
The control performance, in terms of settling time, is evaluated in four simulation cases.
From Figure 4 it emerges that, from the comparison between the case of analytical output
noise ωest,yan and the case of measured output noise ωest,y , for the cavitated operating
condition, the frequency estimation performance is very close. The residual vibration in the
analytical case is due to the discrepancy between y and y an , but it has a very low amplitude
so that, macroscopically, the two curves cannot be distinguished. The comparison of the
frequency estimation for the uncavitated condition is presented in Figure 5, showing that
the performance degradation in terms of the settling of time, for analytical output noise
ωest,yan with respect to measured output noise ωest,y , is restrained in 2 s. In Figure 6,
both curves refer to the analytical output noise, and it is visible that different convergence
performances are registered in cavitated and uncavitated cases. This difference is also
recovered by the performance in phase estimation (Figure 7). As expected, the frequency
estimation procedure represents a warm-up phase of the overall closed-loop disturbance
estimation. In fact, according to Remark 6, the convergence of the asymptotical frequency
estimate is standalone with respect to the rest of the control loop. Hence, only once
the frequency estimate has reached its steady-state value, the phase estimation starts its
exponential convergence ((thesis iii) of Proposition 1) in the last segment of all four curves
(Figures 4, 5 and 7). The worst performance is registered in the uncavitated condition
with analytical noise (solid black line in Figure 7) since the phase estimation settling time
exceeds the other three simulation cases by 6 s.
The numerical analysis of the frequency and phase estimation timeseries gives the
fitting results shown in Figure 8, considering 95% confidence in the goodness of fit (GOF)
statistics. The order of convergence calculated for the frequency estimation timeseries are
pω,unc = 0.9983 and pω,cav = 1.005 under uncavitated and cavitated conditions, respec-
tively (Figure 8A,B), with acceptable GOF where the R2ω,unc = 0.9998 and R2ω,cav = 1.0.
The frequency estimation timeseries are sampled in the overall simulation time (degree of
freedom, dfe = 2398). Aiming to show that the phase estimation converges exponentially
only after the initial warm-up (according to Proposition 1), in the case of phase estima-
tion (Figure 8C,D), each timeseries sampling is restricted to the interval in which its own
warm-up phase ends: 8.41 (s) (dfe = 716) and 1.59 (s) (dfe = 2079) under uncavitated and
cavitated conditions, respectively, as shown in Figure 7. The convergence of phase estima-
tion approximates the exponential behavior better than frequency estimation, obtaining
p ϕ,unc = 1.001 and p ϕ,cav = 1 under uncavitated and cavitated conditions, respectively
(Figure 8C,D) with acceptable GOF where R2ϕ,unc = 0.9983 and R2ϕ,cav = 0.9989. The final
phase estimation error is registered within 0.43%. Overall, with respect to the fitting curve,
both frequency and phase estimation error timeseries exhibit less dispersion of the points
in the cavitated case than the points sampled in the uncavitated condition.
Figure 9 illustrates that, although an additive disturbance δyi affects the analytical
output y an , due to the disturbance δyi amplitude, which is bounded and considerably
smaller than the output noise itself (nine orders of magnitude less), y an fits the real distur-
..
bance dei (t, ω ) very well in both cavitated and uncavitated conditions (Figure 9A,C). This
enforces Assumption 1, and Proposition 1 applies. Figure 10 illustrates the comparison of
the reduced-order observer performance (Equation (20c)) in the center journal coordinates’
estimation in the four simulation cases. It emerges that the performance degradation in
the journal vibration absorption in the two cases with the analytical output (Figure 10A,C)

157
Mathematics 2022, 10, 1703

is restrained in a few tenths of millimeters with respect to the ideal case with measured
..
output (Figure 10B,D). The exponential hooking of the disturbance dei (t, ω ) by the adaptive
..
noise cancellation control dˆei (t, ωest ) (Equation (20b)) is illustrated in Figure 11. For sake
of brevity, only the worst case (uncavitated condition with analytical output noise) is re-
ported. Analogous results are obtained in the other three simulation cases. The preliminary
phase of the overall disturbance estimations is dictated by the asymptotical convergence
of the frequency estimate ωest . From that instant, the same closed-loop behavior of [24]
is recovered, with the last segment of the disturbance convergence, which is exponential.
..
The disturbance adaptation error d+ei (t, ω ), reported in Figure 12, in the four cases recovers
the observer behavior (Figure 10). Moreover, the behavior and performance of the control
input (Figure 13) recover the behaviors of the displacement observation error (Figure 10)
and the sinusoidal estimation errors (Figure 11) in the four cases. A zoomed plot is reported
to highlight the sinusoidal trend of the two control vector components, oscillating at the
same frequency as the rejected disturbance (as shown also in [24]). Figure 13 shows that
the control amplitude stabilizes at the same time (respectively, in the four cases) in which
all the estimation errors converge. Figure 14 illustrates that, from the comparison of the
rotor center coordinate stabilization in the four cases, it emerges that the performance
degradation in terms of vibration amplitude, for analytical output noise with respect to the
(ideal) measured output noise case, is restrained in a few tenths of millimeters.
Overall, better estimation performance is obtained in the cavitated condition with
respect to the uncavitated case. As the main figures of merit for the proposed adaptive
closed-loop control, the rotor vibration attenuation time (about 8.5 (s) in the worst case) and
the vibration residual amplitude (order of 10−1 (mm)) are considered, which are acceptable
values for a practical application. Similar results, in terms of estimation error convergence
time, are obtained in [17,18] (order of seconds), which presented tests of an experimental
setup on thin steel strips [17] (order of millimeters) and simulative in silico validation on a
reversible cold strip rolling mill [18].

5.2. Conclusions
The aim of this study was to provide a mathematical model for a novel adaptive
noise cancellation (ANC) technique designed to stabilize a flexible rotor shaft supported by
two hydrodynamic full (short) journal bearings and affected by a sinusoidal disturbance
output noise with unknown frequency. The disturbance frequency corresponds to the rotor
operating angular speed (ω), which is driven by an external actuator, so that uncertainties
may arise in the frequency actual value due to actuation operating point fluctuations.
The adaptation with respect to the frequency estimation (FE) generalizes the ANC to the
novel ANC-FE.
As the main novelty in this study, inspired from immersion and invariance (I&I)
techniques, an asymptotical frequency estimation (FE) module is designed as a combination
of state-observer and asymptotical parameter estimation. The FE operation represents
the warm-up phase of the overall adaptive noise cancellation control. The FE module is
externalized as an additive plug-in block, which processes the analytical reconstruction of
the output data downstream of the ANC closed-loop system (Figure 2).
The mathematical structural proof of the ANC-FE control theoretical formulation,
Proposition 1, is provided under Assumption 1, which requires that the analytical output
noise reconstruction is a signal at the same frequency of disturbance. It follows that: the
frequency estimation convergence is asymptotical; the disturbance phase estimation and
the rotor center coordinate stabilization are exponential.
The mathematical results have been validated experimentally in silico by numerical
simulations performed in four scenarios: cavitated and uncavitated conditions with analyt-
ical and measured output noise. Moreover, a data fitting analysis with the Inexact Newton
method (with 95% of confidence) is performed on frequency and phase estimation error
point series, in order to also validate numerically Proposition 1, demonstrating that the

158
Mathematics 2022, 10, 1703

phase estimate succession approximates the exponential behavior better than frequency
estimate succession. The orders of convergence obtained by the frequency estimation
timeseries are pω,unc = 0.9983 and pω,cav = 1.005; the order of convergence obtained by
the phase estimation timeseries is p ϕ = 1 for both uncavitated and cavitated conditions.
Considering the analytical output noise as the case of practical interest, the settling time of
disturbance rejection and then of the rotor center coordinates stabilization is about 76% less
in the cavitated than in the uncavitated condition, 2 (s) and 8.5 (s), respectively.
Simulation shows that the analytical output noise is very close to the measured output
noise, enforcing Assumption 1. The warm-up phase length is restrained in less than 10 s,
which is acceptable as an initialization process duration in a real application. For future
investigations the possible application of the proposed control technique to other dynamical
complex systems can be considered.

Author Contributions: Conceptualization, G.A., R.D. and A.R.; methodology, G.A. and R.D.; soft-
ware, G.A.; validation, G.A. and R.D.; formal analysis, G.A., R.D. and A.R.; investigation, G.A., R.D.
and A.R.; data curation, G.A.; writing—original draft preparation, G.A. and R.D.; writing—review
and editing, G.A., R.D. and A.R.; visualization, G.A. and R.D.; supervision, A.R. All authors have
read and agreed to the published version of the manuscript.
Funding: This research is funded by FARB 2019 University of Salerno (Italy).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Friswell, M.I.; Penny, J.E.T.; Garvey, S.D.; Lees, A.W. Dynamics of Rotating Machines; Cambridge University Press: Cambridge, UK,
2010; ISBN 9780511780509.
2. Lund, J.W. Review of the Concept of Dynamic Coefficients for Fluid Film Journal Bearings. J. Tribol. ASME US 1987, 109, 37–41.
[CrossRef]
3. Hamrock, B.J.; Schmid, S.R. Fundamental of Fluid Film Lubrication, 2nd ed.; CRC Press: Boca Raton, FL, USA, 2004; ISBN 0824753712.
4. Salazar, J.G.; Santos, I.F. Active tilting-pad journal bearings supporting flexible rotors: Part I—The hybrid lubrication. Tribol. Int.
2017, 107, 94–105. [CrossRef]
5. Chen, Y.; Yang, R.; Sugita, N.; Mao, J.; Shinshi, T. Identification of bearing dynamic parameters and unbalanced forces in a flexible
rotor system supported by oil-film bearings and active magnetic devices. Actuators 2021, 10, 216. [CrossRef]
6. Muszynska, A. Stability of whirl and whip in rotor/bearing systems. J. Sound Vib. 1988, 127, 49–64. [CrossRef]
7. D’Agostino, V.; Guida, D.; Ruggiero, A.; Senatore, A. An analytical study of the fluid film force in finite-length journal bearings.
Part I. Lubr. Sci. 2001, 13, 329–340. [CrossRef]
8. Poritsky, H. Contribution to the theory of oil whip. Trans. ASME 1953, 75, 1153–1161.
9. Tripathy, D.; Bhattacharyya, K. Analysis of a Hydrodynamic Journal Bearing of Circular Cross Section Lubricated by a Magne-
tomicropolar Fluid. In Proceedings of the Lecture Notes in Mechanical Engineering; Springer: Singapore, 2022; pp. 1495–1502.
10. Das, S.; Guha, S.K. Non-linear stability analysis of micropolar fluid lubricated journal bearings with turbulent effect. Ind. Lubr.
Tribol. 2019, 71, 31–39. [CrossRef]
11. Bhattacharjee, B.; Chakraborti, P.; Choudhuri, K. Evaluation of the performance characteristics of double-layered porous
micropolar fluid lubricated journal bearing. Tribol. Int. 2019, 138, 415–423. [CrossRef]
12. Harika, E.; Bouyer, J.; Fillon, M.; Hélène, M. Measurements of lubrication characteristics of a tilting pad thrust bearing disturbed
by a water-contaminated lubricant. Proc. Inst. Mech. Eng. Part J. J. Eng. Tribol. 2013, 227, 16–25. [CrossRef]
13. Vania, A.; Pennacchi, P.; Chatterton, S. Dynamic Effects Caused by the Non-Linear Behavior of Oil-Film Journal Bearings in
Rotating Machines. In Proceedings of the Volume 7: Structures and Dynamics, Parts A and B; ASME: New York, NY, USA, 2012; p. 657.
14. Das, S.; Guha, S.K.; Chattopadhyay, A.K. Linear stability analysis of hydrodynamic journal bearings under micropolar lubrication.
Tribol. Int. 2005, 38, 500–507. [CrossRef]
15. Sukumaran Nair, V.P.; Prabhakaran Nair, K. Finite element analysis of elastohydrodynamic circular journal bearing with
micropolar lubricants. Finite Elem. Anal. Des. 2004, 41, 75–89. [CrossRef]
16. Prabhakaran Nair, K.; Sukumaran Nair, V.P.; Jayadas, N.H. Static and dynamic analysis of elastohydrodynamic elliptical journal
bearing with micropolar lubricant. Tribol. Int. 2007, 40, 297–305. [CrossRef]
17. Marko, L.; Saxinger, M.; Steinboeck, A.; Kugi, A. Cancellation of unknown multi-harmonic disturbances in multivariable flexible
mechanical structures. Automatica 2022, 137, 110123. [CrossRef]

159
Mathematics 2022, 10, 1703

18. Liu, L.; Shao, N.; Deng, R.; Ding, S. Immersion and invariance adaptive decentralized control for the speed and tension system of
the reversible cold strip rolling mill. Int. J. Adapt. Control Signal Process. 2022, 36, 785–801. [CrossRef]
19. Burrows, C.R.; Sahinkaya, M.N. Vibration control of multi-mode rotor-bearing systems. Proc. R. Soc. Lond. A 1983, 386, 77–94.
20. Ruggiero, A.; D’Amato, R.; Magliano, E.; Kozak, D. Dynamical simulations of a flexible rotor in cylindrical uncavitated and
cavitated lubricated journal bearings. Lubricants 2018, 6, 40. [CrossRef]
21. Zhou, S.; Shi, J. Active balancing and vibration control of rotating machinery: A survey. Shock Vib. Dig. 2001, 33, 361–371.
[CrossRef]
22. Stanway, R.; Burrows, C.R. Active vibration control of a flexible rotor on flexibly-mounted journal bearings. J. Dyn. Syst. Meas.
Control 1981, 103, 383–388. [CrossRef]
23. Zheng, S.; Li, H.; Peng, C.; Wang, Y. Experimental Investigations of Resonance Vibration Control for Noncollocated AMB Flexible
Rotor Systems. IEEE Trans. Ind. Electron. 2017, 64, 2226–2235. [CrossRef]
24. D’Amato, R.; Amato, G.; Wang, C.; Ruggiero, A. A Novel Tracking Control Strategy with Adaptive Noise Cancellation for Flexible
Rotor Trajectories in Lubricated Bearings. IEEE ASME Trans. Mechatron. 2022, 27, 753–765. [CrossRef]
25. Lei, S.; Palazzolo, A. Control of flexible rotor systems with active magnetic bearings. J. Sound Vib. 2008, 314, 19–38. [CrossRef]
26. Marko, L.; Saxinger, M.; Steinboeck, A.; Kemmetmüller, W.; Kugi, A. Frequency-adaptive cancellation of harmonic disturbances
at non-measurable positions of steel strips. Mechatronics 2020, 71, 102423. [CrossRef]
27. Marino, R.; Tomei, P. Adaptive disturbance rejection for unknown stable linear systems. Trans. Inst. Meas. Control 2016, 38,
640–647. [CrossRef]
28. Kumar, P.; Tiwari, R. Dynamic analysis and identification of unbalance and misalignment in a rigid rotor with two offset discs
levitated by active magnetic bearings: A novel trial misalignment approach. Propuls. Power Res. 2020, 10, 58–82. [CrossRef]
29. D’Amato, R.; Amato, G.; Ruggiero, A. Adaptive Noise Cancellation-Based Tracking Control for a Flexible Rotor in Lubricated
Journal Bearings. In Proceedings of the 2019 23rd International Conference on Mechatronics Technology (ICMT), Salerno, Italy,
23–26 October 2019; pp. 1–5. [CrossRef]
30. Vance, J.M. Rotordynamics of Turbomachinery; Wiley: Hoboken, NJ, USA, 1988; ISBN 0471802581.
31. Avramov, K.V.; Borysiuk, O.V. Nonlinear dynamics of one disk asymmetrical rotor supported by two journal bearings. Nonlinear
Dyn. 2012, 67, 1201–1219. [CrossRef]
32. Carnevale, D.; Astolfi, A. A minimal dimension observer for global frequency estimation. Proc. Am. Control Conf. 2008, 2,
5236–5241. [CrossRef]
33. Carnevale, D. Robust hybrid estimation and rejection of multi-frequency signals. Int. J. Adapt. Control Signal Process. 2016, 30,
1649–1673. [CrossRef]
34. Hoad, K.; Robinson, S.; Davies, R. Automating warm-up length estimation. J. Oper. Res. Soc. 2010, 61, 1389–1403. [CrossRef]
35. Ballnus, B.; Schaper, S.; Theis, F.J.; Hasenauer, J. Bayesian parameter estimation for biochemical reaction networks using
region-based adaptive parallel tempering. Bioinformatics 2018, 34, i494–i501. [CrossRef]
36. Ypma, T.J. Local Convergence of Inexact Newton Methods. SIAM J. Numer. Anal. 2006, 21, 583–590. [CrossRef]
37. Sastry, S.; Bodson, M. Adaptive Control—Stability, Convergence, and Robustness; Prentice-Hall: Hoboken, NJ, USA, 1989; p. 201.
38. Marino, R.; Tomei, P. Output Regulation for Unknown Stable Linear Systems. IEEE Trans. Autom. Control 2015, 60, 2213–2218.
[CrossRef]
39. Sun, X.; Member, S.; Wu, M.; Yin, C.; Wang, S.; Tian, X. Multiple-Iteration Search Sensorless Control for Linear Motor in Vehicle
Regenerative Suspension. IEEE Trans. Transp. Electrif. 2021, 7, 1628–1637. [CrossRef]
40. Marino, R.; Tomei, P. Nonlinear Control Design: Geometric, Adaptive and Robust; Prentice Hall: London, UK, 1995; Volume 1.
41. Marino, R.; Tomei, P.; Verrelli, C.M. Induction Motor Control Design; Springer Science & Business Media: Berlin/Heidelberg,
Germany, 2010; ISBN 1849962847.

160
 mathematics

Article
Complete Balancing of the Six-Bar Mechanism Using Fully
Cartesian Coordinates and Multiobjective Differential
Evolution Optimization
María T. Orvañanos-Guerrero 1 , Mario Acevedo 2 , Claudia N. Sánchez 1 , Daniel U. Campos-Delgado 3 , Amir
Aminzadeh Ghavifekr 4 , Paolo Visconti 1,5 and Ramiro Velázquez 1, *

1 Facultad de Ingeniería, Universidad Panamericana, Aguascalientes 20290, Mexico;

torvananos@up.edu.mx (M.T.O.-G.); cnsanchez@up.edu.mx (C.N.S.); paolo.visconti@unisalento.it (P.V.)
2 Facultad de Ingeniería, Universidad Panamericana, Zapopan 45010, Mexico; macevedo@up.edu.mx
3 Facultad de Ciencias, Instituto de Investigación en Comunicación Óptica, Universidad Autónoma de San
Luis Potosí, San Luis Potosí 78295, Mexico; ducd@fciencias.uaslp.mx
4 Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz 5166616471, Iran;
aa.ghavifekr@tabrizu.ac.ir
5 Department of Innovation Engineering, University of Salento, 73100 Lecce, Italy
* Correspondence: rvelazquez@up.edu.mx; Tel.: +52-449-910-6200

Abstract: The high-speed operation of unbalanced machines may cause vibrations that lead to noise,
wear, and fatigue that will eventually limit their efficiency and operating life. To restrain such
vibrations, a complete balancing must be performed. This paper presents the complete balancing
Citation: Orvañanos-Guerrero, M.T.;
optimization of a six-bar mechanism with the use of counterweights. A novel method based on fully
Acevedo, M.; Sánchez, C.N.;
Cartesian coordinates (FCC) is proposed to represent such a balanced mechanism. A multiobjective
Campos-Delgado, D.U.; Ghavifekr,
optimization problem was solved using the Differential Evolution (DE) algorithm to minimize the
A.A.; Visconti, P.; Velázquez, R.
shaking force (ShF) and the shaking moment (ShM) and thus balance the system. The Pareto front is
Complete Balancing of the Six-Bar
Mechanism Using Fully Cartesian
used to determine the best solutions according to three optimization criteria: only the ShF, only the
Coordinates and Multiobjective ShM, and both the ShF and ShM. The dimensions of the counterweights are further fine-tuned with
Differential Evolution Optimization. an analysis of their partial derivatives, volumes, and area–thickness relations. Numerical results show
Mathematics 2022, 10, 1830. https:// that the ShF and ShM can be reduced by 76.82% and 77.21%, respectively, when importance is given
doi.org/10.3390/math10111830 to either of them and by 45.69% and 46.81%, respectively, when equal importance is given to both. A
comparison of these results with others previously reported in the literature shows that the use of
Academic Editors: Higinio Rubio
Alonso, Alejandro Bustos Caballero,
FCC in conjunction with DE is a suitable methodology for the complete balancing of mechanisms.
Jesus Meneses Alonso and Enrique
Soriano-Heras Keywords: six-bar mechanism; dynamic balancing; fully Cartesian coordinates; multiobjective
optimization; differential evolution; Pareto front
Received: 23 April 2022
Accepted: 21 May 2022
MSC: 91B03; 70B15
Published: 26 May 2022

Publisher’s Note: MDPI stays neutral

with regard to jurisdictional claims in
published maps and institutional affil- 1. Introduction
iations. The high-speed motion of the links of a mechanism generates dynamic reactions that
are transmitted as vibrations to the fixed frame producing undesirable effects such as noise,
wear, fatigue, and energy losses that will eventually limit the mechanism’s performance and
operating life. In addition to these technical problems, vibrations may induce a constant
Copyright: © 2022 by the authors.
machine maintenance involving wastes that have an environmental impact and noise
Licensee MDPI, Basel, Switzerland.
This article is an open access article
pollution that can lead to health problems [1].
distributed under the terms and
The dynamic reactions under discussion are produced by both external and inertial
conditions of the Creative Commons forces. The former are naturally forces generated outside the mechanism such as those
Attribution (CC BY) license (https:// induced by an actuator. The latter are due to the large accelerations of the links. Some
creativecommons.org/licenses/by/ representative methods for calculating dynamic reactions in mechanisms are those of
4.0/). Uicker [2] and Waldron [3].

Mathematics 2022, 10, 1830. https://doi.org/10.3390/math10111830 161 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 1830

The force and the moment that occur at the mechanism’s fixed frame are typically
known as shaking force (ShF) and shaking moment (ShM), respectively. Thus, a traditional
but still open challenge in machine theory is how to effectively remove or minimize the
dynamic reactions, i.e., the ShF and ShM, derived from the mechanism’s motion. This
process is known as complete (or dynamic) balancing.
The dynamic balancing of mechanisms has been widely studied in mechanical engi-
neering. Some recent historical reviews addressing the topic can be found in [1,4–6].
Among the first publications, we can find the work of Fischer introducing in 1902 a
method called principal vectors [7]. This method allows for the balancing of the ShF by
analyzing each of the links of the mechanism and determining the points, called principal
points, in which the static balancing can be reached. The work of Fischer provided the
basis for the methods studying the motion of the centers of mass (CoM) of the links in a
mechanism. The method of principal vectors was subsequently used by Goryachkin [8],
Yudin [9], and Kreutzinger [10].
Another early method proposed for the dynamic balancing of mechanisms was the
static substitution of masses. Its aim is to statically replace the mass of the coupler by
concentrated masses, thus transforming the problem of mechanism balancing into a simpler
problem of balancing the rotating links. The works of Maxwell [11], Smith [12], and
Talbourdet [13] are based on this method.
The 1920s were marked by a special interest in the balancing of engines [14,15] and
machines related to agriculture [16]. The Lanchester balancer stands out among these
works. It is still being used for balancing four-stroke engines.
During the 1940s, some methods based on function approximation were proposed to
achieve partial equilibrium. The work of Gheronimus [17] is a representative example. In
this work, the balancing conditions were formulated by minimizing the root mean square
(RMS) or the maximum values of the ShM. The duplicate mechanism method [18] was also
proposed in this decade. It achieves a complete balancing with the addition of axial and
mirror symmetric mechanism duplicates.
The balancing methods based on harmonic analysis appeared in the 1960s with the
use of the crank-slider mechanism in internal combustion engines [19,20]. Such methods
reduce the ShM by balancing certain harmonics of both the ShF and ShM. To carry out
this process, the unbalanced forces and moments are divided into Fourier series and then
analyzed in parts.
In 1969, Berkof and Lowen [21] proposed a new solution for dynamic balancing using
a method called linearly independent vectors. This method consists of formulating an
equation considering a vector representation describing the position of the total CoM of the
mechanism in conjunction with the equation representing the closed kinematic chain. A
system of equations containing linearly independent vectors is then obtained allowing to
find the balancing conditions of the mechanism by reducing to zero the time-dependent
coefficients. This method was subsequently explored in [22,23].
In the 1970s, the dynamic balancing theory achieved significant advances. Berkof [24]
formulated the first method for complete balancing using counterweights and counter-
inertias to eliminate the undesired forces and moments of the moving links, respectively. It
then became clear that a complete balancing of mechanisms is feasible at the expense of
considering complex design modifications and an inevitable increase of the total mass and
volume of the mechanism.
Methods proposing a partial balancing were then proposed searching to keep the
added mass reasonably small. In [25], Wiederrich and Roth formulated simple general
conditions for determining the inertia properties of a four-bar mechanism and achieved
its partial balancing. Dresig examined the partial balancing conditions for several 2D
mechanims with six and eight-bar linkages [26].
Later approaches involved the use of planetary [27] and auto-balanced systems [28].
More recent ones propose the use of the instantaneous dynamic balancing conditions [29],
the trajectory planning of the CoM of the mechanism [30], and the reformulation of the

162
Mathematics 2022, 10, 1830

balancing conditions following a Taylor-based approach [29]. Finally, the most recent and
novel balancing method encompasses the flexibility of the links of the mechanism [31].
We focus our attention on Berkof’s counterweight approach. After the partial balancing
methods, optimization techniques were then explored searching a tradeoff between the
added mass and the reduction of the ShF and ShM.
In 1998, Segla presented the static balancing optimization of a robotic mechanism [32].
The balancing conditions were first obtained and then, a basic genetic algorithm pro-
grammed in Fortran was used for the optimization.
Currently, the most popular optimization techniques for mechanical problems are
based on metaheuristic methods: evolutionary [33], differential evolution [34], genetic [35],
and Firefly [36] algorithms. These proposals were designed to find heuristics (i.e., partial
solutions) that may provide sufficiently good tradeoffs for the dynamic balancing problem.
To our knowledge, all previous works addressing optimization techniques for the
dynamic balancing of mechanisms make use of Cartesian coordinates (CC) to obtain
the expressions of the dynamic reactions. A major drawback of CC is that they involve
trigonometric functions that derive into complex mathematical expressions that are compu-
tationally burdensome.
Our research approach differs from the existing literature in that it presents an alterna-
tive to CC: fully Cartesian coordinates (FCC), which are also called natural coordinates [37].
By using FCC, the dynamic reactions at the fixed frame can be formulated by means of
equations of less complexity (no angular variables). Then, the resulting ShF and ShM
expressions are used to optimize the dynamic balancing of mechanisms through the use
of counterweights.
Our previous work [38,39] successfully reported the dynamic balancing of a four-bar
mechanism and its optimization with Projected Gradient Descent. In this paper, we address
the dynamic balancing of a more complex system, a six-bar mechanism, and explore the
differential evolution (DE) algorithm as our optimization method.
Parallel mechanisms are increasingly being used in robotic applications [40,41]. The
six-bar mechanism is a typical parallel manipulator. This single degree-of-freedom planar
linkage is typically used as a variable-speed transmission mechanism where the input crank
rotates at constant speed and the output link works as an overrunning clutch mounted on
the output shaft [42].
The rest of the paper is organized as follows: In Section 2, the FCC-based mass-matrix
definition for the six-bar mechanism with counterweights is introduced together with
the expressions of the ShF and ShM. In Section 3, the details of the multiobjective DE
optimization are presented. In Section 4, a numerical example is presented to illustrate the
proposed approach. In Section 5, results are discussed and compared to others previously
reported in the literature. Finally, Section 6 concludes, summarizing the main contributions
and giving the future work perspectives of this research.

2. Mechanical Analysis
2.1. FCC-Based Definition of the Mass Matrix of an Element Defined by Three Basic Points
This section details the method used to obtain the mass matrix M3P of elements
defined by three basic points using FCC and the concept of virtual work.
Consider an element defined by three basic points i, j, and k, as shown in Figure 1.
The element is located in a global coordinate system ( x, y) and in a local coordinate system
( x̄, ȳ) with its origin at point i and the x̄ axis directed toward point j.

163
Mathematics 2022, 10, 1830

Figure 1. A 2D element with three basic points (i, j, and k).

The location of any point P in this element is defined by a vector r in the global
reference system and a vector r̄ in the local coordinate system. In this way, r can be
expressed, according to Equation (1).

r = ri + Ar̄ (1)
where A is the rotation matrix. Being the element rigid, the local position of vector r̄
remains constant regardless of the element’s motion. Thus, the position of point P can be
defined according to Equation (2).

r = ri + Ar̄ = ri + c1 (rj − ri) + +c2 (rk − ri ) (2)

where c1 and c2 are the components of vector r̄ in the local coordinate system. The compo-
nents of vector r can be expressed in matrix form as shown in Equation (3).
⎧ ⎫
⎪ xi ⎪
⎪
⎪ ⎪
⎪ yi ⎪
⎪ ⎪
⎪
( )  ⎪
⎨x ⎪ ⎬
x 1 − c1 − c2 0 c1 0 c2 0 j
r= = = Cq (3)
y 0 1 − c1 − c2 0 c1 0 c2 ⎪
⎪ yj ⎪
⎪
⎪
⎪ ⎪
⎪
⎪
⎪ x ⎪
⎩ k⎪ ⎭
yk

where qT = xi yi x j y j xk yk is the vector that contains the FCC of the element.

Note that matrix C is constant for a given point P and does not change with the system’s
motion, thus fulfilling Equations (4) and (5).

ṙ = Cq̇ (4)

r̈ = Cq̈ (5)
Coefficients c1 and c2 in matrix C can be expressed in terms of the coordinates of
points i, j, and k in the local reference frame according to Equation (6).

r̄ = c1 (rj − ri ) + c2 (rk − ri ) (6)

Since r¯i = 0 is located at the local reference origin, Equation (6) can be rewritten as
shown in Equation (7).
( )
c
r̄ = [ r¯j | r¯k ] 1 = X̄c (7)
c2

164
Mathematics 2022, 10, 1830

Vector c contains the coefficients c1 and c2 , and matrix X̄ has the components of vectors
r¯j and r¯k in its columns (Equation (8)).
   
x xk l Kx
X̄ = [ r¯j | r¯k ] = j = ij (8)
y j yk 0 Ky
Now, it is possible to define the virtual work W ∗ generated by the inertial forces
(Equation (9)).
.
W ∗ = −ρ ṙ∗T r̈dΩ (9)
Ω
where ρ is the density of the element’s material. Substituting Equations (4) and (5) into
Equation (9) yields to the definition of virtual work (Equation (10)):
.
W ∗ = −ρ q̇∗T CT Cq̈dΩ (10)
Ω

Since vectors q̇∗T and q̈ are independent of Ω, they can be taken out of the integral as
expressed by Equation (11).
 . 
W ∗ = −q̇∗T ρ CT CdΩ q̈ (11)
Ω

On the other hand, taking into account the definition of virtual work proposed in [43]
(Equation (12)) and comparing it to Equation (11), the mass matrix M3P can be expressed
by Equation (13).
W ∗ = −q̇∗T Mq̈ (12)
.
M3P = ρ CT CdΩ (13)
Ω

Further development of the product of CT C in Equation (13) yields to Equation (14).

⎡ ⎤
ce 0 c f 0 c g 0
⎢0 c 0 c f 0 cg ⎥
⎢
. ⎢
e ⎥
⎥
⎢ c f 0 c h 0 ci 0 ⎥
M3P = ρ ⎢ ⎥dΩ (14)
Ω⎢ 0 c f 0 c h 0 ci ⎥
⎢ ⎥
⎣ c g 0 ci 0 c j 0 ⎦
0 c g 0 ci 0 c j
with:
ce = c21 + 2c1 c2 − 2c1 + c22 − 2c2 + 1 (15)
cf = −c21 − c1 c2 + c1 (16)

cg = −c1 c2 − c22 + c2 (17)

ch = c21 (18)
c i = c1 c2 (19)
cj = c22 (20)
Note that Equation (14) involves solving the integrals of Equations (21)–(23).
.
ρdΩ = m (21)
Ω

165
Mathematics 2022, 10, 1830

. . .  
1
− KKy xl x̄
ρX̄−1 r̄dΩ = ρ
lij
ρcdΩ = 1
ij
dΩ =
Ω Ω Ω 0 Ky
ȳ
⎡ ⎤ (22)
. x̄
− ȳK x m x̄ g mȳ K
− Kygl x
ρ lij
ȳ
Ky lij
dΩ = ⎣ lij
mȳ g
ij ⎦
Ω
Ky Ky

. .  .  2  
x̄ x̄ ȳ
ρccT dΩ = X̄−1 ρr̄r̄T dΩ X̄−T = X̄−1 ρ 2 dΩ X̄−T =
Ω Ω Ω x̄ ȳ ȳ
⎡ 2 ⎤
  Ix K x 2Ixy K x Iy I (23)
2 l2 − K l2 + l2 − KIx2Kl x + Kyxyl
I I ⎢ ⎥
X̄−1 X̄−T = ⎣ ij
y xy K y y y ij ij
ij
I
ij
⎦
Ixy Ix − KIx2Kl x + Kyxyl Ix
y ij K2 ij y

where m is the total mass of the element, r̄ represents the local coordinates of the center of
gravity, and Ix , Iy , and Ixy are the moments and products of inertia with respect to local
coordinates with origin at the basic point i.
By substituting the integrals of Equations (21)–(23) into Equation (14), we finally obtain
the mass matrix M3P (Equation (24)).
⎡ ⎤
e 0 f 0 g 0
⎢0 e 0 f 0 g⎥
⎢ ⎥
⎢f 0 h 0 i 0⎥
M3P =⎢
⎢0
⎥ (24)
⎢ f 0 h 0 i⎥⎥
⎣g 0 i 0 j 0⎦
0 g 0 i 0 j
with:

Ix K2x 2Ix Kx Ix 2Ixy Kx 2Ixy Iy 2Kx mȳ g 2m x̄ g 2mȳ g

e= − 2 + 2− + + 2 + +m− − (25)
Ky2 l 2 Ky lij Ky Ky lij2 Ky lij lij Ky lij lij Ky

Ix K2x Ix K x 2Ixy Kx Ixy Iy Kx mȳ g m x̄ g

f =− + 2 + − − 2 − + (26)
Ky2 lij2 Ky lij Ky lij2 Ky lij lij Ky lij lij

Ix K x Ix Ixy mȳ g
g= − 2− + (27)
Ky2 lij Ky Ky lij Ky

Ix K2x 2Ixy Kx Iy
h= − + 2 (28)
Ky2 lij2 Ky lij2 lij

Ix K x Ixy
i=− + (29)
Ky2 lij Ky lij
Ix
j= (30)
Ky2

2.2. Six-Bar Mechanism

Figure 2 shows the six-bar mechanism. This kind of mechanism exhibits one degree of
freedom and five mobile links.
Each of the links has a local coordinate system with its origin located at point i and
the x axis directed toward point j. Each of the links has a mass mbn and a center of gravity
located at local coordinates (xbn , ybn ) for 1 ≤ n ≤ 5. The distribution of the points is
detailed in Table 1.

166
Mathematics 2022, 10, 1830

For this mechanism, it is possible to define a vector q representing the positions of its
basic points (Equation (31)).

q = [ A x Ay Bx By Cx Cy Dx Dy Ex Ey Fx Fy Gx Gy ] T (31)

By time-deriving q, it is possible to obtain a new vector q̇ representing the velocities

of each of the basic points of the mechanism (Equation (32)). Similarly, by time-deriving
the velocity vector, the acceleration vector q̈ can be obtained (Equation (33)).

q̇ = [VAx VAy VBx VBy VCx VCy VDx VDy VEx VEy VFx VFy VGx VGy ] T (32)

q̈ = [ A Ax A Ay A Bx A By ACx ACy A Dx A Dy A Ex A Ey A Fx A Fy AGx AGy ] T

(33)

Figure 2. The six-bar mechanism.

Table 1. Point distribution for the six-bar mechanism.

Element Point i Point j Point k

1 C E D
2 A C
3 B F E
4 D G
5 F G

2.3. Counterweight Addition

To proceed with the dynamic balancing of the six-bar mechanism, a set of cylindrical
counterweights is added to its structure (Figure 3).
To simplify the procedure, let us assume that all counterweights are coincident with
the basic point i of each of the elements of the mechanism. The center of gravity of each

167
Mathematics 2022, 10, 1830

counterweight is located at the local coordinates (xcn , ycn ) for 1 ≤ n ≤ 5. Their thickness is
defined by tcn for 1 ≤ n ≤ 5.

Figure 3. Six-bar mechanism with cylindrical counterweights.

It is then possible to define the mass mcn of each counterweight as a function of

its density ρcn = ρ, its thickness tcn , and the location of its CoM (xcn , ycn ), as shown in
Equation (34).
mcn = ρ V = ρπtcn ( xcn2
+ y2cn ) (34)
It is also possible to obtain the mass moments for each of the counterweights with
respect to the local coordinate system origin, as shown in Equations (35) and (36).

1 1 1
Ixcn = mcn ( xcn
2
+ y2cn ) + my2cn = mcn ( xcn
2
+ 5y2cn ) = ρπtcn ( xcn
2
+ y2cn )( xcn
2
+ 5y2cn ) (35)
4 4 4
1 1 1
Iycn = mcn ( xcn
2
+ y2cn ) + mxcn
2
= mcn (5xcn
2
+ y2cn ) = ρπtcn ( xcn
2
+ y2cn )(5xcn
2
+ y2cn ) (36)
4 4 4
The polar moment of inertia of each counterweight Izcn with respect to the local
coordinate system origin can be defined by Equation (37).

3 3
Izcn = Ixcn + Iycn = mcn ( xcn
2
+ y2cn ) = ρπtcn ( xcn
2
+ y2cn )2 (37)
2 2
Similarly, the product of inertia of each counterweight with respect to the local coordi-
nate system origin can be calculated with Equation (38).

Ixycn = mcn xcn ycn = ρπtcn ( xcn

2
+ y2cn ) xcn ycn (38)

2.4. Mass Matrix for the Six-Bar Mechanism with Counterweights

To formulate the mass matrix of the balanced mechanism, it is necessary to obtain the
mass matrices for each of the linkages (link and counterweight) first. The details of these
individual mass matrices can be found in Appendix A.

168
Mathematics 2022, 10, 1830

Given that the mechanism consists of seven basic points, each of them represented by
one (x, y) coordinate, the resulting mass matrix will consist of 14 columns and 14 rows.
Equation (39) shows the mass matrix M representing the six-bar mechanism with
counterweights.
⎡ ⎤
a2 0 0 0 b2 c2 0 0 0 0 0 0 0 0
⎢ 0 − c2 ⎥
⎢ a2 0 0 b2 0 0 0 0 0 0 0 0 ⎥
⎢ 0 ⎥
⎢ 0 e3 0 0 0 0 0 g3 0 f3 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 0 e3 0 0 0 0 0 g3 0 f3 0 0 ⎥
⎢ ⎥
⎢ b2 c2 0 0 d 2 + e1 0 g1 0 f1 0 0 0 0 0 ⎥
⎢ ⎥
⎢ − c2 b2 0 0 0 d 2 + e1 0 g1 0 f1 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 g1 0 a4 + j1 0 i1 0 0 0 b4 c4 ⎥
M=⎢
⎢ 0
⎥
⎥ (39)
⎢ 0 0 0 0 g1 0 a4 + j1 0 i1 0 0 − c4 b4 ⎥
⎢ 0 0 g3 0 f1 0 i1 0 h1 + i3 0 i3 0 0 0 ⎥
⎢ ⎥
⎢ 0 h1 + i3 ⎥
⎢ 0 0 g3 0 f1 0 i1 0 0 i3 0 0 ⎥
⎢ ⎥
⎢ 0 0 f3 0 0 0 0 0 i3 0 a5 + h3 0 b5 c5 ⎥
⎢ ⎥
⎢ 0 0 0 f3 0 0 0 0 0 i3 0 a5 + h3 − c5 b5 ⎥
⎢ ⎥
⎣ 0 0 0 0 0 0 b4 c4 0 0 b5 c5 d4 + d5 0 ⎦
0 0 0 0 0 0 − c4 b4 0 0 − c5 b5 0 d4 + d5

where an , bn , . . . , jn are the terms of the different mass matrices for linkages n = 1, . . . , 5 (see
Appendix A).

2.5. Linear Momentum and Shaking Force

Once vectors q, q̇, and q̈ have been defined (Equations (31)–(33)), it is possible to
calculate the linear momentum L associated to the balanced mechanism (Equation (40)).
 
Li
= BMq̇ (40)
Lj

where B (Equation (41)) is a matrix formed by identity matrices matching the number of
basic points in the mechanism.
 T
1 0 1 0 1 0 1 0 1 0 1 0 1 0
B= (41)
0 1 0 1 0 1 0 1 0 1 0 1 0 1

By solving Equation (40) and considering that the velocity of the fixed points is always
zero (VA X = 0, VAY = 0, VBX = 0, VBY = 0), the expressions of the linear momentum (Li
and Lj ) can be obtained.
The ShFi and ShF j of the mechanism can be computed by time-deriving the equations
Li and Lj (Equation (40)) (expressions are not included in the paper because of their length).
To guarantee the equilibrium, the result of these derivatives must be constant (normally
zero) in the analyzed period of time.

2.6. Angular Momentum and Shaking Moment

The use of FCC allows us to express the angular momentum H of the mechanim, as
shown in Equation (42):
H = q × (Mq̇) = rMq̇ (42)
where r is given as a function of the basic points and can be expressed as Equation (43):
 T
r = − AY AX − BY BX −CY CX − DY DX − EY EX − FY FX − GY GX (43)
The ShM can then be calculated by time-deriving H (Equations (44) and (45)).

dH d(q̇) dr
ShM = = rM( ) + ( )Mq̇ (44)
dt dt dt

169
Mathematics 2022, 10, 1830

dH
ShM = = rMq̈ + ṙMq̇ (45)
dt
with:

ṙ =[−VAY VAX − VBY VBX − VCY VCX − VDY VDX

(46)
− VEY VEX − VFY VFX − VGY VGX ]T

To guarantee the dynamic equilibrium of the mechanism, the ShM must be constant,
i.e., the time-derivative of H (Equation (45)) must be zero.
The ShM of the mechanism is finally obtained by solving Equation (45) and considering
VAX = 0, VAY = 0, VBX = 0 y VBY = 0 (equation is not included in the paper because of
its length).

3. Optimization
3.1. Objective Function
Two dimensionless balancing indexes β i , containing the motion parameters (q, q̇, and
q̈) of the six-bar mechanism, can be used to define the optimization’s objective function:
β ShF and β ShM .
β ShF (Equation (47)) is defined by the RMS value of the ShF reaction of the opti-
mized mechanism (rms(o ShF )) with respect to the RMS value of the original mechanism
(rms(ShF )), which are both considered over a time period T.
0
1 N o
rms(o ShF ) 1 ∑k=1 ( ShFik2 +o ShFjk2 )
β ShF = =2 N (47)
rms(ShF ) ∑k=1 (ShFik2 + ShFjk2 )

β ShM (Equation (48)) can be calculated in a similar way. Nevertheless, the reaction
produced by ShM must also be considered.

∑kN=1 o ShMk2
β ShM = (48)
∑kN=1 ShMk2
where o ShM is the shaking moment of the optimized mechanism and ShM is a constant
representing the shaking moment of the unbalanced mechanism.
A multiobjective optimization problem emerges as it is desired to minimize both β ShF
and β ShM considering the variables boundaries (i.e., the physical limits for the locations of
the CoM (xcn and ycn ) and the thickness (tcn ) of each counterweight). To solve this problem,
a linear combination of the objectives is performed as proposed in Equation (49).

f ( X ) = γ ∗ β ShM + (1 − γ) ∗ β ShF (49)

where γ is a scalar value that assigns the importance to each optimization objective. Thus,
the 15 variables to be optimized are: xc1 , yc1 , tc1 , xc2 , yc2 , tc2 , xc3 , yc3 , tc3 , xc4 , yc4 , tc4 , xc5 , yc5 ,
and tc5 . The boundaries for optimization are defined according to Equation (50).
min max
xcn xcn xcn
ymin
cn ycn ymax
cn (50)
tmin
cn tcn tmax
cn

3.2. Algorithm
Once the objective function has been proposed, an optimization method can be applied.
In this paper, we explore Differential Evolution (DE) [44].
Being an evolutionary algorithm, DE uses approaches inspired by the theory of evolu-
tion. It optimizes a problem by proposing a population of candidate solutions and creating
new candidate solutions with the ones that obtained the best scores. Thus, the new gen-

170
Mathematics 2022, 10, 1830

erations are better than the previous ones. Recently, DE has been successfully applied to
disruptive fields such as the oil market [45] and genome studies [46].
The DE algorithm proposed for the complete balancing optimization of the six-bar
mechanism is presented in Algorithm 1. It was programmed in Python.

Algorithm 1: Differential Evolution (DE).

Input : N = 255, F = 0.8, CR = 0.7, k max = 100
1 An initial population is generated S = { X1k , X2k , . . . , X Nk }
2 Xbest = solution with the lowest value in the objective function
3 for k = 0 until k max do
4 Q = {}
5 for i = 1 until N do
6 Selection of three random solutions in S γ, δ, η ∈ {1, . . . , N }, γ = δ = η = i
7 X̂ik = Xγk + F ( Xδk − Xηk )
8 X̂ik = Clip( X̂ik )
9 for j = 1 until D do
10 R = Random value between 0 and 1 with uniform distribution
11 if R ≤ CR then
12 Yikj = X̂ikj
13 else
14 Yikj = Xikj
15 end
16 end
17 if f (Yik ) < f ( Xik ) then
18 Q = Q ∪ Yik
19 else
20 Q = Q ∪ Xik
21 end
22 If f (Yik ) < f ( Xbest ) then Xbest = Yik
23 end
24 S=Q
25 k = k+1
26 end
27 return Xbest

Given that the number of variables to optimize is equal to 15, a population of

N = 152 = 225 solutions is created, identifying the nth solution of generation k with the
vector X Nk . Solutions are initialized following a uniform distribution bounded by limits be-
tween the allowed ranges for each variable. Three solutions are then selected to perform the
random mutation: Xγk , Xδk , and Xηk to generate a new solution: X̂ik = Xγk + F ( Xδk − Xηk )
with F being a random value between 0 and 2.
The crossover is performed with a probability CR = 0.7. Then, Yik and Xik are
evaluated in the function to be optimized. To be considered part of the new generation, the
one with the best results is chosen. The generations are repeated 100 times, resulting in the
solution exhibiting the lowest value.

4. Results
This section presents the numerical results of the dynamic balancing of the six-bar
mechanism with FCC and DE optimization. To better visualize the influence of the counter-
weights on the procedure, a Pareto front analysis is proposed. This tool allows to restrict
our attention to the set of best solutions and ease the decision-making process.

171
Mathematics 2022, 10, 1830

Table 2 summarizes the physical parameters of the six-bar mechanism shown in

Figure 2. Steel links have been considered with a density of 7800 kg/m3 . The counter-
weights are considered to be made of brass with a density ρcn = 8500 kg/m3 .

Table 2. Parameters of the six-bar mechanism used in the example. Those indicated with a ‘-’ are not
necessary for the numerical analysis.

Link 1 2 3 4 5
n
Mass
0.6935 0.1022 0.9636 0.1825 0.1679
m bn
[kg]
Length
0.19 0.14 0.13416408 0.25 0.23
ln
[m]
Inertia
0.00116161 - 0.00622646 - -
Ixbn
[kg m/s2 ]
Inertia
0.00556534 - 0.00657336 - -
Iybn
[kg m/s2 ]
Inertia
- 0.00066856 - 0.00380360 0.00296204
Izbn
[kg m/s2 ]
Inertia
0.00167596 - 0.00522914 - -
Ixybn
[kg m/s2 ]
CoM 0.08 0.07 0.07751702 0.125 0.115
xbn [m]
CoM 0.03333333 0.0 0.06559133 0.0 0.0
ybn [m]
K x [m] 0.05 - 0.09838699 - -
Ky [m] 0.1 - 0.196677398 - -

A motor located at point A is responsible for actuating the mechanism. Its rotating
speed has been fixed at 500 rpm. Using direct kinematics, it is possible to obtain a sample
of the positions, velocities, and accelerations at each of the basic points of the mechanism.
By replacing all known parameters in the equations of the balancing indexes
(Equations (47) and (48)), it is possible to define the objective function (Equation (49)).
According to the mechanical characteristics of this particular example, the boundaries
considered for the optimization are shown in Equation (51):

−0.16 m xcn , ycn 0.16 m

(51)
0.005 m tcn 0.04 m

4.1. Optimization with Five Counterweights

The DE algorithm was executed until 200 valid solutions (i.e., solutions resulting from
the objective function optimization with values between 0 and 1) were obtained considering
γ as a random value with uniform distribution in the range (0, 1), following the random
search of the hyper-parameters presented in [47].
Figure 4a shows the relationship between the β ShF and the β ShM values for all the
solutions found. Different colors are used to represent the values used for γ in the objective
function f(X) (Equation (49)).

172
Mathematics 2022, 10, 1830

In Figure 4b, the dark points represent the Pareto front while the light ones represent
the dominated solutions.

(a) (b)

Figure 4. Pareto analysis of the optimization objectives: β ShM and β ShF . (a) Optimized β ShM and
β ShF according to γ. (b) Pareto Front of the optimized objectives.

Among all the solutions found in the Pareto front, it is possible to select the one that
is the most appropriate according to the desired type of balancing. To exemplify this
statement, let us consider three different solutions from the Pareto front.
In the first one, a greater importance is given to balancing the ShM (β ShM = 0.235917108,
β ShF = 0.932850297). In the second one, a greater importance is given to balancing the ShF
(β ShM = 0.924195224, β ShF = 0.270900009). Finally, in the third solution, the same importance
is given to balancing both the ShF and ShM (β ShM = 0.580111266, β ShF = 0.558041831). These
solutions are detailed below.
1. In the first solution, a greater importance is given to balancing the ShM. This is
achieved by choosing the minimum value of index β ShM (β ShM = 0. 235917108), which
allows us to obtain an improvement of 76.4% without considering any improvement
of the ShF (β ShF = 0.932850297). The following variables values correspond to this
solution:
xc1 = −0.02611622 yc1 = −0.033186164 tc1 = 0.021370695
xc2 = −0.06 yc2 = −0.012314794 tc2 = 0.039459546
xc3 = −0.06 yc3 = 0.016752092 tc3 = 0.039814955
xc4 = −0.002045161 yc4 = −0.000151565 tc4 = 0.005
xc5 = 0.001614171 yc5 = 0.00639164 tc5 = 0.005106848
2. The second chosen solution in the Pareto front is the one with the minimum value
in β ShF (β ShF = 0.270900009), which achieves an improvement of 72.91% in balancing
the ShM. This choice assigns no importance to the balancing of the ShM (β ShM =
0.924195224). This solution yields to the following variable values:
xc1 = −0.049437705 yc1 = −0.04279353 tc1 = 0.0074319
xc2 = −0.06 yc2 = −0.001485936 tc2 = 0.038953174
xc3 = −0.050813691 yc3 = −0.017724236 tc3 = 0.039660979
xc4 = 0.001669535 yc4 = 0.005996916 tc4 = 0.005
xc5 = 0.000387207 yc5 = 0.013090811 tc5 = 0.005
3. The third chosen solution is the one in the Pareto front where both indexes are
optimized (β ShM = 0.580111266, β ShF = 0.558041831). By using this solution, the ShM
is reduced by 41.99% and the ShF is reduced by 44.2%. It corresponds to the following
variable values:

173
Mathematics 2022, 10, 1830

xc1 = −0.06 yc1 = −0.030069707 tc1 = 0.006021448

xc2 = −0.06 yc2 = −0.00383543 tc2 = 0.04
xc3 = −0.06 yc3 = 0.004650961 tc3 = 0.028206906
xc4 = 0.000606455 yc4 = −0.001322538 tc4 = 0.005416754
xc5 = 0.000545683 yc5 = −0.0000185 tc5 = 0.005
To measure the method’s efficiency, we assessed two parameters: convergence and
computing time. Figure 5 shows the process of convergence of the objective function f ( X )
as a function of generations. Six executions are presented with different γ values. Note
that all executions exhibit values higher than 1.0 at the beginning of the execution (i.e., not
optimized at all). After the 40th generation, all reach convergence (the closer the value to
zero, the better), proving that the DE method actually optimizes the objective function.
Computing time is a major concern in the efficiency of optimization algorithms. To
measure it, we executed 50 times the DE optimization procedure on a desktop computer
with an Intel Core i7 processor (2.40 GHz), 8 GB RAM, and Windows 10 OS. The average
computing time was 15.22 min with a standard deviation of 0.25 min. Being a mechan-
ical design application, the optimization procedure does not require a fast or real-time
computing but rather a practical one.

Figure 5. The convergence of the objective function f ( X ) with the DE method.

4.2. Dimensions of the Counterweights

An analysis is now presented to determine if the proposed boundaries for the opti-
mization are the most suitable or if they should be modified (in case there is the possibility
of changing them due to the mechanical constraints of the mechanism and its surround-
ing space).
Figure 6 shows the box plots of the partial derivatives of the objective function with re-
spect to each of the variables to be optimized xn , yn , and tn for each of the n counterweights
(1 ≤ n ≤ 5) when evaluated at the optimal solutions.
We can state that optimal boundaries have been found when all the values of the
partial derivatives of the objective function are close to zero. In the box plots of Figure 6, it
can be seen that the partial derivatives of the variables x1 , y1 , y2 , y3 , x4 , y4 , t4 , x5 , y5 , and t5
are close to zero, implying that the proposed boundaries allow such variables to reach their
optimal values.
However, note that variables t1 , x2 , t2 , x3 , and t3 have partial derivative values that are
not close to zero. For example, in t1 and even more significantly in x2 and x3 , the values
tend to be greater than zero; hence, it can be deduced that better optimization results could
be obtained if: (1) the thicknesses of counterweight 1 were smaller than the fixed boundary
of 0.005 m (i.e., practically eliminating counterweight 1) and (2) the x-axis positions of
counterweights 1 and 3 were less than the limit of −0.16 m (i.e., moving their CoM in their
–x local axis).

174
Mathematics 2022, 10, 1830

Figure 6. Box plots of partial derivatives of the objective function with respect to each optimization
variable when using five counterweights.

On the other hand, the values of the partial derivatives of variables t2 and t3 tend to
be less than zero, which means that, if possible, it would be convenient to further extend
the boundaries of the optimization of these variables to values greater than 0.04 m.
Note that the information obtained from the Pareto front of the partial derivatives of
the objective function is extremely useful to make decisions concerning the boundaries of
the optimization, allowing us to foresee pertinent changes in the linkages as far as their
mechanical conditions allow it.
Figure 7a shows the histograms of the volumes of the counterweights obtained from
the different optimization results. Together with the analysis of the counterweights’ area
and thickness shown in Figure 7b, it is possible to conclude that counterweight 4 can be
eliminated: it has a very small volume compared to the other counterweights, and its
thickness and area are practically negligible.
In addition, for both counterweights 2 and 3, it can be appreciated that their thickness
tends to remain at the upper limit of the optimization, thus confirming the result of the
partial derivatives analysis: if the mechanical characteristics of the mechanism allow it, it
would be advisable to extend the upper limit of their thicknesses.

175
Mathematics 2022, 10, 1830

(a)

(b)

Figure 7. Dimensional analysis for the five counterweights balancing the six-bar mechanism. (a) Vol-
umes of the five counterweights. (b) Relationship between the area and the thickness for the five
counterweights.

176
Mathematics 2022, 10, 1830

4.3. Implementation
The optimization method and the dimensional analysis conclude with the design of a
solution that effectively reduces the dynamic reactions of the six-bar mechanism.
Let us consider the third case where both ShF and ShM are optimized. Figure 8
shows the proposed implementation. As suggested by the analysis, counterweight 4 has
been eliminated. Note that counterweight 5 is not visible because it exhibits very small
dimensions in relation to the mechanism. Yet, it has been considered for the dynamic
balancing of the mechanism.

Figure 8. Conceptual implementation of the third solution from the Pareto front for the six-bar
mechanism using five counterweights Ci .

4.4. Optimization with Four or Less Counterweights

The previous analysis suggested that not all five counterweights are strictly needed to
balance the six-bar mechanism: counterweight 4 is meaningless and counterweight 5 is so
small that it can be practically eliminated. When seeking a simple solution, i.e., one that
does not involve adding too much volume to the mechanism, it may be useful to reduce
the number of counterweights as long as they still provide acceptable results.
We performed the same analysis for four, three, two, and one counterweights. Figure 9
shows a comparison between the Pareto fronts of the different optimization results as
a function of the number of counterweights. Note that as expected, similar results are
obtained when using five, four, and three counterweights. So, a general conclusion for the
implementation would be to use three counterweights.

Figure 9. Pareto front comparison of the optimization objectives β ShM and β ShF according to the
number of counterweights.

177
Mathematics 2022, 10, 1830

Table 3 summarizes the results obtained. The best results are highlighted in blue.
Again, the best choice depends on the desired balancing: only the ShF, only the ShM,
or both.

Table 3. Summary of the optimization results for the six-bar mechanism balancing.

Counterweights ShF Optimization ShM Optimization ShF and ShM Optimization

Cn % ShF, % ShM % ShF, % ShM % ShF, % ShM
All 5 72.91, 7.59 6.72, 76.41 44.2, 41.99
C1, C2, C3, C4 76.82, 3.59 3.21, 76.97 44.31, 40.73
C1, C2, C3 75.95, 0.5 3.53, 77.21 45.69, 46.81
C2, C3 56.98, 6.52 0.17, 67 41.66, 34.18
C3 44.97, 2.26 1.03, 61 28.69, 30.23

5. Discussion
The implementation of well-designed counterweights allows us to reduce the ShF and
ShM by 76.82% and 77.21%, respectively, when importance is given to either of them and
by 45.69% and 46.81%, respectively, when equal importance is given to both of them.
Table 4 compares our balancing results with others previously reported in the literature.
Note that the comparison is limited to the cases where either the ShF or the ShM are
optimized as not all the studies report the joint optimization. Table 4 includes other types
of mechanisms such as the crank-slider and the four-bar mechanism, which are indeed
simpler structures and thus easier to optimize. Six-bar mechanisms and their balancing are
more complex and less frequently found in the literature because, as explained in Section 2,
they exhibit a higher number of links with some defined by three basic points.

Table 4. Comparison of balancing results for several types of mechanisms and optimization methods.

Reference, Mechanism Optimization % of ShF % of ShM

CC or FCC Algorithm Optimization Optimization
[48], CC crank-slider Differential Evolution 61.42 65.96
[49], CC crank-slider Teaching–Learning 48 44
[50], CC crank-slider Genetic 46 99
[51], FCC crank-slider Differential Evolution 97.76 94.58
[52], CC four-bar Genetic 50 68
[36], CC four-bar Firefly 86.3 83.39
[38], FCC four-bar Gradient Descent 99.70 83.99
[53], CC six-bar Genetic 48.5 32.35
This work, FCC six-bar Differential Evolution 76.82 77.21

Note that the use of FFC to define the mass matrix of a mechanism and thus obtain the
expressions representing the dynamic reactions in its base, in conjunction with an optimiza-
tion algorithm, is a suitable methodology for the complete balancing of mechanisms. Our
previous work in crank-sliders [51] and four-bar [38] mechanisms confirm the efficiency of
the proposed approach. The results obtained surpass those of approaches using Cartesian
coordinates (CC).
The DE optimization method was successfully applied to solve the balancing problem of
a six-bar mechanism. DE is a simple yet robust approach to address multiobjective problems.
Comparison between Pareto fronts has proven to be a useful tool for better visualizing
the impact of each counterweight on the dynamic balancing of the mechanism.
The effectiveness of the analysis of the boundaries by means of box plots of the partial
derivatives of the variables to be optimized together with the histograms of volumes and
relations between area and thickens were also demonstrated. They ease the visualization of
possible improvements on the counterweights and allow us to make useful decisions on
their implementation.

178
Mathematics 2022, 10, 1830

Balancing six-bar mechanisms is highly appreciated in areas such as mobile robotics,

since this type of linkage is used in humanoid robots [54] and bionic legs [55,56] as well as
in rehabilitation engineering [57], in which balancing could minimize the reactions caused
by the devices’ motion, thus improving their performance and usability.

6. Conclusions
This paper has presented a novel approach for the complete (or dynamic) balanc-
ing of mechanisms: the use of fully Cartesian coordinates (FCC) in conjunction with an
optimization method such as Differential Evolution (DE).
Among the main contributions of this paper is the development of the two-dimensional
mass matrix for elements defined by three basic points. To our knowledge, this matrix has
not been proposed so far in the literature and it can be applied to a vast number of more
complex mechanisms that use type of linkages. By using FCC, this work has demonstrated
that even for the most complex mechanisms, it is possible to obtain relatively simple non-
trigonometric equations that define the ShF and ShM and to further optimize their dynamic
balancing with algorithms such as DE with very good results.
As future work, it is expected to continue exploring FCC together with other meta-
heuristic methods to optimize more complex mechanisms in two and three dimensions.
Our approach promises to be highly efficient for optimizing the balancing conditions of
even more complex mechanisms once their mass matrices have been defined.

Author Contributions: Conceptualization, M.A. and M.T.O.-G.; methodology, M.T.O.-G. and C.N.S.;
simulation, M.T.O.-G.; validation, D.U.C.-D., P.V. and R.V.; formal analysis, M.A., R.V. and A.A.G.;
investigation, M.T.O.-G., M.A. and R.V.; writing—original draft preparation, M.T.O.-G.; writing—
review and editing, R.V. and P.V.; visualization, D.U.C.-D. and A.A.G.; supervision, M.A. and R.V. All
authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations
The following abbreviations are used in this manuscript:

CC Cartesian Coordinates
FCC Fully Cartesian Coordinates
ShF Shaking Force
ShM Shaking Moment
CoM Center of Mass
DE Differential Evolution

Appendix A. Mass-Matrix for Individual Linkages

It is necessary to define the mass matrix for each of the linkages in order to assemble
the mass matrix of the entire mechanism.
Given that each counterweight is firmly attached to its corresponding link and that
their position with respect to the local reference system does not change with the mech-
anism’s motion, it is possible to consider five linkages to be optimized, each of them
encompassing a link and a counterweight.
Linkages 1 and 3 have mass matrices consisting of three basic points, while linkages 2,
4, and 5 have mass matrices involving two basic points.
To define the different mass matrices, it is necessary to substitute the terms m, x̄ g , ȳ g ,
Ix , Iy , Ixy , and Iz in the equations that define the mass matrices of two and three points,

179
Mathematics 2022, 10, 1830

considering the contribution of both the link and the counterweight. Thus, for each n
linkage, Equations (A1)–(A7) can be written.

mn = mbn + mcn = mbn + ρπtcn ( xcn

2
+ y2cn ) (A1)

xbn mbn + xcn mcn

x gn = (A2)
mbn + mcn
ybn mbn + ycn mcn
y gn = (A3)
mbn + mcn
1
Ixn = Ixbn + Ixcn = Ixbn + ρπtcn ( xcn
2
+ y2cn )( xcn
2
+ 5y2cn ) (A4)
4
1
Iyn = Iybn + Iycn = Iybn + ρπtcn ( xcn
2
+ y2cn )(5xcn2
+ y2cn ) (A5)
4
Ixyn = Ixybn + Ixycn = Ixybn + ρπtcn ( xcn
2
+ y2cn ) xcn ycn (A6)
Izn = Izbn + Izcn = Ixbn + Iybn + Ixcn + Iycn (A7)

Appendix A.1. Mass Matrix for Linkages 1 and 3

To define the mass matrix of linkage 1, let us consider point C as i, point E as j, and
point D as k. Similarly, for linkage 3, point B is considered as i, point F as j, and point E as k.
Figure A1 shows the distances Kx1 , Ky1 , Kx3 , and Ky3 . Note that for linkage 1, lij = l1
corresponds to the distance CE ¯ while for linkage 2, lij = l2 corresponds to the distance BF.¯

(a) (b)

Figure A1. Linkages defined by three basic points. (a) Linkage 1. (b) Linkage 3.

By using the mass matrix M3P (Equation (24)), the terms of the mass matrix for linkages
n = 1 and n = 3 can be defined according to Equations (A8)–(A13).

180
Mathematics 2022, 10, 1830

Ixbn K2xn 2I Kxn I 2Ixybn Kxn 2Ixybn Iybn πK2xn ρcn tcn xcn
4
en = 2 2
− xbn
2
+ xbn
2
− 2
+ + 2 + 2 2
Kyn ln Kyn ln Kyn Kyn ln Kyn ln ln 4Kyn ln
3πK2xn ρcn tcn xcn
2 5πK2xn ρcn tcn y4cn 2Kxn mbn ybn 2πKxn ρcn
+ 2 l2
y2cn + 2 l2
+ + 2
tcn xcn ycn
2Kyn n 4Kyn n Kyn ln Kyn ln
2πKxn ρcn 2πKxn ρcn 2πKxn ρcn πKxn ρcn tcn xcn
4
+ tcn y3cn − 2
3
tcn xcn ycn − 2
tcn xcn y3cn − 2
Kyn ln Kyn ln Kyn ln 2Kyn ln
3πKxn ρcn 5πKxn ρcn tcn y4cn 2m2bn xbn
− 2 2
tcn xcn ycn − −
2
Kyn ln 2
2Kyn ln ln mbn + πln ρcn tcn xcn
2 + πl ρ t y2
n cn cn cn

2m2bn ybn 2πmbn ρcn tcn xbn xcn

2
− −
Kyn mbn + πKyn ρcn tcn xcn
2 + πK ρ t y2
yn cn cn cn l m
n bn + πl ρ t x 2 + πl ρ t y2
n cn cn cn n cn cn cn
2πmbn ρcn tcn xbn y2cn 2πmbn ρcn tcn xcn3
− −
ln mbn + πln ρcn tcn xcn + πln ρcn tcn ycn
2 2 ln mbn + πln ρcn tcn xcn + πln ρcn tcn y2cn
2

2πmbn ρcn tcn xcn

2 y
bn
−
Kyn mbn + πKyn ρcn tcn xcn
2 + πK ρ t y2
yn cn cn cn
2πmbn ρcn tcn xcn
2 y
cn
− (A8)
Kyn mbn + πKyn ρcn tcn xcn + πKyn ρcn tcn y2cn
2

2πmbn ρcn tcn xcn y2cn 2πmbn ρcn tcn ybn y2cn
− −
ln mbn + πln ρcn tcn xcn + πln ρcn tcn ycn
2 2 Kyn mbn + πKyn ρcn tcn xcn
2 + πK ρ t y2
yn cn cn cn
2πmbn ρcn tcn y3cn 2π 2 ρ2cn t2cn xcn
5
− −
Kyn mbn + πKyn ρcn tcn xcn + πKyn ρcn tcn ycn
2 2 ln mbn + πln ρcn tcn xcn + πln ρcn tcn y2cn
2

2π 2 ρ2cn t2cn xcn

4 y
cn 4π 2 ρ2cn t2cn xcn
3 y2
− − cn
Kyn mbn + πKyn ρcn tcn xcn + πKyn ρcn tcn ycn
2 2 ln mbn + πln ρcn tcn xcn + πln ρcn tcn y2cn
2

4π 2 ρ2cn t2cn xcn

2 y3 2π 2 ρ2cn t2cn xcn y4cn
− cn
−
Kyn mbn + πKyn ρcn tcn xcn + πKyn ρcn tcn ycn
2 2 ln mbn + πln ρcn tcn xcn
2 + πl ρ t y2
n cn cn cn
2π 2 ρ2cn t2cn y5cn 5πρcn tcn 4 3πρcn tcn 2 2
− +1+ xcn + xcn ycn
Kyn mbn + πKyn ρcn tcn xcn2 + πK ρ t y2
yn cn cn cn 4ln2 2ln2
πρcn tcn y4cn 2πρcn tcn 3 2πρcn tcn πρcn tcn xcn
4 3πρcn tcn 2 2
+ + x ycn + xcn y3cn + + xcn ycn
4ln2 Kyn ln cn Kyn ln 4Kyn2 2
2Kyn
5πρcn tcn 4
+ 2
ycn
4Kyn

Ixbn K2xn I Kxn 2Ixybn Kxn Ixybn Iybn πK2xn ρcn tcn xcn
4
fn = − 2 2
+ xbn2 + 2
− − 2 − 2 2
Kyn ln Kyn ln Kyn ln Kyn ln ln 4Kyn ln
3πK2xn ρcn tcn xcn
2 5πK2xn ρcn tcn y4cn Kxn mbn ybn πKxn ρcn tcn 2
− 2 l2
y2cn − 2 l2
− − xcn ycn
2Kyn n 4K yn n K yn l n Kyn ln
πKxn ρcn tcn 3 2πKxn ρcn 2πKxn ρcn πKxn ρcn tcn xcn
4
− ycn + 3
tcn xcn ycn + tcn xcn y3cn + (A9)
Kyn ln Kyn ln2 Kyn ln2 2 l
4Kyn n

3πKxn ρcn tcn xcn

2 5πKxn ρcn tcn y4cn m x πρcn πρcn
+ 2
y2cn + 2 l
+ bn bn + 3
tcn xcn + tcn xcn y2cn
2Kyn ln 4Kyn n ln ln ln
5πρcn tcn 4 3πρcn tcn 2 2 πρcn tcn y4cn πρcn tcn ycn 3 πρcn tcn xcn 3
− 2
xcn − 2
xcn ycn − − xcn − ycn
4ln 2ln 4ln2 Kyn ln Kyn ln

181
Mathematics 2022, 10, 1830

Ixbn Kxn I Ixybn πKxn ρcn tcn xcn

4 3πKxn ρcn tcn xcn
2
gn = 2
− xbn
2
− + 2
+ 2
y2cn
Kyn ln Kyn Kyn ln 4Kyn ln 2Kyn ln
5πKxn ρcn tcn y4cn m y πρcn πρcn πρcn tcn ycn 3
+ 2 l
+ bn bn + 2
tcn xcn ycn + tcn y3cn − xcn (A10)
4Kyn n Kyn Kyn Kyn Kyn ln
πρcn tcn xcn 3 πρcn tcn xcn
4 3πρcn tcn 2 2 5πρcn tcn 4
− ycn − 2
− 2
xcn ycn − 2
ycn
Kyn ln 4Kyn 2Kyn 4Kyn

Ixbn K2xn 2Ixybn Kxn Iybn πK2xn ρcn tcn xcn

4 3πK2xn ρcn tcn xcn
2
hn = 2 l2
− 2
+ 2 + 2 l2
+ 2 l2
y2cn
Kyn n K l
yn n l n 4K yn n 2K yn n
5πK2xn ρcn tcn y4cn 2πKxn ρcn 2πKxn ρcn 5πρcn tcn 4
+ 2 l2
− 3
tcn xcn ycn − tcn xcn y3cn + xcn (A11)
4Kyn n Kyn ln2 Kyn ln2 4ln2
3πρcn tcn 2 2 πρcn tcn y4cn
+ 2
xcn ycn +
2ln 4ln2

Ixbn Kxn Ixybn πKxn ρcn tcn xcn

4 3πKxn ρcn tcn xcn
2 5πKxn ρcn tcn y4cn
in = − 2
+ − 2
− 2
y2cn − 2 l
Kyn ln Kyn ln 4Kyn ln 2Kyn ln 4Kyn n
(A12)
πρcn tcn ycn 3 πρcn tcn xcn 3
+ xcn + ycn
Kyn ln Kyn ln

Ixbn πρcn tcn xcn

4 3πρcn tcn 2 2 5πρcn tcn 4
jn = 2
+ 2
+ 2
xcn ycn + 2
ycn (A13)
Kyn 4Kyn 2Kyn 4Kyn

Appendix A.2. Mass Matrix for Linkages 2, 4, and 5

Linkages 2, 4, and 5 are defined by two basic points. For linkage 2, point A is consid-
ered as i while point C is considered as j; for linkage 4, point D is considered as i while point
G is considered as j; for linkage 5, point F is considered as i while point G is considered as j.
Each of them have their origin at point i and the x axis directed toward point j. For each
linkage, ln is the distance between points i and j.
By substituting the corresponding terms in the mass matrix of the elements defined
by two basic points M2P (equation can be found in [38]), it is possible to obtain the terms
of the mass matrix of each of these elements consisting of a counterweight and a linkage
(Equations (A14)–(A17)).
  1
an =mbn + πρcn tcn xcn
2
+ y2cn −
ln (mbn + πρcn tcn ( xcn
2 + y2 ))
cn
    
2.0mbn + 2.0πρcn tcn xcn + ycn
2 2
mbn xbn + πρcn tcn xcn xcn
2
+ y2cn (A14)
      2 
1
+ 2 Ibn + πρcn tcn xcn2 2
xcn + y2cn + πρcn tcn y2cn xcn
2
+ y2cn + 0.5πρcn tcn xcn
2
+ y2cn
ln

1  
bn = mbn xbn + πρcn tcn xcn xcn2
+ y2cn
ln
      2  (A15)
1
− 2 Ibn + πρcn tcn xcn 2 2
xcn + y2cn + πρcn tcn y2cn xcn2
+ y2cn + 0.5πρcn tcn xcn2
+ y2cn
ln
1   
cn = −mbn − πρcn tcn xcn
2
+ y2cn
ln (mbn + πρcn tcn ( xcn + ycn ))
2 2
(A16)
  
mbn ybn + πρcn tcn ycn xcn2
+ y2cn

182
Mathematics 2022, 10, 1830

      2 
1
dn = Ibn + πρcn tcn xcn
2 2
xcn + y2cn + πρcn tcn y2cn xcn
2
+ y2cn + 0.5πρcn tcn xcn
2
+ y2cn (A17)
ln2

References
1. Arakelian, V.; Briot, S. Balancing of Linkages and Robot Manipulators. Mechanisms and Machine Science; Springer International
Publishing: Cham, Switzerland, 2015; Volume 27.
2. Uicker, J.J.; Pennock, G.R.; Shigley, J.E. Theory of Machines and Mechanisms, 5th ed.; Oxford University Press: New York, NY,
USA, 2016.
3. Waldron, K.J.; Kinzel, G.L. Kinematics, Dynamics, and Design of Machinery, 2nd ed.; John Wiley: Hoboken, NJ, USA, 2004.
4. Arakelian, V.; Dahan, M.; Smith, M.A. Historical review of the evolution of the theory on balancing of mechanisms. In Symposium
on History of Machines and Mechanisms Proceedings HMM 2000; Ceccarelli, M., Ed.; Springer: Dordrecht, The Netherlands, 2000;
pp. 291–300.
5. Arakelian, V. Inertia forces and moments balancing in robot manipulators: A review. Adv. Robot. 2017, 31, 717—726. [CrossRef]
6. Wei, B.; Zhang, D. A review of dynamic balancing for robotic mechanisms. Robotica 2021, 39, 55–71. [CrossRef]
7. Fisher, O. Über die reduzierten Systeme und die Hauptpunkte der Glieder eines Gelenkmechanismus und ihre Bedeutung für die
technische Mechanik. Z. Für Angew. Math. Und Phys. 1902, 47, 429–466.
8. Goryachkin, V.P. The forces of inertia and their balancing. In Collection of Scientific Works; Kolos: Moscow, Russia, 1914; pp. 283–418.
9. Yudin, V. The Balancing of Machines and Their Stability; Edition of Academy of Red Army: Moscow, Russia, 1941; 124p.
10. Kreutzinger, R. Über die bewegung des Schwerpunktes beim Kurbelgetriebe. Getriebetechnik 1942, 10, 397–398.
11. Maxwell, R.L. Kinematics and Dynamics of Machinery, 1st ed.; Prentice Hall: Englewood Cliff, NJ, USA, 1960.
12. Smith, M.R.; Maundert, L. Inertia forces in a four-bar linkage. J. Mech. Eng. Sci. 1967, 9, 218–225. [CrossRef]
13. Talbourdet, G.L.; Shepler, P.R. Mathematical solution of 4-bar linkages-IV. Balancing of linkages. Mach. Des. 1941, 13, 73–77.
14. Lanchester, F.W. Engine Balancing. Inst. Automob. Eng. 1914, 8, 195–271. [CrossRef]
15. Root, R.E. Dynamics of Engine and Shaft; John Wiley: New York, NY, USA, 1932.
16. Artobolevsky, I.I.; Edelshtein, B.V. Methods of Inertia Calculation for Mechanisms of Agricultural Machines; Selkhozizdate: Moscow,
Russia, 1935.
17. Gheronimus, Y.L. An approximate method of calculating a counterweight for the balancing of vertical inertia forces. Mechanisms
1968, 3, 283–288. [CrossRef]
18. Doucet, E. Équilibrage dynamique des moteurs en ligne. Tech. Automob. Arienne 1946, 37, 30–31.
19. Emod, I.; Jurek, A. Massenausgleich am Kurbelgetriebe von Sechszylinder-viertakt-V-motoren mit 6 Kurbeln und 60 Zylinder-
winkeln. Period. Polytech. Mech. Eng. 1967, 3–4, 205–211.
20. Semenov, M.V. The synthesis of partly balanced plane mechanisms. Mechanisms 1968, 3, 339–353. [CrossRef]
21. Berkof R.S.; Lowen, G.G. A new method for complete force balancing simple linkages. J. Eng. Ind. 1969, 91B, 21–26. [CrossRef]
22. Smith, M.R. Dynamic analysis and balancing of linkages with interactive computer graphics. Comput. Aided Des. 1975, 7, 15–19.
[CrossRef]
23. Tepper, F.R.; Lowen, G.G. General theorems concerning full force balancing of planar linkages by internal mass redistribution. J.
Eng. Ind. 1972, 94, 789–796. [CrossRef]
24. Berkof, R.S. Complete force and moment balancing of inline four-bar linkages. Mech. Mach. Theory 1973, 8, 397–410. [CrossRef]
25. Wiederrich, J.L.; Roth, B. Momentum balancing of four-bar linkages. J. Manuf. Sci. Eng. 1976, 4, 1289–1295. [CrossRef]
26. Dresig, H.; Jacobi, P. Vollständiger trägheitskraftausgleich von ebenen koppelgetrieben durch anbringen eines zweischlages.
Maschinenbautechnik 1974, 23, 5–8.
27. Feng, G. Complete Shaking Force and Shaking Moment balancing of four types of six-bar linkages. Mech. Mach. Theory 1989, 24,
275–287. [CrossRef]
28. Kochev, I.S. Active balancing of the frame Shaking Moment in high speed planar machines. Mech. Mach. Theory 1992, 27, 53–58.
[CrossRef]
29. de Jong, J.J.; van Dijk, J.; Herder, J.L. A screw based methodology for instantaneous dynamic balance. Mech. Mach. Theory 2019,
141, 267–282. [CrossRef]
30. Acevedo, M.; Orvañanos-Guerrero, M.T.; Velázquez, R.; Arakelian, V. An alternative method for Shaking Force balancing of the
3RRR PPM through acceleration control of the center of mass. Appl. Sci. 2020, 10, 1351. [CrossRef]
31. Meijaard, J.P.; van der Wijk, V. Dynamic balancing of mechanisms with flexible links. Mech. Mach. Theory 2022, 172, 104784.
[CrossRef]
32. Segla, S.; Kalker-Kalman, C.M.; Schwab, A.L. Statical balancing of a robot mechanism with the aid of a genetic algorithm. Mech.
Mach. Theory 1998, 33, 163–174. [CrossRef]
33. Farmani, M.; Jaamialahmadi, A.; Babaie, M. Multiobjective optimization for force and moment balance of a four-bar linkage using
evolutionary algorithms. J. Mech. Sci. Technol. 2011, 25, 2971–2977. [CrossRef]
34. Zamuda, A.; Brest, J.; Boskovic, B.; Zumer V. Differential evolution for multiobjective optimization with self adaptation. In
Proceedings of the 2007 IEEE Congress on Evolutionary Computation, Singapore, 25–28 September 2007.

183
Mathematics 2022, 10, 1830

35. Erkaya, S. Investigation of balancing problem for a planar mechanism using genetic algorithm. J. Mech. Sci. Technol. 2013, 27,
2153–2160. [CrossRef]
36. Bošković, M.; Šalinić, S.; Bulatović, R.; Miodragović, G. Multiobjective optimization for dynamic balancing of four-bar mechanism.
In Proceedings of the 6th International Congress of Serbian Society of Mechanics, Mountain Tara, Serbia, 19–21 June 2017.
37. García de Jalón, J. Twenty-five years of natural coordinates. Multibody Syst. Dyn. 2007, 18, 15–33. [CrossRef]
38. Orvañanos-Guerrero, M.T.; Sánchez, C.N.; Rivera, M.; Acevedo, M.; Velázquez, R. Gradient descent-based optimization method
of a four-bar mechanism using Fully Cartesian coordinates. Appl. Sci. 2019, 9, 4115. [CrossRef]
39. Acevedo, M.; Orvañanos-Guerrero, M.T.; Velázquez, R.; Haro, E. Optimum balancing of the four-bar linkage using Fully Cartesian
coordinates. IEEE Lat. Am. Trans. 2019, 17, 983–990. [CrossRef]
40. Bourbonnais, F.; Bigras, P.; Bonev, I.A. Minimum-time trajectory planning and control of a pick-and-place five-bar parallel robot.
IEEE/ASME Trans. Mechatron. 2015, 20, 740–749. [CrossRef]
41. Wang, D.; Wang, L.; Wu, J.; Ye, H. An experimental study on the dynamics calibration of a 3-DOF parallel tool head IEEE/ASME
Trans. Mechatron. 2019, 24, 2931–2941.
42. Pennock, G.R.; Israr, A. Kinematic analysis and synthesis of an adjustable six-bar linkage Mech. Mach. Theory 2009, 44, 306–323.
[CrossRef]
43. García de Jalón, J.; Bayo, E. Kinematic and Dynamic Simulation of Multibody Systems: The Real-Time Challenge; Springer: New York,
NY, USA, 1994.
44. Storn, R.; Price, K. Differential evolution-a simple and efficient heuristic for global optimization over continuous spaces. J. Glob.
Optim. 1997, 11, 341–359. [CrossRef]
45. Das, A.K.; Mishra, D.; Das, K.; Mallick, P.K.; Kumar, S.; Zymbler, M.; El-Sayed, H. Prophesying the short-term dynamics of
the crude oil future price by adopting the survival of the fittest principle of improved grey optimization and extreme learning
machine. Mathematics 2022, 10, 1121. [CrossRef]
46. Álvarez Gutiérrez, D.; Sánchez Lasheras, F.; Martín Sánchez, V.; Suárez Gómez, S.L.; Moreno, V.; Moratalla-Navarro, F.; Molina de
la Torre, A.J. A new algorithm for multivariate genome wide association studies based on differential evolution and extreme
learning machines. Mathematics 2022, 10, 1024. [CrossRef]
47. Bergstra, J.; Bengio, Y. Random search for hyper-parameter optimization. J. Mach. Learn. Res. 2012, 13, 281–305.
48. Etesami, G.; Felezi, M.E.; Nariman-Zadeh, N. Pareto optimal multi-objective dynamical balancing of a slider-crank mechanism
using differential evolution algorithm. Int. J. Automot. Eng. 2019, 9, 3021–3032.
49. Chaudhary, K.; Chaudhary, H. Optimal design of planar slider-crank mechanism using teaching-learning-based optimization
algorithm. J. Mech. Sci. Technol. 2015, 29, 5189–5198. [CrossRef]
50. Chaudhary, K.; Chaudhary, H. Optimum balancing of slider-crank mechanism using equimomental system of point-masses.
Procedia Technol. 2014, 14, 35–42. [CrossRef]
51. Orvañanos-Guerrero, M.T.; Acevedo, M.; Sánchez, C.N.; Giannoccaro, N.I.; Visconti, P.; Velázquez, R. Efficient balancing
optimization of a simplified slider-crank mechanism. In Proceedings of the 2020 IEEE ANDESCON, Quito, Ecuador, 13–16
October 2020.
52. Chaudhary, K.; Chaudhary, H. Shape optimization of dynamically balanced planar four-bar mechanism. Procedia Comput. Sci.
2015, 57, 519–526. [CrossRef]
53. Belleri B.K.; Kerur, S.B. Balancing of planar six-bar mechanism with genetic algorithm. J. Mech. Energy Eng. 2020, 4, 303–308.
[CrossRef]
54. Hernández, E.; Velázquez, R.; Macías-Quijas, R.; Pissaloux, E.; Giannoccaro, N.I.; Lay-Ekuakille, A. Kinematic computations for
small-size humanoid robot KUBO. ARPN J. Eng. Appl. Sci. 2017, 12, 7311–7320.
55. Xu, K.; Liu, H.; Zhu, X.; Song, Y. Kinematic analysis of a novel planar six-bar bionic leg. Mech. Mach. Sci. 2019, 73, 13–21.
56. Velázquez, R.; Garzón-Castro, C.L.; Acevedo, M.; Orvañanos-Guerrero, M.T.; Ghavifekr, A.A. Design and characterization of a
miniature bio-inspired mobile robot. In Proceedings of the 2021 12th International Symposium on Advanced Topics in Electrical
Engineering, Bucharest, Romania, 25–27 March 2021.
57. Shao, Y.; Xiang, Z.; Liu, H.; Li, L. Conceptual design and dimensional synthesis of cam-linkage mechanisms for gait rehabilitation.
Mech. Mach. Theory 2016, 104, 31–42. [CrossRef]

184
 mathematics

Article
Singularities of Serial Robots: Identification and Distance
Computation Using Geometric Algebra
Isiah Zaplana 1, *, Hugo Hadfield 2 and Joan Lasenby 2

1 Department of Mechanical Engineering, KU Leuven, 3000 Leuven, Belgium

2 Department of Engineering, University of Cambridge, Cambridge CB2 1PZ, UK; hh409@cam.ac.uk (H.H.);
jl221@cam.ac.uk (J.L.)
* Correspondence: isiah.zaplana@kuleuven.be

Abstract: The singularities of serial robotic manipulators are those configurations in which the robot
loses the ability to move in at least one direction. Hence, their identification is fundamental to enhance
the performance of current control and motion planning strategies. While classical approaches entail
the computation of the determinant of either a 6 × n or n × n matrix for an n-degrees-of-freedom
serial robot, this work addresses a novel singularity identification method based on modelling the
twists defined by the joint axes of the robot as vectors of the six-dimensional and three-dimensional
geometric algebras. In particular, it consists of identifying which configurations cause the exterior
product of these twists to vanish. In addition, since rotors represent rotations in geometric algebra,
once these singularities have been identified, a distance function is defined in the configuration
space C , such that its restriction to the set of singular configurations S allows us to compute the
distance of any configuration to a given singularity. This distance function is used to enhance how
the singularities are handled in three different scenarios, namely, motion planning, motion control
Citation: Zaplana, I.; Hadfield, H.; and bilateral teleoperation.
Lasenby, J. Singularities of Serial
Robots: Identification and Distance Keywords: serial robotic manipulators; singularity identification; geometric algebra; rotor group;
Computation Using Geometric distance to a singularity
Algebra. Mathematics 2022, 10, 2068.
https://doi.org/10.3390/ MSC: 15A66; 15A75; 70B15; 70Q05
math10122068

Academic Editors: Higinio Rubio

Alonso, Alejandro Bustos Caballero,
Jesus Meneses Alonso and Enrique
1. Introduction
Soriano-Heras A serial robot manipulator is an open kinematic chain made up of a sequence of rigid
bodies, called links, connected by means of actuated kinematic pairs, called joints, that
Received: 15 April 2022
provide relative motion between consecutive links. At the end of the last link, there is a tool
Accepted: 13 June 2022
or device known as the end-effector. Only two types of joints are considered throughout
Published: 15 June 2022
this work: revolute joints, which only perform rotations, and prismatic joints, which only
Publisher’s Note: MDPI stays neutral perform translations. If joint i is revolute (prismatic), the amount it rotates (translates) is
with regard to jurisdictional claims in encoded by an angle θi (a displacement di ). These scalars are known as the joint variables
published maps and institutional affil- of the robot.
iations. From a kinematic point of view, the end-effector position and orientation (also known
as the pose) can be expressed as a differentiable function f : C → X, where C denotes the
space of joint variables, called the configuration space of the robot, and X is the space of all
positions and orientations of the end-effector with respect to a reference frame, which is
Copyright: © 2022 by the authors.
usually called the operational space. A serial robot is said to have n degrees of freedom
Licensee MDPI, Basel, Switzerland.
This article is an open access article
(DoF) if its configuration can be minimally specified by n variables. For a serial robot, the
distributed under the terms and
number and nature of the joints determine the number of the DoF. For the task of positioning
conditions of the Creative Commons and orientating its end-effector in the three-dimensional space, the manipulators with more
Attribution (CC BY) license (https:// than six DoF are called redundant while the rest are non-redundant.
creativecommons.org/licenses/by/ In order to describe its relative position and orientation, a frame {o, x, y, z} is attached
4.0/). to each joint (Figure 1). The relations between consecutive joint frames are conventionally

Mathematics 2022, 10, 2068. https://doi.org/10.3390/math10122068 185 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 2068

described by homogeneous transformation matrices. In particular, i−1 Ti relates frame {i }

to frame {i − 1} (the first joint frame is related to a fixed reference frame, known as the
world frame). Therefore, the end-effector pose 0 Tn of a robot with n DoF with respect to
the world frame can be represented as:
n −1
0
Tn = 0 T1 1 T2 · · · Tn (1)

with  
R p
0
Tn = , (2)
0 1
where R is a rotation matrix that describes the end-effector orientation with respect to the
world frame, while p is a position vector describing the end-effector position with respect
to the world frame. This description is equivalent to the one provided by f , known as
the kinematic function of the serial robot. Thus, f (q) = x, where x denotes the vector
describing the end-effector pose and q = (q1 , . . . , qn ) denotes the vector whose components
are the joint variables, also known as the configuration of the robot. Clearly, either qi = θi if
joint i is revolute or qi = di if joint i is prismatic.

  

 

  






 


Figure 1. A frame {o, x, y, z} is attached to each joint of the serial robot to describe its relative position
and orientation.

Deriving the kinematic relation defined by f with respect to time, we obtain another
relation:
ẋ = J (q)q̇, (3)
where ẋ denotes the end-effector velocity vector, q̇; the vector of the joint velocities; and J,
the Jacobian matrix of f . If J = [ J1 · · · Jn ], then each column Ji can also be computed as:
 
z i × ( o n − oi )
Ji = if joint i is revolute,
zi
  (4)
zi
Ji = if joint i is prismatic.
0

Definition 1. Given a serial robot with n DoF, a singularity or kinematic singularity is a configu-
ration q ∈ C satisfying ρ( J (q)) < min{n, 6}, where ρ(·) denotes the rank of the matrix argument.
The set of all singular configurations is a subset of C that is usually denoted by S and known as the
singular set.

186
Mathematics 2022, 10, 2068

Using the relation (3), it is easy to see that if q ∈ C is a singularity of a given serial
robot, then the following two statements hold:
• The robot loses at least one degree of freedom or, equivalently, its end-effector cannot
be translated along or rotated around at least one Cartesian direction.
• Finite linear and angular velocities of the end-effector may require infinite joint velocities.
In addition, Gottlieb [1] and Hollerbach [2] have independently proven that any serial
manipulator with n > 2 DoF has singularities. The identification of such singularities is
achieved by solving the following non-linear equation:

det( J (q)) = 0 (5)

if the robot is non-redundant and

det( J (q) J T (q)) = 0 (6)

if it is redundant.
In general, if the serial robot possesses at least one revolute joint, several coefficients
of the Jacobian matrix are non-linear expressions, and thus, neither Equation (5) nor
Equation (6) are easy to formulate and solve. However, for manipulators with a spherical
wrist, a simplification can be made. For these robots, the axes of their last three joints
intersect at a common point, known as the wrist center point, or are parallel (the intersection
point and, hence, the wrist center point, is the point at the infinity). Since the origin of
the frame attached to the end-effector can be placed at the wrist center point, a zero block
appears in J (q) by definition (see Equation (4)). Hence:
 
J11 (q) 0
J (q) = , (7)
J21 (q) J22 (q)

where J11 (q), J21 (q) are blocks of order 3 × (n − 3), J22 (q) is a block of order 3 × 3 and 0
denotes a block of order 3 × 3, whose entries are all zero. Now, Equation (5) is simplified to:

det( J (q)) = det( J11 (q)) det( J22 (q)), (8)

from which the singularities can be obtained as solutions of either det( J11 (q)) = 0 or
det( J22 (q)) = 0. These two equations allow us to decouple the singularities into position
and orientation singularities as follows:
• Position singularities PS = {q ∈ C : det( J11 (q)) = 0};
• Orientation singularities OS = {q ∈ C : det( J22 (q)) = 0}.
Similarly, we can make the same decoupling for redundant robots:
• Position singularities PS = {q ∈ C : rank( J11 (q)) < 3};
• Orientation singularities OS = {q ∈ C : det( J22 (q)) = 0}.

Remark 1. The Jacobian matrix J (q) is represented with respect to the world frame (usually located
at the base of the robot). However, sometimes it is useful to represent J (q) in a different frame B . To
do so, the following identity is used:

J (q)B = BJ (q), (9)

where:  
R0B 0
B= , (10)
0 R0B

with R0B = ( R0B ) T and where R0B denotes the rotation matrix that relates the orientation of B with
respect to the orientation of the world frame.

187
Mathematics 2022, 10, 2068

Singularity identification is one of the fundamental research fields in robot kinemat-

ics since, as stated before, it affects the motion of the robot and its performance when
executing different tasks. Therefore, such identification is fundamental to enhancing the
performance of current control and motion planning strategies by designing approaches to
handle them. For instance, current applications of the subject include the handling of singu-
larities for a robust control architecture in human–robot collaboration [3], the planning of
singularity-free trajectories [4,5] and the smooth trajectory generation for rotating extensible
manipulators or painting robots [6,7]. However, such identification is still problematic. The
majority of the current approaches are based on the computation of either det( J (q)) = 0
or det( J (q) J T (q)) = 0 or on the manipulation of the singular values of J (q) [8–10]; hence,
they are not computationally efficient. In addition, there is no efficient way of computing
how close an arbitrary configuration is to a given singularity (which is fundamental for
defining a threshold from where the strategies to handle them start to work). In this context,
geometric algebra turns out to be very useful. In addition, it is currently applied to several
problems in robot kinematics and geometry [11–15].
There is not much literature regarding the identification of singularities using geomet-
ric algebra, and the majority of the contributions focus on parallel mechanisms. For serial
robots, Corrochano & Sobczyk [16] extend the Lie bracket of two vectors defined in any
Lie algebra to what they call the superbracket of the lines 1 , . . . , 6 , [1 , . . . , 6 ], where the
line i denotes the axis of joint i. For serial robots with six DoF, the main idea is to split
the superbracket into smaller superbrackets, called bracket monomials, that are equated to
zero. The singularities are the solutions of these monomial bracket equations. Following
the same idea, Kanaan et al. [17] define the superbracket in a Grassmann–Caley algebra.
Since the Lie bracket is well defined in every Grassmann–Caley algebra, the superbracket
is also well defined. However, splitting the superbracket into the bracket monomials in this
context is not intuitive and, as a consequence, is not always realizable. In addition, there is
no standard procedure for computing these brackets’ monomials.
For parallel mechanisms, the majority of works [18–22] focus on approaches developed
for some particular parallel robots. The main idea consists of computing, for each leg of the
mechanism, the exterior product of the twists defined by its joints and equating it to zero.
For those legs with less than six actuated joints, combinations of two, three or more legs are
considered. The main problem with these approaches is their lack of generalisability. Each
approach is designed for the specific parallel robot the authors work with.
Huo et al. [23] present a mobility analysis applying conformal geometric algebra and
a singularity analysis using an idea similar to the ones presented in the above-mentioned
contributions. A mobility analysis of overconstrained parallel mechanisms is performed
using Grassmann–Cayley algebra by Chai et al. [24], while Yang and Li [25] propose a
novel identification method for the constraint singularities of parallel robots based on
differential manifolds. Finally, Kim et al. [26] apply conformal geometric algebra to the
identification of the singularities of a particular type of parallel manipulator, the SPS-
parallel manipulator. Several lines and planes are defined using the different joint axes.
Then, the relative positions of different combinations of these geometric entities are studied
to geometrically find the singularities. However, this method cannot be extended to other
classes of parallel or serial robots, nor can it be implemented as an algorithm due to its
complex geometrical nature.
In this paper, a novel approach for singularity identification based on the six-dimensional
and three-dimensional geometric algebras is introduced. It extends the works developed for
parallel robots and reviewed above. In particular, one of the novelties of this method is that
it can be applied to both redundant and non-redundant serial robots of any geometry. We
first model the twists defined by the joint axes as vectors of the six-dimensional geometric
algebra. Then, we manipulate the exterior product of these twists. In addition, this method
can be simplified for serial robots with a spherical wrist using, instead of the six-dimensional
geometric algebra G6 , the three-dimensional geometric algebra G3 . Once the singularities
have been identified and since rotors describe the transformations between arbitrary

188
Mathematics 2022, 10, 2068

multivectors in geometric algebra, a distance function D can be defined in the configuration

space C that can be used to determine the distance of any arbitrary configuration q ∈ C to a
given singularity qs ∈ S . This is the first time, to the best of the authors’ knowledge, that
such a distance function has been defined. It is well-known that there are several indices
that can be used to check whether a given configuration is close to a singularity or not. For
instance, the Jacobian matrix J (q) allows us to define the manipulability index wm as:
#
wm = det( J (q) T J (q)) = σn · · · σ1 , (11)

where 0 ≤ σ1 ≤ · · · ≤ σn are the singular values of J (q). Alternatively, we can also define
the condition number of J (q), wc = σn /σ1 . Clearly, the former is close to zero when the
configuration is close to a singularity, while the value of the latter increases as the robot
approaches to a singular configuration. Although there are several approaches based on
the use of such indices [27,28], none of them define a distance function, and as stated in [29],
they do not provide a realistic measure of how close a singularity is, only whether it is
close or not. On the other hand, Yao et al. [30] propose a different index of closeness to
singularities for planar parallel robots based on the volume of the workspace. Although
interesting, it is still not a distance function, and it cannot be easily applied to serial
robots. Similarly, Nawratil [31] defines a distance function for parallel manipulators of the
Stewart–Gough type. However, it measures how close a given pose of the end-effector is
to a singular pose (i.e., the pose associated with a singular configuration). Hence, such a
distance function is not defined in the configuration space C but in the operational space
X. Finally, Bu [32] defines an angle between the velocity vector associated with one of the
joints and the manifold generated by the others. Again, such an angle acts as a measure of
closeness but not as a distance function, and thus, it does not provide a realistic measure of
how close a singularity is.
The rest of the paper is organized as follows: Section 2 presents an overview of
geometric algebra that will be useful for understanding the proposed contribution. In
Section 3, the novel singularity identification approach and the simplification for serial
robots with a spherical wrist are fully developed, while the novel distance function is
constructed in Section 4. The application of these results to the Kuka LWR 4+, a redundant
serial robot with a spherical wrist, is given in Section 5. Section 6 lists three different
applications where both the singularity identification and the novel distance function can
be applied in order to illustrate their utility. Finally, the conclusions are given in Section 7.

2. Mathematical Preliminaries: Geometric Algebra

One of the main problems of vector spaces is that linear transformations between
them are represented through matrices, which entails a high computational cost when
implemented. To overcome these and related problems, geometric algebra provides an
excellent framework. It was first introduced by W. K. Clifford [33], whom based his own
work on the previous works of H. Grassmann [34] and W. R. Hamilton [35]. Throughout
this section, a brief overview of geometric algebra is presented. More detailed treatments
of the subject can be found in [36–39].

Definition 2. Given two vectors x1 , x2 ∈ Rn , the outer or exterior product of x1 and x2 , x1 ∧ x2 ,

is a new element that can be seen as the oriented area of the parallelogram obtained by sweeping the
vector x1 along x2 . The exterior product is bilinear, associative and anticommutative. In particular,
x ∧ x = 0 for every x ∈ Rn .

The new element defined by the exterior product is called a bivector, and it is defined
to have grade two. By extension, the outer product of a bivector with a vector is known as
a trivector, is denoted by x1 ∧ x2 ∧ x3 and is defined to have grade three. Trivectors can be
seen as the oriented volume obtained by sweeping the bivector x1 ∧ x2 along x3 .

189
Mathematics 2022, 10, 2068

This can be generalized to an arbitrary dimension. Thus,

x1 ∧ x2 ∧ · · · ∧ x k (12)

denotes a k-blade, i.e., an element of grade k. Linear combinations of k-blades are known as
k-vectors, while linear combinations of k-vectors (for different k) are known as multivectors.
In his work [33], Clifford extends the exterior product by adding a scalar product
between vectors, the inner product. He defines the geometric product (also known as the
Clifford product) as follows:

x1 x2 = x1 · x2 + x1 ∧ x2 ( x1 , x2 ∈ Rn ). (13)

Thus, the geometric product between two vectors has two components: the scalar compo-
nent given by the inner product and the bivector component given by the exterior product.
Clearly, it also inherits the associativity and bilinearity of the exterior product.
When applied to an orthonormal basis B = {e1 , . . . , en } of Rn , the geometric product
acts as follows: (
1 for i = j
ei e j = (14)
ei ∧ e j for i = j
Thus, for each 0 ≤ k ≤ n, the set of k-vectors is spanned by:
k =0 {1} (scalars).
k =1 {e1 , . . . , en } (vectors).
k =2 {ei ∧ e j }1≤i< j≤n (bivectors).
k =3 {ei ∧ e j ∧ ek }1≤i< j<k≤n (trivectors).
..
.
k = r {ei1 ∧ · · · ∧ eir }1≤i1 <···<ir ≤n (r-vectors).
..
.
k = n {e1 ∧ · · · ∧ en } (pseudoscalar).
Then, for each 0 ≤ k ≤ n, there are exactly C (n, k ) generators for the set of k-
vectors, and thus, the set of k-vectors defines a vector space with basis Bk = {ei1 ∧ · · · ∧
eik }1≤i1 <···<ik ≤n and dimension C (n, k).

Definition 3. Let Rn denote the real vector space of dimension n. Then, the vector space spanned
by the basis
B = {ei1 ∧ · · · ∧ eir }1≤i1 <···<ir ≤n (15)
0≤r ≤ n

endowed with the geometric product defined in (13) is an algebra over R known as the geometric
algebra (GA) of Rn . Such an algebra is denoted by Gn and has the dimension C (n, 0) + C (n, 1) +
· · · + C (n, n) = 2n .

Remark 2. Since the grading structure of multivectors is a property associated with the exterior
product, the elements of Gn can still be called k-blades, k-vectors and multivectors.

An important family of linear operators in Gn are the grade-k projection operators,

denoted by ·k for 0 ≤ k ≤ n. Applied to an arbitrary multivector A,  Ak projects onto
the grade-k components in A, i.e., it returns the components of A that can be expressed as a
linear combination of {ei1 ∧ · · · ∧ eik }1≤i1 <···<ik ≤n . Obviously, if Ak denotes a k-vector, then
 Ak k = Ak .
Using these operators, general multivectors A ∈ Gn can be expressed as:

A =  A 0 +  A 1 + · · · +  A  n . (16)

190
Mathematics 2022, 10, 2068

Hence, the set of all k-vectors for a given 1 ≤ k ≤ n is a vector subspace of Gn denoted
by Gn k and spanned by Bk = {ei1 ∧ · · · ∧ eik }1≤i1 <···<ik ≤n .
The multivector representation (16) is very useful in defining another important
operator in Gn . This linear operator is known as the reversion operator and is denoted by
the superscript ∼. The reversion is defined over the geometric product of m vectors as:

( a1 · · · a m ) ∼ = a m · · · a1 . (17)

Applied to k-vectors, we have that:

k ( k −1)
+k = (−1)
A 2 Ak (18)

due to the anticommutativity of the exterior product. Finally, since reversion is a linear
operator, the reverse of an arbitrary multivector is:
3 4 3 4 n ( n −1)
A + +···+ A
+= A + =  A +  A −  A + · · · + (−1) 2  A . (19)
0 1 2 n
0 n

Notice that the reversion operator corresponds simply to matrix transposition when a
matrix representation of the n-dimensional algebra is considered.
Finally, another operator of great interest is the dual operator. Every grade-n element
of Gn is of the form α(e1 ∧ · · · ∧ en ) for a scalar α ∈ R. For each α ∈ R, α(e1 ∧ · · · ∧ en ) is
known as the volume element Eα of Gn , while the generator e1 ∧ · · · ∧ en is known as the
pseudoscalar of Gn and is usually denoted by I. Pseudoscalars allow us to define the dual
operator, whose action over a k-vector Ak is:

A∗k = I Ak , (20)

where A∗k is an (n − k)-vector.

Now, let us go back to the bivectors of Gn since they will be fundamental in the
modelling of the rotations in Rn . An important property of these bivectors is that they
always square to a scalar. Therefore, given a bivector B, the unit bivector associated with
B is B = B/| B2 |. Unit bivectors of Gn always square to -1. This allows us to compute the
following series:
∞
(αB )m
exp(αB ) = ∑ , (21)
m =0 m!
where α ∈ R. Expanding Equation (21), we have that:

α2 α3 B
exp(αB ) = 1 + αB − − +···
2 3!
   
α2 α3 (22)
= 1− + · · · + B α − +···
2 3!
= cos(α) + B sin(α).

Equation (22) indicates that exp(αB ) could be related to rotations. Indeed, we have
the following result.

Proposition 1. Let B be a unit bivector and 0 ≤ θ ≤ 2π; then, R = exp(−(θB)/2) =

cos(θ/2) − B sin(θ/2) ∈ Gn 0 + Gn 2 defines a rotation by an angle θ with a rotation plane
represented by B. It acts over an element X ∈ Gn through the sandwiching product:

+
X  = RX R. (23)

Such an element R is termed a rotor.

191
Mathematics 2022, 10, 2068

Rotors satisfy the following properties:

(1) RR+ = 1;
(2) + = (− R) x(− R
Rx R + ) for x ∈ Rn ;
(3) Rxy R+ = Rx RRy + R + for x, y ∈ Rn ;
(4) Given two frames {e1 , . . . , en } and { f 1 , . . . , f n }, there always exists a rotor R that
transforms one into the other, i.e., there exists a rotor R such that ei = R f i R + for
i = 1, . . . , n. This rotor is computed as [36,40]:

1 + e1 f 1 + · · · + e n f n
R= , (24)
1 + e1 f 1 + · · · + e n f n 

where {e1 , . . . , en } denotes the reciprocal frame of {e1 , . . . , en }.

The first property is the analogous version of the property defining the orthogonal
matrices with determinant equal to 1, which are known to represent rotations. The second
property proves that both R and − R encode the same rotation, while the third property is
known as the geometric covariance of rotors.
In general, rotors define a group R with the geometric product as the group product:

+ = 1}.
R = { R ∈ Gn 0 + Gn 2 : R R (25)

Therefore, the product of two different rotors R1 and R2 also encodes a rotation. In
particular, it is the rotation resulting from the composition of the rotations encoded by
R1 and R2 , respectively. In addition, the second property states that R provides a double
covering of the rotation group.
Finally, one of the most important geometric algebras is the spatial geometric algebra
G3 , whose basis is:
{1, e1 , e2 , e3 , e12 , e13 , e23 , I }, (26)
where {e1 , e2 , e3 } is an orthonormal basis of R3 and eij = ei ∧ e j .

3. Identification of Singularities Using Geometric Algebra

Since six degrees of freedom are required to describe the position and orientation
of a rigid body in the three-dimensional space, the more natural way of formulating the
singularity problem is through the six-dimensional geometric algebra G6 , which extends
naturally the three-dimensional algebra G3 introduced in Section 2. Screw theory [41,42]
provides an intuitive and geometrical description of the differential kinematics of serial
and parallel manipulators using six-dimensional vectors. Because of this, throughout this
chapter, some concepts taken from this theory will be employed. This will provide the
initial framework to completely understand the approach introduced in this section.
As stated in the introduction, we are going to work with three-dimensional rigid
motions, i.e., three-dimensional orientation-preserving isometries. They form a Lie group,
called the special Euclidean group, denoted by SE(3). Its associated Lie algebra is:
(  )
Ω v
se(3) = ξ̂ ∈ M4 : ξ̂ = , (27)
0 0

where Ω is a skew-symmetric matrix of order 3, v ∈ R3 , and M4 denotes the vector space

of order 4 square matrices with real entries. Since every skew-symmetric matrix Ω can
be represented as a vector ω, we can express an element ξ̂ ∈ se(3) as a six-dimensional
vector ξ = [ω v] T , termed a twist. Therefore, twists are the infinitesimal generators of rigid
motions via the exponential map, i.e., exp(ξ̂t) = f (t) with f ∈ SE(3). The next theorem is
a fundamental result in screw theory.

Theorem 1 (Chasles, 1830 [43]). Every rigid motion f ∈ SE(3) can be realized as a rotation
around an axis followed (preceded) by a translation along the same axis.

192
Mathematics 2022, 10, 2068

Definition 4. A screw motion consists of a rotation around an axis followed (preceded) by a

translation along the same axis, the screw axis . The ratio between the translational and the
rotational parts of the motion is known as the pitch and is denoted by h. In particular, if a point
is rotated around  by an angle θ = 0 and translated along  by an amount d, then h = d/θ. By
convention, if θ = 0, h = ∞.

˙ θ̇.
Remark 3. For infinitesimal motions, if θ = 0, then the pitch is defined as h = d/

Hence, every rigid motion is a screw motion. Particular cases of screw motions are pure
rotations (pure translations) where the translation (rotation) is the identity or, equivalently,
h = 0 (h = ∞). In addition, every screw motion can be characterized by the triple (, h, q),
where q denotes the magnitude of the motion. If h = ∞, then θ = q and d = hθ, while if
h = ∞, then θ = 0 and d = q. We call this triple the screw associated with the screw motion,
and we denote it using $.

Proposition 2. Given a screw $ = (, h, q) with screw axis , pitch h and magnitude q, there
exists a twist ξ such that the rigid motion it generates is the screw motion associated with $.

Proposition 2 states a correspondence between twists and screws that is useful for
our purposes. In particular, if p is a point on  and v is its direction unit vector, then
 = { p + vλ : λ ∈ R} and we have that:
 
v
ξ=θ for a general screw motion,
p × v + hv
 
v
ξ=θ for a pure rotation, (28)
p×v
 
0
ξ=d for a pure translation.
v

A twist ξ associated with a magnitude 1 screw $ is said to be a unit twist. Hence, any
twist ξ can be seen as a unit twist multiplied by the magnitude of the associated screw axis:
 
v
ξ = θξ U = θ , (29)
p × v + hv

where ξ U is a unit twist. Clearly, ξ is associated with $ = (, h, q), while ξ U is associated
with $ = (, h, 1).

Proposition 3. Let us consider a rigid body performing a screw motion represented by the screw
$ = (, h, q(t)), where the magnitude q(t) is a time-dependent variable. Its velocity during the
screw motion is given by the associated twist ξ, where, now, the pitch is defined as in Remark 3.
In particular:  
v
ξ = θ̇ (t) if θ = 0,
p × v + hv
  (30)
0
ξ = d˙(t) if θ = 0,
v

where, here, θ̇ (t) (d˙(t)) is known as the twist amplitude.

Now, let us consider a serial robot with n DoF, where ω, v denote the angular and
linear velocity vectors of its end-effector. If Equation (3) is expanded, the following is
obtained:

193
Mathematics 2022, 10, 2068

 
v
= J1 (q)q̇1 + · · · + Jn (q)q̇n , (31)
ω
where Ji denotes the i-th column of the Jacobian matrix J. Notice that the right side of
Equation (31) can be seen as the addition of the twists associated with the joints of the robot,
where q̇i plays the role of the twist amplitude and where the linear and angular parts are
interchanged. However, for the sake of formality, let us consider the unit twist ξ i associated
with the i-th joint of the robot (Since, from now on, we are going to work exclusively with
unit twists, the subindex U is omitted for simplicity). Then:
⎧  
⎪
⎪ zi
⎨ q̇i if joint i is revolute
ξ i (q)q̇i =  zi ×
 ( o n − oi ) (32)
⎪
⎪ 0
⎩ q̇i if joint i is prismatic
zi

where, as stated in the introduction, zi is the direction vector of the joint axis, on (oi ) is the
origin of the frame attached to the end-effector (i-th joint) and q̇i = θ̇i if joint i is revolute
and q̇i = d˙i if joint i is prismatic.

Remark 4. The unit twists ξ i (q) defined in Equation (32) are represented with respect to the world
frame, not with respect to the local frame attached to the previous joint. If the unit twists are defined
with respect to a local frame, we need to use the adjoint transformation to represent them with
respect to the world frame. In particular, ξ i (q) = Ad f ξ i (q), where Ad f : R6 → R6 is the adjoint
transformation associated with the rigid motion f , i.e., the rigid motion transforming the reference
frame to the local frame in which the twist is initially represented.

The following is a key result:

Theorem 2 (Tsai, 1999 [44]). Given a serial robot with n DoF:

 
ω
= ξ 1 (q)q̇1 + · · · + ξ n (q)q̇n = [ξ 1 (q) · · · ξ n (q)]q̇, (33)
v

where, again, ω, v denote the angular and linear velocity vectors of the robot’s end-effector and
q̇ = (q̇1 , . . . , q̇n ).

The main advantage of the screw-based Jacobian matrix defined in Equation (33) is that
it allows a geometrical identification of the singularities. Moreover, if an approach based
on geometric algebra is used, an intuitive geometrical and computer-friendly algebraic
identification of the singularities is possible. For that purpose, let us consider the geometric
algebra G6 , where, for every i = 1, . . . , n, the unit twist ξ i (q) can be modelled as a vector.
Indeed, we make the identification ξ i (q) = [ξ i1 · · · ξ i6 ] T with the vector x = ξ i1 e1 + · · · +
ξ i6 e6 ∈ G6 , where e1 , . . . , e6 are the basis vectors of G6 .
The following gives the main result of this section.

Theorem 3. Let ξ i (q) denote the unit twist defined by the i-th joint expressed as a vector of
G6 . Then:
ξ 1 (q) ∧ · · · ∧ ξ 6 (q) = det([ξ 1 (q) · · · ξ 6 (q)])e1 ∧ · · · ∧ e6 . (34)

Theorem 3 can be seen as a particular case of a more general result:

Theorem 4. Let a1 , . . . , an be a set of n vectors of Gn . Then:

a1 ∧ · · · ∧ an = det([ a1 · · · an ])e1 ∧ · · · ∧ en (35)

194
Mathematics 2022, 10, 2068

Theorem 4 can be easily deduced from Equation (4.143) in [36] (p. 108):

F ( I ) = det( F ) I, (36)

where F is a linear transformation in Gn , F : Gn → Gn , and, as stated in Section 2, I denotes

the pseudoscalar of Gn .
In particular, Theorem 4 is true for any set of six vectors a1 , . . . , a6 of G6 , which proves
Theorem 3. Now, the following corollary of Theorem 3 allows us to characterize the
singularities of any serial robot of 6 DoF.

Corollary 1. Given a serial robot with 6 DoF and associated unit twists ξ 1 (q), . . . , ξ 6 (q), then
q ∈ S if, and only if, ξ 1 (q) ∧ · · · ∧ ξ 6 (q) = 0.

Proof. Taking the dual of Equation (34), the following identity is obtained:

(ξ 1 (q) ∧ · · · ∧ ξ 6 (q))∗ = det([ξ 1 (q) · · · ξ 6 (q)]) (37)

and, therefore, the singularities of the serial robot are those configurations q ∈ C verify-
ing that:
(ξ 1 (q) ∧ · · · ∧ ξ 6 (q))∗ = 0. (38)
Now, since for a given non-zero multivector M ∈ Gn , M∗ = 0 if, and only if, M = 0,
Equation (38) can be simplified to:

ξ 1 (q) ∧ · · · ∧ ξ 6 (q) = 0. (39)

Thus, q ∈ S if, and only if, ξ 1 (q) ∧ · · · ∧ ξ 6 (q) = 0.

Remark 5. Notice that what has also been proven in Corollary 1 is that the outer product of n
vectors of Gn is zero if, and only if, the n vectors are linearly dependent.

In addition, Corollary 1 allows us to re-define the singular set as:

S = { q ∈ C : ξ 1 ( q ) ∧ · · · ∧ ξ 6 ( q ) = 0}. (40)

Remark 6. What Theorem 3 states is that, for instance, if two unit twists ξ 1 and ξ 2 satisfy
ξ 1 ∧ ξ 2 = 0, then they represent the same twist, and hence, they generate the same screw motion.
This means that if such a screw motion is a pure translation, then the translational axes are either
parallel or coincident, while if the screw motion is a pure rotation, the rotational axes are coincident
(Since the twists contain the term (zi × (o6 − oi )) for i = 1, 2, they cannot be parallel). Regarding
the kinematic singularities of serial robots, this implies that two prismatic joints whose axes are
either parallel or coincident give rise to a singularity and, equivalently, that two revolute joints
whose axes are coincident give rise to a singularity. This is, in fact, in agreement with what is
known about kinematic singularities since two parallel revolute joint axes do not give rise to a
singularity. Obviously, the same geometrical interpretation can be made for three, four or more unit
twists satisfying that their outer product is zero.

With respect to redundant serial robots, it is clear that, for n > 6, ξ 1 (q) ∧ · · · ∧ ξ n (q) = 0
for any q ∈ C . Hence, Corollary 1 on its own does not allow us to characterize the
singularities of redundant robots. However, this problem can be easily overcome by
studying all the possible combinations of six unit twists in {ξ 1 (q), . . . , ξ n (q)}. We denote
the set of all combinations of six elements that can be drawn from {1, . . . , n} by S. Clearly,
S has C (n, 6) = (n6 ) elements of the form {i1 , . . . , i6 }, where 1 ≤ i1 < · · · < i6 ≤ n and
1 ≤ i ≤ C (n, 6).

195
Mathematics 2022, 10, 2068

Theorem 5. Given a serial robot with n DoF and associated unit twists ξ 1 (q), . . . , ξ n (q), then
q ∈ S if, and only if, for each 1 ≤ i ≤ C (n, 6):

ξ i1 (q) ∧ · · · ∧ ξ i6 (q) = 0, (41)

where {i1 , . . . , i6 } is the i-th element of S.

Proof. It follows that, for q ∈ C :

ξ i1 (q) ∧ · · · ∧ ξ i6 (q) = 0 for every 1 ≤ i ≤ C (n, 6)

(1)
⇐⇒ det([ξ i1 (q) · · · ξ i6 (q)]) = 0 for every 1 ≤ i ≤ C (n, 6) (42)
(2)
⇐⇒ ρ([ξ 1 (q) · · · ξ n (q)]) < 6

where (1) uses Equations (37) and (2) uses the fact that all the minors of order 6 of the
matrix [ξ 1 (q) · · · ξ n (q)] have null determinants. Clearly, ρ([ξ 1 (q) · · · ξ n (q)]) < 6 if, and
only if, ρ( J (q)) < 6, which, in turn, is equivalent to q ∈ S (by Definition 1).

The computation of either Equation (39) for non-redundant robots or Equation (41)
for redundant ones is computationally more efficient than the computation of either
det( J (q)) = 0 or det( J (q) J T (q)) = 0. The main reason for this lies in the computational
complexity of the operations needed to obtain the expressions (39) or (41) with respect to
the complexity of the operations needed for obtaining det( J (q)) = 0 or det( J (q) J T (q)) = 0.
It is clear that the outer product of n vectors of Gn behaves like the addition and product
of real numbers, and hence, it has complexity O(n) + O(n2 ), while the determinant has
complexity O(n3 ) or O(n4 ) depending on the algorithm used. In addition, for redundant
robots, there are two main operations: the product between J (q) and J T (q) and the deter-
minant of the product matrix. This implies that, for this case, the complexity increases to
O ( n3 ) + O ( n4 ).

Special Case: Serial Robots with a Spherical Wrist

Similarly to what happens with the Jacobian matrix J, a simplification can be achieved
for robots that have a spherical wrist. As stated in Section 1, the singularities of these robots
can be decoupled into position and orientation singularities. Position singularities involve
the first n − 3 joints and are computed by studying the rank of the following matrix:
 
J p = s1 · · · s n −3 , (43)

where si = zi × (on − oi ) if joint i is revolute and si = zi if joint i is prismatic. On the other

hand, orientation singularities involve the last three joints and are computed through the
determinant of the following matrix:
 
Jo = zn−2 zn−1 zn . (44)

Now, let us consider the three-dimensional geometric algebra G3 . As proven in

Theorem 4 for n = 3, a1 ∧ a2 ∧ a3 = det([ a1 a2 a3 ])e1 ∧ e2 ∧ e3 for any three vectors
a1 , a2 , a3 ∈ R3 . Hence, analogously to what has been performed before, the following
characterization for the position and orientation singularities can be deduced.

Theorem 6. Given a serial robot with n DoF and a spherical wrist, if either zi × (on − oi ) or zi
are denoted by si for i = 1, . . . , n − 3, then:
• q ∈ C is a position singularity if, and only if, si1 (q) ∧ si2 (q) ∧ si3 (q) = 0 for each 1 ≤
i ≤ C (n − 3, 3), where {i1 , i2 , i3 } is the i-th combination of three elements drawn from
{1, . . . , n − 3}.

196
Mathematics 2022, 10, 2068

• q ∈ C is an orientation singularity if, and only if,

zn−2 (q) ∧ zn−1 (q) ∧ zn (q) = 0. (45)

Proof. The proof is completely analogous to the proof of Corollary 1 and Theorem 5.

Remark 7. Since the last three joint axes either intersect at a single point or are parallel, there is
only one orientation singularity, namely, when these three joint axes are coplanar. This can also be
easily deduced from Equation (45). A schematic representation of such singularity, also called wrist
singularity, is depicted in Figure 2.

ࢠ௡ିଵ

ߠ௡ିଶ
ࢠ௡ିଶ ࢠ௡
ߠ௡

ߠ௡ିଵ

Figure 2. Schematic representation of the wrist singularity.

4. Distance to Singularities
Let q1 , q2 ∈ C be two arbitrary configurations of a serial robot with n DoF and let ξ 1 , . . . , ξ n
be the unit twists associated with its joints. Then, there exist R1 (q1 , q2 ), . . . , Rn (q1 , q2 ), where,
for each 1 ≤ i ≤ n, Ri (q1 , q2 ) is a configuration-dependent rotor in the six-dimensional
geometric algebra G6 such that (Figure 3):

+ i ( q1 , q2 ).
ξ i ( q2 ) = R i ( q1 , q2 ) ξ i ( q1 ) R (46)

The reason why these rotors exist is simple: unit twists are modelled as vectors in
G6 and there always exists a rotor relating any pair of vectors in any geometric algebra
Gn . In particular, there is always a rotor relating the same unit twist ξ in two different
configurations q1 , q2 .

  
  

 
 





 

 


Figure 3. Rotor Ri relating the twist ξ i in two different configurations q1 and q2 .

197
Mathematics 2022, 10, 2068

Now, let qs ∈ S denote a singularity of a serial robot. As explained in the previous

section, if the serial robot has a spherical wrist, then qs only involves a maximum of two
or three joints and, therefore, two or three unit twists. If, conversely, the robot has no
spherical wrist, then it can involve a maximum of six joints. Let us suppose, without loss
of generality, that a given singularity qs involves the joints i1 , . . . , ir with associated unit
twists ξ i1 (qs ), . . . , ξ ir (qs ) for 2 ≤ r ≤ 6. Then, for any configuration q ∈ C , there exist
Ri1 (q, qs ), . . . , Rir (q, qs ) such that:

+ i (q, qs ) for each 1 ≤ j ≤ r.

ξ i j (qs ) = Ri j (q, qs )ξ i j (q) R (47)
j

The notation chosen for these rotors expresses a configuration dependence that is not
a functional dependency, i.e., there is no analytical expression for these rotors with q as
a variable.
Now, it is clear that Ri j (q, qs ) = 1 if, and only if, q = qs for every j = 1, . . . , r. However,
since for each j, Ri j (q, qs ) does not define a function on q, a distance function cannot be
defined. However, the measure of how close a given configuration q is to a singularity can
be set as:
q ≈ qs ⇐⇒ Ri j (q) ≈ 1 for every j = 1, . . . , r. (48)

Example 1. Let qs ∈ S be a singularity of a serial robot that only involves the second and third
joints. Then, for any configuration q ∈ C , there exist R2 (q, qs ) and R3 (q, qs ) such that:

+2 (q, qs ),
ξ 2 (qs ) = R2 (q, qs )ξ 2 (q) R
(49)
+3 (q, qs ).
ξ 3 (qs ) = R3 (q, qs )ξ 3 (q) R

Therefore, q is close to qs if, and only if:

,
R2 (q, qs ) ≈ 1
(50)
R3 (q, qs ) ≈ 1

and is singular if, and only if: ,

R2 (q, qs ) = 1
(51)
R3 (q, qs ) = 1
where, in general, R2 (q, qs ) = R3 (q, qs ).

These rotors can be constructed in many different ways. The easiest way consists of
considering, for each 1 ≤ i ≤ n, the frame {i } attached to joint i and constructed from ξ i .
This three-dimensional frame varies with the configuration q. Hence, for two different
configurations q1 and q2 , there are two frames {i } attached to joint i. As shown in Section 2,
we can recover the three-dimensional rotor that transforms one of the frames into the other.
Since each frame {i } depends continuously on the configuration q, the rotor Ri (q) is a
continuous function defined as follows:

Ri : C → R
(52)
q  → Ri ( q )

Thus, these configuration-dependent rotors exhibit a functional dependency on the

configuration, which allows us to define a distance function. Such distance is based on the
norm of a multivector X ∈ Gn , defined by the relation:
3 4
 X 2 = X X + . (53)
0

To prove that  ·  is a norm, the following two lemmas are necessary.

198
Mathematics 2022, 10, 2068

3 4
+ ∈ R+ = [0, +∞).
Lemma 1. For any given multivector X ∈ Gn , X X
0

Proof. According to Equation (19), it follows that:

3 4 3 4 3 4
X+= X + + X + +···+ X + . (54)
0 1 n

Then:
n n 3 4
+=
XX ∑ ∑  X i + .
X (55)
j
i =0 j =0

Now, note that, for each i = 1, . . . , n,  X i is a i-vector, i.e., it only contains terms of
grade i. The geometric product of two k-vectors (with different k) is stated as follows [36]
(p. 103):
Ar Bs =  Ar Bs |r−s| +  Ar Bs |r−s|+2 + · · · +  Ar Bs r+s , (56)
5 3 46
Therefore, it is clear that  X i X + = 0 for i = j. Thus:
j 0

3 4 n 3 3 44
+
XX = ∑ +
 X i X . (57)
0 i 0
i =0

Now, each  X i can be expanded as follows:

C (n,i )
 X i = ∑ α j (i )e j1 · · · e ji , (58)
j =1

where, for every 1 ≤ j ≤ C (n, i ) = (ni), α j (i ) ∈ R and e j1 · · · e ji are the base elements of
Gn i . Therefore:
3 4 C (n,i )
+ = ∑ α j (i ) e j · · · e j
X (59)
i 1
i
j =1

and, thus:

3 4 C (n,i ) C (n,i )
+ = ∑ α j (i ) e j · · · e j ∑ α j (i ) e j · · · e j
 X i X 1 i i 1
i
j =1 j =1
(60)
C (n,i ) C (n,i )
= ∑ ∑ α j (i )αk (i )e j1 · · · e ji eki · · · ek1 ,
j =1 k =1
7 8
where, clearly, e j1 · · · e ji eki · · · ek1 0 = δjk with δjk the Kronecker delta. Then:

3 3 44 C (n,i )
+
 X i X = ∑ α2j , (61)
i 0
j =1

which, for every 1 ≤ i ≤ n, is a positive scalar. This implies that the sum of Equation (57) is
also a positive scalar.

Lemma 2. Given three strictly positive real numbers a1 , a2 , a3 ∈ R+ \ {0}, the following proper-
ties hold:
√ √ √
• a + a2 − a3 ≤ a1 + a2 ;
√ 1 √ √ √
• a1 + a2 + a3 ≤ a1 + a2 if, and only if, a3 ≤ 2 a1 a2 .

Proof. Both properties can be obtained by a straightforward computation.

199
Mathematics 2022, 10, 2068

3 4
+ is a norm
Proposition 4. The function  ·  : Gn → R+ defined by the identity  X 2 = X X
0
in Gn , i.e.,
(i)  X  ≥ 0 for all X ∈ Gn . In particular,  X  = 0 if, and only if, X = 0.
(ii) λX  = |λ| X  for all X ∈ Gn and λ ∈ R.
(iii)  X + Y  ≤  X  + Y  for all X, Y ∈ Gn (usually known as the triangle inequality).

Proof. 3 4
(i) + is equivalent to:
Given a multivector X, identity  X 2 = X X
0
"3 4
X = ± +
XX . (62)
0

Thus, it is clear by Lemma 1 that the positive branch of Equation (62) is well defined
and that  X  ≥ 0. In particular, if  X  = 0, then:
"3 4 3 4 n 3 3 44
+
XX + = 0 =⇒
= 0 =⇒ X X ∑ +
 X i X = 0, (63)
0 0 i 0
i =0

where all the terms of the last equation are positive by Lemma 1 and, thus, all of them
are equal to zero. Now, note that each addend is the geometric product of an i-vector
with its reverse. Therefore, if such product is zero, the corresponding i-vector must be
zero. Since all the terms are zero, all the i-vectors that form X are zero, and thus, X
is zero.
(ii) If λ ∈ R and X ∈ Gn , then:
"3 "3 4 4
λX  = +)
(λX )(λ X
= +
λ2 X X
0 0
" 3 4 "3 4 (64)
(1)
= λ2 X X+ = |λ| XX+ = |λ| X ,
0 0

where (1) uses the linearity of the grade-0 projection operator (as stated in Section 2).
(iii) Given two different multivectors X and Y, they can be expanded as linear combina-
tions of the basis elements of Gn as follows:
2n
X= ∑ αi e j1 · · · e ji ,
i =0
(65)
2n
Y= ∑ βi e j1 · · · e ji .
i =0

Now, it follows that:

2n
X+Y = ∑ (αi + βi )e j1 · · · e ji (66)
i =0

and hence:
0
# 1 2n
1 (1)
X + Y = ( X + Y )( X + Y ) 0 = 2 ∑ (αi + β i )2
∼
i =0
0
1 (67)
1 2n 2n 2n
1
=1
1 ∑ α2i + ∑ β2i + 2 ∑ αi β i ,
2 i =0 i =0 i =0
A B C

200
Mathematics 2022, 10, 2068

where (1) uses Lemma 1, while A, B and C are just notations given to simplify the
different manipulations. Since A, B > 0 (If either A, B are equal to zero, then either
X = 0 or Y = 0, which will make the condition  X + Y  ≤  X  + Y  trivial):
0 0
1 2n 1 2n
√ (1) √ √ 1 1
A + B + C ≤ A + B = 2 ∑ α2i + 2 ∑ β2i =  X  + Y , (68)
i =0 i =0

where (1) uses the first or second property of Lemma 2 depending

√ on whether C < 0
or C > 0. It only remains to check if C > 0, 2C ≤ 2 AB. Indeed, the previous
inequality is equivalent to that C2 ≤ AB. Now:

2n 2n 2n 2n 2n 2n 2n
AB = ∑ α2i ∑ β2i = ∑ ∑ α2i β2j = ∑ α2i β2i + ∑ ∑ α2i β2j ,
i =0 i =0 i =0 j =0 i =0 i =0 j =0
j  =i
 2 (69)
2n 2n 2n 2n
C =
2
∑ αi β i = ∑ α2i β2i +∑ ∑ αi β i α j β j
i =0 i =0 i =0 j =0
j  =i

and, thus, C2 ≤ AB turns to:

2n 2n 2n 2n 2n 2n
∑ α2i β2i + ∑ ∑ αi βi α j β j ≤ ∑ α2i β2i + ∑ ∑ α2i β2j , (70)
i =0 i =0 j =0 i =0 i =0 j =0
j  =i j  =i

which is equivalent to:

2n 2n 2n 2n
∑ ∑ αi β i α j β j ≤ ∑ ∑ α2i β2j (71)
i =0 j =0 i =0 j =0
j  =i j  =i

which, in turn, is equivalent to:

2n 2n 2n 2n
0≤ ∑ ∑ α2i β2j − ∑ ∑ αi βi α j β j
i =0 j =0 i =0 j =0
j  =i j  =i
n n n n n n
1 2 2 2 2 1 2 2 2 2 2 2

2 i∑ ∑ αi β j + 2 ∑ ∑ αi β j − ∑ ∑ αi β i α j β j
=
=0 j =0 i =0 j =0 i =0 j =0
(72)
j  =i j  =i j  =i
n n
1 2 2
= ∑ ∑ ( α i β j − α j β i )2 .
2 i =0 j =0
j  =i

Since this last inequality is always true, the triangle inequality is also true.

Now, a distance function D can be defined for rotors.

Theorem 7. The function D : R × R → R+ defined by the identity D ( R1 , R2 ) =  R1 − R2  is

a distance in R, i.e.,
(i) D ( R1 , R2 ) ≥ 0 for all R1 , R2 ∈ R. In particular, D ( R1 , R2 ) = 0 if, and only if, R1 = R2 ;
(ii) D ( R1 , R2 ) = D ( R2 , R1 ) for all R1 , R2 ∈ R;
(iii) D ( R1 , R3 ) ≤ D ( R1 , R2 ) + D ( R2 , R3 ) for all R1 , R2 , R3 ∈ R.

201
Mathematics 2022, 10, 2068

Proof. The proof is straightforward and uses the fact that  ·  is a norm. Given two
different rotors R1 and R2 :
(i) D ( R1 , R2 ) =  R1 − R2  ≥ 0. In particular:

(1)
D ( R1 , R2 ) = 0 ⇐⇒  R1 − R2  = 0 ⇐⇒ R1 − R2 = 0 ⇐⇒ R1 = R2 , (73)

where (1) uses the first property of a norm.

(ii) We have that:
#
D ( R1 , R2 ) =  R1 − R2  = ( R1 − R2 )( R1 − R2 )∼ 0
# (74)
= ( R2 − R1 )( R2 − R1 )∼ 0 =  R2 − R1  = D ( R2 , R1 ).

(iii) Given a third rotor R3 , we have that:

D ( R1 , R3 ) =  R1 − R3  =  R1 − R2 + R2 − R3 
(1) (75)
≤  R1 − R2  +  R2 − R3  = D ( R1 , R2 ) + D ( R2 , R3 ),

where (1) uses the third property of a norm.

As stated before, the end-effector pose of a serial robot and the pose of each one of
its joints are described by the configuration-dependent rotors R(q) and Ri (q), respectively.
Thus, one can be tempted to extend the distance function D to C as follows:

D : C × C → R+
(76)
D (q1 , q2 ) =  R(q1 ) − R(q2 )

This function verifies all the requirements of a distance function with the exception of:

D (q1 , q2 ) = 0 ⇐⇒ q1 = q2 . (77)

The reason is simple: a given pose of the end-effector can be associated with up to
16 different configurations if the serial robot is non-redundant and an infinite number if
it is redundant. In particular, this means that R(q1 ) = R(q2 ) with q1 = q2 . However, this
problem can be overcome as follows:
• For each joint i, denote by Ci the configuration space of the subchain formed by the
first i joints. It is clear that, if the robot has n degrees of freedom, Ci ⊂ C for every
1 ≤ i ≤ n. Then, the following set of functions can be defined:

Di : C i × C i → R +
(78)
Di (q1 , q2 ) =  Ri (q1 ) − Ri (q2 )

where, as stated before, Ri is the rotor that describes the pose of joint i. Again, these
functions are not distance functions for the same reason as D (Equation (76)) is not a
distance function.
• The function:

D : C × C → [0, +∞)
(79)
D (q1 , q2 ) = D1 (q11 q21 ) + · · · + Dn (q1n , q2n )

where q1i (q2i ) denotes the first i coordinates of the configuration vector q1 (q2 ), defines
a distance function in C .

202
Mathematics 2022, 10, 2068

Proof. Since, for each 1 ≤ i ≤ n, Di satisfies the requirements (ii ) and (iii ) of a
distance function, it is clear that D also satisfies them. In addition, Di (q1i , q2i ) ≥ 0
for each 1 ≤ i ≤ n and q1i , q2i ∈ Ci . Therefore, D (q1 , q2 ) ≥ 0 for arbitrary q1 , q2 ∈ C .
Finally, if D (q1 , q2 ) = 0, then, since any term of Equation (79) is a positive scalar, it can
be deduced that Di (q1i , q2i ) = 0 for every 1 ≤ i ≤ n. Thus, q1 and q2 not only have
the same end-effector pose, but the same pose for of each of its joints, which clearly
implies that q1 = q2 .

This distance function can be restricted to S just by considering the joints involved in
a given singularity qs .

Definition 5. Let qs ∈ S be a singularity of a serial robot that involves joints i1 , . . . , ir . Then, the
function D : C × S → R+ is defined by the expression:

D (q, qs ) = Di1 (qi1 , qsi ) + · · · + Dir (qir , qsir ), (80)

1

where, for each i1 ≤ k ≤ ir , Dk is the function defined in (78) and is a distance function in C .

5. Application to the Serial Robot Kuka Lwr 4+

To show the advantages of the proposed method, an illustrative example is developed
in this section, making use of the Kuka LWR 4+, an anthropomorphic robotic arm with
seven degrees of freedom and a spherical wrist. It is schematically depicted in Figure 4.
Since it has a spherical wrist, its singularities can be decoupled into position and orientation
singularities. Hence, Theorem 6 can be applied in order to find out such singularities. The
computations with the vectors of G3 have been carried out using the Clifford Multivector
Toolbox of MATLAB [45].

‫଺ݖ‬ ߠ଺
‫ݖ‬ହ
‫ݔ‬ଷ
‫ݔ‬ଶ ‫ݖ‬ଵ ‫ݔ‬ହ
ߠଷ ‫ݖ‬ଷ ߠ଻
ߠହ
‫ݖ‬ଶ ‫ݔ‬ଵ ‫ݖ‬ସ ‫ݔ‬ସ
ߠସ ‫଻ݔ‬
ߠଶ
‫଻ݖ‬
ߠଵ

Figure 4. Schematic representation of the Kuka LWR 4+.

With respect to the position singularities, the following system of C (4, 3) = 4 equations
should be solved:
⎫
(z1 × (o7 − o1 )) ∧ (z2 × (o7 − o2 )) ∧ (z3 × (o7 − o3 )) = 0 ⎪ ⎪
⎪
⎪
(z1 × (o7 − o1 )) ∧ (z2 × (o7 − o2 )) ∧ (z4 × (o7 − o4 )) = 0 ⎬
(81)
(z1 × (o7 − o1 )) ∧ (z3 × (o7 − o3 )) ∧ (z4 × (o7 − o4 )) = 0 ⎪ ⎪
⎪
⎪
⎭
(z × (o − o )) ∧ (z × (o − o )) ∧ (z × (o − o )) = 0
2 7 2 3 7 3 4 7 4

203
Mathematics 2022, 10, 2068

where, as computed in [46], we have that:

z1 × (o7 − o1 )
⎡ ⎤
−400c2 s1 − 390s4 (c1 s3 + c3 s1 s2 ) − 390c2 c4 s1
⎣
= 400c1 c2 − 390s4 (s1 s3 − c1 c3 s2 ) + 390c1 c2 c4 ⎦,
0

z2 × (o7 − o2 )
⎡ ⎤
−c1 (400s2 + 390c4 s2 − 390c2 c3 s4 )
= −s1 (400s2 + 390c4 s2 − 390c2 c3 s4 ) ⎦,
⎣
A1
(82)
z3 × (o7 − o3 )
⎡ ⎤
c2 s1 (390c4 s2 − 390c2 c3 s4 ) − s2 (390s4 (c1 s3 + c3 s1 s2 ) + 390c2 c4 s1 )
=⎣ −s2 (390s4 (s1 s3 − c1 c3 s2 ) − 390c1 c2 c4 ) − c1 c2 (390c4 s2 − 390c2 c3 s4 ) ⎦,
c1 c2 (390s4 (c1 s3 + c3 s1 s2 ) + 390c2 c4 s1 ) + c2 s1 (390s4 (s1 s3 − c1 c3 s2 ) − 390c1 c2 c4 )

z4 × (o7 − o4 )
⎡ ⎤
(390c4 s2 − 390c2 c3 s4 )(c1 c3 − s1 s2 s3 ) − c2 s3 (390s4 (c1 s3 + c3 s1 s2 ) + 390c2 c4 s1 )
= ⎣(390c4 s2 − 390c2 c3 s4 )(c3 s1 + c1 s2 s3 ) − c2 s3 (390s4 (s1 s3 − c1 c3 s2 ) − 390c1 c2 c4 )⎦,
A2

where

A1 = c1 (400c1 c2 − 390s4 (s1 s3 − c1 c3 s2 ) + 390c1 c2 c4 ) + s1 (400c2 s1

+ 390s4 (c1 s3 + c3 s1 s2 ) + 390c2 c4 s1 ),
A2 = (c1 c3 − s1 s2 s3 )(390s4 (s1 s3 − c1 c3 s2 ) − 390c1 c2 c4 ) − (390s4 (c1 s3 + c3 s1 s2 )
+ 390c2 c4 s1 )(c3 s1 + c1 s2 s3 ),

and where ci = cos(θi ) and si = sin(θi ). However, in order to simplify these expressions,
the system of Equation (81) is expressed with respect to the frame attached to the fourth
joint of the Kuka LWR 4+. To do so, a relation analogous of relation (9) is applied. Here,
instead of pre-multiplying by the corresponding rotation matrix, the system of Equation (81)
is multiplied by the three-dimensional rotor R that performs the rotation between the frame
attached to the end-effector and the frame attached to the fourth joint. For instance, the
first equation of the system (81) becomes:

+ = 0,
R(z1 × (o7 − o1 )) ∧ (z2 × (o7 − o2 )) ∧ (z3 × (o7 − o3 )) R (83)

which, using the geometric covariance property for rotors introduced in Section 2, becomes:

+ ∧ R(z2 × (o7 − o2 )) R
R(z1 × (o7 − o1 )) R + ∧ R(z3 × (o7 − o3 )) R
+ = 0. (84)

Therefore, the system of Equation (81) becomes:

⎫
a1 ∧ a2 ∧ a3 = 0 ⎪
⎪
⎪
⎪
a1 ∧ a2 ∧ a4 = 0 ⎬
(85)
a1 ∧ a3 ∧ a4 = 0 ⎪
⎪
⎪
⎪
a ∧a ∧a =0⎭ 2 3 4

204
Mathematics 2022, 10, 2068

where

a1 = (−10c2 s3 (40c4 + 39))e1 + (400c2 s3 s4 )e2 + (400c2 c3 + 390s2 s4 + 390c2 c3 c4 )e3 ,

a2 = (10c3 (40c4 + 39))e1 + (−400c3 s4 )e2 + (10s3 (40c4 + 39))e3 ,
(86)
a3 = (390s4 )e3 ,
a4 = (−390)e1 .

Now, the system of Equation (85) becomes:

⎫
0=0 ⎪
⎪
⎪
⎪
⎬
40c2 s4 + 39s2 c3 s24 + 39c2 c4 s4 = 0
(87)
c2 s3 s24 = 0 ⎪
⎪
⎪
⎪
⎭
c3 s24 = 0

which clearly has two different solutions:

• s4 = 0 or, equivalently, q4 = 0.
• c2 = c3 = 0 or, equivalently, q2 = ± π2 and q3 = ± π2 .
These two solutions correspond to the position singularities of the Kuka LWR 4+.
With respect to the orientation singularities, there is only one equation to solve:

z5 ∧ z6 ∧ z7 = 0. (88)

Again, the expression of each zi for i = 5, 6, 7 can be simplified by expressing those

vectors with respect to the frame attached to the fourth joint. Thus, Equation (88) becomes:

e2 ∧ (−s5 e1 − c5 e3 ) ∧ (c5 s6 e1 + c6 e2 − s5 s6 e3 )
= (−s5 e2 ∧ e1 − c5 e2 ∧ e3 ) ∧ (c5 s6 e1 + c6 e2 − s5 s6 e3 ) (89)
(1)
= −s25 s6 e1 ∧ e2 ∧ e3 − c25 s6 e1 ∧ e2 ∧ e3 = −s6 e1 ∧ e2 ∧ e3 = 0,

where (1) uses the anticommutativity of the outer product. Clearly, the last expression of
Equation (89) is zero if, and only if, s6 = 0 or, equivalently, if, and only if, q6 = 0. Thus, the
Kuka LWR 4+ only has one orientation singularity (the wrist singularity, as explained in
Remark 7).
Finally, the distance function defined in Definition 5 can be applied to any of the
already obtained singular configurations. Let us consider, for instance, the position sin-
gularity q4 = 0. Then, the distance between an arbitrary configuration q ∈ C and this
singularity is given by the expression:

D (q, qs ) =  R4 (q) − R4 (qs ), (90)

where qs denotes the singular configuration q4 = 0 and R4 is the rotor defining the pose of
the fourth joint of the Kuka LWR 4+.
In particular, R4 can be found as explained in Section 2. Indeed, if {e1 , e2 , e3 } denotes
the orthogonal basis defined by the world frame and { f 1 , f 2 , f 3 } (respectively, { f 1 , f 2 , f 3 }),
the orthogonal basis defined by the frame attached to the fourth joint under the effect of
configuration q (respectively, singular configuration qs ), then:

1 + e1 f 1 + e2 f 2 + e3 f 3
R4 ( q ) = ,
1 + e1 f 1 + e2 f 2 + e3 f 3 
(91)
1 + e1 f 1 + e2 f 2 + e3 f 3
R4 ( q s ) = ,
1 + e1 f 1 + e2 f 2 + e3 f 3 

205
Mathematics 2022, 10, 2068

where {e1 , e2 , e3 } is the reciprocal frame of {e1 , e2 , e3 }. Since {e1 , e2 , e3 } is also an orthonor-
mal set of vectors, such a reciprocal frame is:

e 1 = e1 ,
e 2 = e2 , (92)
e 3 = e3 .

Thus, Equation (91) turns to:

1 + e1 f 1 + e2 f 2 + e3 f 3
R4 ( q ) = ,
 1 + e1 f 1 + e2 f 2 + e3 f 3 
(93)
1 + e1 f 1 + e2 f 2 + e3 f 3
R4 ( q s ) = .
1 + e1 f 1 + e2 f 2 + e3 f 3 

Evaluating Equation (93) for the KUKA LWR 4+, we obtain:

a 1 + a 2 e1 ∧ e2 + a 3 e1 ∧ e3 + a 4 e2 ∧ e3
R4 ( q ) = # ,
a21 + a22 + a23 + a24
(94)
b + b2 e1 ∧ e2 + b3 e1 ∧ e3 + b4 e2 ∧ e3
R4 ( q s ) = 1 # ,
b12 + b22 + b32 + b42

where {e1 ∧ e2 , e1 ∧ e3 , e2 ∧ e3 } are the basic bivectors of G3 and

a1 = c2 s3 + c4 s1 s3 − c4 c3 c1 s2 + s3 s4 c1 + s4 s1 s2 c3 + s4 c1 c2 + c2 c4 s1 ,
a2 = c2 s1 s4 − c4 c1 s3 − c4 c3 s1 s2 − c1 c2 c4 + s4 s1 s3 − s4 s2 c1 c3 ,
a3 = s2 s4 + c2 c3 c4 + c3 s1 + c1 s2 s3 ,
a4 = c4 s2 − c2 c3 s4 − c1 c3 + s1 s2 s3 ,
(95)
b1 = c2 s3 + s1 s3 − c3 c1 s2 + c2 s1 ,
b2 = −c1 s3 − c3 s1 s2 − c1 c2 ,
b3 = c2 c3 + c3 s1 + c1 s2 s3 ,
b4 = s2 − c1 c3 + s1 s2 s3 .

Therefore, by Proposition 4 and the decomposition used in the Proof of Lemma 1, the
distance of an arbitrary configuration q to the position singularity q4 = 0 is given by:
#
D (q, qs ) = ( a1 − b1 )2 + ( a2 − b2 )2 + ( a3 − b3 )2 + ( a4 − b4 )2 , (96)

where
ai bi
ai = # and bi = # . (97)
a21 + a22 + a23 + a24 b12 + b22 + b32 + b42

6. Handling of Singularities
Once the set of singular configurations S has been identify, several methods can be
applied to handle the singularities. The detailed treatment of this topic is beyond the scope
of this work. However, in order to show the possibilities of the distance function proposed
in Section 4, we comment on three different situations, namely, motion planning, motion
control and bilateral teleoperation. In each one of these situations, the distance function
defined in Definition 5 plays an important role for handling the singularities.

206
Mathematics 2022, 10, 2068

6.1. Singularity Handling in Motion Planning

Motion planning consists of programming collision-free motions for a given robotic
manipulator from a start position to a goal position among a collection of static obstacles.
The subset of robot configurations that do not cause collision with such obstacles is termed
the free-of-obstacles configuration space, and it is denoted by Cfree . The main methods used
for motion planning can be grouped into three categories:
• Potential field methods, where a differentiable real-valued function U : C → R, called
the potential function, is defined. Such a function has an attractive component that
pulls the trajectory towards the goal configuration and a repulsive component that
pushes the trajectory away from the start configuration and from the obstacles.
• Sampling-based multi-query methods, where a roadmap is constructed over Cfree . The
nodes represent free-of-obstacles configurations, while the edges represent feasible
local paths between those configurations. Once the roadmap is constructed, a search
algorithm finds out the best solution trajectory by selecting and joining the local paths
through an optimization process.
• Sampling-based single-query methods, where tree-structure data are constructed by
searching new configurations (nodes) in Cfree and connecting them through local paths
(edges). Its main difference with respect to the multi-query methods is that, while the
multi-query methods work in two steps (construction of the roadmap and searching
of a solution trajectory), in the single-query methods, both steps are taken together.
Each new configuration added to the set of nodes is connected by a local path and
evaluated in order to check its feasibility.
For any method of these three categories, the distance function D defined in Definition 5
can be applied to construct solution trajectories that also avoid the singularities. Indeed:
• For a potential field method, it is sufficient to add a repulsive component that pushes
the trajectory, not only away from obstacles, but also away from singularities. To do
so, the most efficient method is to define, for each singularity qs , a quadratic repulsive
component as follows:
⎧  
⎨ κ 1 1 2
− if D (q, qs ) ≤ D0
Ur,qs (q) =
⎩ 2 D (q, qs ) D0 (98)
0 if D (q, qs ) > D0

where D0 is set as a threshold for the distance D and κ ∈ R.

• For a sampling-based method with multiple queries, it is sufficient to remove from the
roadmap those nodes associated with singular configurations. During the construction
of the roadmap, each configuration q ∈ C is evaluated to determine whether q is free-
of-obstacles or not. Similarly, the idea is to evaluate each q ∈ C in order to determine
whether q is close to a singularity or not. To speed up the process, both evaluations
can be carried out together:
(1) Select a value D0 > 0 that will work as a threshold.
(2) Given a discretization of the configuration space C , each q of this discretization is
evaluated to check whether:
* It is free-of-obstacles;
* It is far from any singularity. This can be achieved simply by evaluating
whether D (q, qs ) > D0 or D (q, qs ) ≤ D0 ;
(3) If q is free-of-obstacles and far from any singularity, then it can be added to the
set of nodes of the roadmap.
• For a sampling-based method with a single query, the approach is completely analo-
gous to the one used for methods with multiple-queries due to the similarities between
both categories.

207
Mathematics 2022, 10, 2068

6.2. Singularity Handling in Motion Control

Motion control consists of making the end-effector of a robot follow a time-varying
trajectory specified within the manipulator workspace. A typical Inverse Dynamics Control
scheme (depicted as a block diagram in Figure 5) can be described as:
• An input, i.e., the desired or target configuration qd together with its velocity q̇d ;
• A controller based on the dynamical model of the robot:

τ = M(q)q̈ + C (q, q̇)q̇ + g(q) (99)

where M (q) denotes the inertia matrix of the robot, C (q, q̇) denotes the matrix of
Coriolis and centrifugal forces and g(q) denotes the gravity vector.
• An output, i.e., the vector of torques τ, that is sent to the robot to perform the de-
sired motion;
• The robot executes the motion and updates the vectors q and q̇;
• The robot sends such updated vectors to the controller (also known as the feedback
of the system).

qd τ q, q̇
Controller Robot

Figure 5. Standard motion control scheme.

To handle the singularities, a restriction can be defined inside the controller:

(1) The target configuration qd is entered into the controller.
(2) qd is checked in order to determine whether it is close to a singularity or not:
– Select a threshold value D0 > 0.
– Evaluate the condition D (qd , qs ) > D0 for each singularity qs .
– If the evaluation returns yes, then τ can be computed from qd using the dynamical
model (Equation (99)) and sent to the robot. Otherwise, qd is substituted by
qd + D0 qd and evaluated again.
A block diagram of this scheme is depicted in Figure 6.

qd τ
D (qd , qs ) > D0 Dynamical model
yes

no
qd = qd + D0 qd

Figure 6. Proposed control scheme in presence of singularities.

6.3. Singularity Handling in Bilateral Teleoperation

Teleoperated robotic systems are characterized by a robot that executes the move-
ments/actions commanded by a human operator. Any high-level or planning decision
is made by a human user, while the robot is responsible for their mechanical implemen-
tation [47]. Teleoperation systems are often, at least conceptually, split into two parts: a
local manipulator and a remote manipulator. The first one refers to the device moved by
the human operator, while the second refers to the robot or robot system that performs
the action.

208
Mathematics 2022, 10, 2068

According to the information flow direction, the teleoperation may be unilateral or

bilateral. In unilateral teleoperation, the local manipulator sends position or force data to
the remote manipulator and only receives, as feedback, visual information from the remote
scene. However, in bilateral teleoperation, the position or force data are also sent from the
remote manipulator in addition to the visual information.
In a bilateral teleoperation system, some strategies for handling of kinematic singu-
larities can make use of the distance function D defined in Definition 5. For instance, the
following scheme could be applied:
(1) Select a value D0 > 0 that will work as a threshold.
(2) The local manipulator sends a pose (force) p ( f ) to the remote manipulator.
(3) The controller of the remote manipulator obtains the associated configuration q.
(4) The distance D (q, qs ) is computed for each qs ∈ S .
(5) If, for some qs , D (q, qs ) < D0 , then the remote manipulator controller computes a
reaction force f s in the same direction of the motion but with inverse sense.
(6) The remote manipulator sends this force f s to the local manipulator.
(7) The human operator will not be able to move the local manipulator in such a direction
implied by f s and, thus, the singularity qs will never be reached.

7. Conclusions
This paper proposes a novel singularity identification method for arbitrary serial
robots based on the six-dimensional geometric algebra G6 . For non-redundant serial robots,
we take the six unit twists ξ 1 , . . . , ξ 6 associated with the joints, and we model them as
vectors of G6 . Hence, the problem reduces to find the configurations causing the exterior
product ξ 1 (q) ∧ · · · ∧ ξ 6 (q) to vanish since, as proven in Corollary 1, ξ 1 (q) ∧ · · · ∧ ξ 6 (q) = 0
if, and only if, q ∈ S . Analogously, for a redundant robot with n DoF, we consider the
C (n, 6) different combinations of six unit twists taken from {ξ 1 , . . . , ξ n }, and we find
the configurations causing all the exterior products of the form ξ j1 (q) ∧ · · · ∧ ξ j6 (q) for
1 ≤ j ≤ C (n, 6) to vanish.
For serial robots with a spherical wrist, a simplification is possible. For these ma-
nipulators, the singularities are of two types: position singularities and orientation sin-
gularities. The former are identified as the configurations causing the exterior products
si1 (q) ∧ si2 (q) ∧ si3 (q) to vanish for 1 ≤ i ≤ C (n − 3, 3), where si j is the linear velocity
component of the unit twist ξ i j and is modelled as a vector of G3 , while the latter are
identified as the configuration causing the exterior product zn−2 (q) ∧ zn−1 (q) ∧ zn (q) to
vanish, where zi is the i-th joint axis and, again, is modelled as a vector of G3 . Thus, the
simplification consists of evaluating the exterior product of three vectors in G3 instead of
six vectors in G6 .
Once the singularities are identified, a distance function is defined such as its restriction
to the singular set S , defined in Definition 5, is also a distance function that allows us
to check how far an arbitrary configuration q is to a singularity. This distance function
exploits the fact that between any two vectors x, y ∈ Gn , there always exists a rotor R such
that y = Rx R. +
The advantages of the strategy introduced in this work are clear. First, it is a computer-
friendly approach that avoids the computation of the determinant of an order 6 × n (for
non-redundant robots) or n × n (for redundant robots) matrix and the Jacobian matrix J. In
addition, the novel distance function defined in Definition 5 can be used to improve the
performance of current control schemes or motion planning algorithms, which, as seen in
the introduction, is still a hot research topic in robotics.

Author Contributions: Conceptualization, I.Z. and J.L.; methodology, I.Z.; investigation, I.Z. and
H.H.; writing—original draft preparation, I.Z. and H.H.; writing—review and editing, I.Z., H.H. and
J.L.; supervision, J.L. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.

209
Mathematics 2022, 10, 2068

Institutional Review Board Statement: Not applicable.

Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Gottlieb, D. Robots and Topology. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), San
Francisco, CA, USA, 7–10 April 1986; pp. 1689–1691.
2. Hollerbach, J. Optimum kinematic design for a seven degree of freedom manipulator. In Robotics Research: The Second International
Symposium; Hanafusa, H., Inoue, H., Eds.; MIT Press: Cambridge, MA, USA, 1985; pp. 215–222.
3. Carmichael, M.; Khonasty, R.; Aldini, S.; Liu, D. Human Preferences in Using Damping to Manage Singularities During Physical
Human-Robot Collaboration. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Paris,
France, 31 May–31 August 2020; pp. 10184–10190.
4. Lopez-Franco, C.; Diaz, D.; Hernandez-Barragan, J.; Arana-Daniel, N.; Lopez-Franco, M. A Metaheuristic Optimization Approach
for Trajectory Tracking of Robot Manipulators. Mathematics 2022, 10, 1051. [CrossRef]
5. Thananjeyan, B.; Tanwani, A.; Ji, J.; Fer, D.; Patel, V.; Krishnan, S.; Goldberg, K. Optimizing Robot-Assisted Surgery Suture Plans
to Avoid Joint Limits and Singularities. In Proceedings of the International Symposium on Medical Robotics (ISMR), Atlanta, GA,
USA, 3–5 April 2019; pp. 1–7.
6. Dupac, M. Smooth trajectory generation for rotating extensible manipulators. Math. Methods Appl. Sci. 2018, 41, 2281–2286.
[CrossRef]
7. Wang, X.; Zhang, D.; Zhao, C.; Zhang, H.; Yan, H. Singularity analysis and treatment for a 7R 6-DOF painting robot with
non-spherical wrist. Mech. Mach. Theory 2018, 126, 92–107. [CrossRef]
8. Ratajczak, J.; Tchoń, K. Normal forms and singularities of non-holonomic robotic systems: A study of free-floating space robots.
Syst. Control. Lett. 2020, 138, 104661. [CrossRef]
9. Almarkhi, A.; Maciejewski, A. Singularity Analysis for Redundant Manipulators of Arbitrary Kinematic Structure. In Proceedings
of the 16th International Conference on Informatics in Control, Automation and Robotics (ICINCO), Prague, Czech Republic,
29–31 July 2019; pp. 42–49.
10. Sharifi, H.; Black, W. Identification Algorithm to Determine the Trajectory of Robots with Singularities. arXiv 2019,
arXiv:1911.06632.
11. Zhu, G.; Wei, S.; Zhang, Y.; Liao, Q. A Novel Geometric Modeling and Calculation Method for Forward Displacement Analysis of
6-3 Stewart Platforms. Mathematics 2021, 9, 442. [CrossRef]
12. Hadfield, H.; Wei, L.; Lasenby, J. The forward and inverse kinematics of a Delta Robot. In Advances in Computer Graphics;
Magnenat-Thalmann, N., Stephanidis, C., Wu, E., Thalmann, D., Sheng, B., Kim, J., Papagiannakis, G., Gavrilova, M., Eds.;
Springer International Publishing: Berlin/Heidelberg, Germany, 2020; pp. 447–458.
13. Thiruvengadam, S.; Tan, J.; Miller, K. A Generalised Quaternion and Clifford Algebra Based Mathematical Methodology to Effect
Multi-stage Reassembling Transformations in Parallel Robots. Adv. Appl. Clifford Algebr. 2021, 31, 39. [CrossRef]
14. Breuils, S.; Tachibana, K.; Hitzer, E. New Applications of Clifford’s Geometric Algebra. Adv. Appl. Clifford Algebr. 2022, 32, 1–17.
[CrossRef]
15. Hitzer, E.; Lavor, C.; Hildenbrand, D. Current survey of Clifford geometric algebra applications. Math. Methods Appl. Sci. 2022, in
press . [CrossRef]
16. Corrochano, E.; Sobczyk, G. Applications of Lie algebras and the algebra of incidence. In Geometric Algebra with Applications in
Science and Engineering; Corrochano, E., Sobczyk, G., Eds.; Birkhäuser Boston: Boston, MA, USA, 2001; pp. 252–277.
17. Kanaan, D.; Wenger, P.; Caro, S.; Chablat, D. Singularity analysis of lower mobility parallel manipulators using Grassmann–Cayley
algebra. IEEE Trans. Robot. 2009, 25, 995–1004. [CrossRef]
18. Tanev, T. Singularity analysis of a 4-DOF parallel manipulator using geometric algebra. In Advances in Robot Kinematics:
Mechanisms and Motion; Lennarčič, J., Roth, B., Eds.; Springer: Dordrecht, The Netherlands, 2006; pp. 275–284.
19. Chai, X.; Xiang, J. Mobility Analysis of Limited-Degrees-of-Freedom Parallel Mechanisms in the Framework of Geometric
Algebra. ASME J. Mech. Robot. 2016, 8, 41005.
20. Yao, H.; Chen, Q.; Chai, X.; Li, Q. Singularity analysis of 3-RPR parallel manipulators using geometric algebra. Adv. Appl. Clifford
Algebr. 2017, 27, 2097–2113. [CrossRef]
21. Chai, X.; Li, Q. Analytical mobility analysis of Bennett linkage using geometric algebra. Adv. Appl. Clifford Algebr. 2017,
27, 2083–2095. [CrossRef]
22. Ma, J.; Chen, Q.; Yao, H.; Chai, X.; Li, Q. Singularity analysis of the 3/6 Stewart parallel manipulator using geometric algebra.
Math. Methods Appl. Sci. 2018, 41, 2494–2506. [CrossRef]
23. Huo, X.; Sun, T.; Song, Y. A geometric algebra approach to determine motion/constraint, mobility and singularity of parallel
mechanism. Mech. Mach. Theory 2017, 116, 273–293. [CrossRef]

210
Mathematics 2022, 10, 2068

24. Chai, X.; Li, Q.; Ye, W. Mobility analysis of overconstrained parallel mechanism using Grassmann-Cayley algebra. Appl. Math.
Model. 2017, 51, 643–654. [CrossRef]
25. Yang, S.; Li, Y. Classification and analysis of constraint singularities for parallel mechanisms using differential manifolds. Appl.
Math. Model. 2020, 77, 469–477. [CrossRef]
26. Kim, J.; Jeong, J.; Park, J. Inverse kinematics and geometric singularity analysis of a 3-SPS/S redundant motion mechanism using
conformal geometric algebra. Mech. Mach. Theory 2015, 90, 23–36. [CrossRef]
27. Huo, L.; Baron, L. The joint-limits and singularity avoidance in robotic welding. Ind. Robot. 2008, 35, 456–464. [CrossRef]
28. Yahya, S.; Moghavvemi, M.; Mohamed, H. Singularity avoidance of a six degree of freedom three dimensional redundant planar
manipulator. Comput. Math. Appl. 2012, 64, 856–868. [CrossRef]
29. Siciliano, B.; Sciavicco, L.; Villani, L.; Oriolo, G. Robotics: Modelling, Planning and Control; Springer Publishing Company:
Berlin/Heidelberg, Germany, 2008.
30. Yao, H.; Li, Q.; Chen, Q.; Chai, X. Measuring the closeness to singularities of a planar parallel manipulator using geometric
algebra. Appl. Math. Model. 2018, 57, 192–205. [CrossRef]
31. Nawratil, G. Singularity distance for parallel manipulators of Stewart Gough type. In Advances in Mechanism and Machine Science;
Uhl, T., Ed.; Springer International Publishing: Berlin/Heidelberg, Germany, 2019; pp. 259–268.
32. Bu, W. Closeness to singularities of robotic manipulators measured by characteristic angles. Robotica 2016, 34, 2105–2115.
[CrossRef]
33. Clifford, W.; Smith, H.; Tucker, R. Mathematical Papers by William Kingdon Clifford-Edited; Macmillan: London, UK, 1882.
34. Grassmann, H. Ausdehnungslehre-Extension Theory (English Re-Edition); American Mathematical Society: Providence, RI, USA, 2000.
35. Hamilton, W.R. Elements of Quaternions; Longmans, Green: London, UK, 1866.
36. Doran, C.; Lasenby, A. Geometric Algebra for Physicists; Cambridge University Press: Cambridge, UK, 2003.
37. Dorst, L.; Fontijne, D.; Mann, S. Geometric Algebra for Computer Science: An Object-Oriented Approach to Geometry; Morgan Kaufmann
Publishers Inc.: Burlington, MA, USA, 2007.
38. Perwass, C. Geometric Algebra with Applications in Engineering; Springer Publishing Company, Incorporated: Berlin/Heidelberg,
Germany, 2009.
39. Hildenbrand, D. Foundations of Geometric Algebra Computing; Springer Publishing Company, Incorporated: Berlin/Heidelberg,
Germany, 2012.
40. Lavor, C.; Xambó-Descamps, S.; Zaplana, I. A Geometric Algebra Invitation to Space-Time Physics, Robotics and Molecular Geometry;
SRMA/Springerbriefs; Springer: Berlin/Heidelberg, Germany, 2018.
41. Murray, R.; Li, Z.; Shankar-Sastry, S. A Mathematical Introduction to Robotic Manipulation; CRC Press: Boca Raton, FL, USA, 1994.
42. Davidson, J.; Hunt, K. Robots and Screw Theory: Applications of Kinematics and Statics to Robotics; Oxford University Press: Oxford,
UK, 2004.
43. Chasles, M. Note sur les propriétés générales du système de deux corps semblables entr’eux. Bull. Sci. Math. Astron. Phys.
Chemiques 1830, 14, 321–326.
44. Tsai, L. Robot Analysis: The Mechanics of Serial and Parallel Manipulators; John Wiley and Sons: Hoboken, NJ, USA, 1999.
45. Sangwine, S.; Hitzer, E. Clifford Multivector Toolbox (for MATLAB). Adv. Appl. Clifford Algebr. 2017, 27, 539–558. [CrossRef]
46. Zaplana, I.; Claret, J.; Basanez, L. Kinematic analysis of redundant robotic manipulators: applications to Kuka LWR 4+ and ABB
Yumi. Rev. Iberoam. Autom. Inform. Ind. 2018, 15, 192–202. [CrossRef]
47. Basañez, L.; Suárez, R. Teleoperation. In Springer Handbook of Automation; Nof, S., Ed.; Springer: Berlin/Heidelberg, Germany,
2009; pp. 449–468.

211
 mathematics

Article
Mathematical Analysis of a Low Cost Mechanical Ventilator
Respiratory Dynamics Enhanced by a Sensor Transducer (ST)
Based in Nanostructures of Anodic Aluminium Oxide (AAO)
Jesús Alan Calderón Chavarri 1,2, *, Carlos Gianpaul Rincón Ruiz 2 , Ana María Gómez Amador 3 ,
Bray Jesús Martin Agreda Cardenas 2 , Sebastián Calero Anaya 2 , John Hugo Lozano Jauregui 2,4 ,
Alexandr Toribio Hinostroza 4 and Juan José Jiménez de Cisneros y Fonfría 2

1 Angewandte Nanophysik, Institut Für Physik, Technische Universität Ilmenau, 98693 Ilmenau, Germany
2 Engineering Department, Pontificia Universidad Católica del Perú, Lima 15088, Peru;
rinconr.carlos@pucp.pe (C.G.R.R.); b.agreda@pucp.edu.pe (B.J.M.A.C.); sebastian.calero@pucp.pe (S.C.A.);
john.lozanoj@pucp.edu.pe (J.H.L.J.); juanjose.cisneros@pucp.pe (J.J.J.d.C.y.F.)
3 Departament of Mechanical Engineering, Universidad Carlos III de Madrid, 28911 Madrid, Spain;
amgomez@ing.uc3m.es
4 Mechatronic Department, Northen (Artic) Federal University Named after M.V. Lomonosov,
Arkhangelsk 163002, Russia; alexandrtoribio@mail.ru
* Correspondence: alan.calderon@pucp.edu.pe

Abstract: Mechanical ventilation systems require a device for measuring the air flow provided to a
Citation: Chavarri, J.A.C.; Ruiz, patient in order to monitor and ensure the correct quantity of air proportionated to the patient, this
C.G.R.; Gómez Amador, A.M.; device is the air flow sensor. At the beginning of the COVID-19 pandemic, flow sensors were not
Cardenas, B.J.M.A.; Anaya, S.C.;
available in Peru because of the international supply shortage. In this context, a novel air flow sensor
Lozano Jauregui, J.H.; Hinostroza,
based on an orifice plate and an intelligent transducer was developed to form an integrated device.
A.T.; Jiménez de Cisneros y Fonfría,
The proposed design was focused on simple manufacturing requirements for mass production in a
J.J. Mathematical Analysis of a Low
developing country. CAD and CAE techniques were used in the design stage, and a mathematical
Cost Mechanical Ventilator
Respiratory Dynamics Enhanced by a
model of the device was proposed and calibrated experimentally for the measured data transduction.
Sensor Transducer (ST) Based in The device was tested in its real working conditions and was therefore implemented in a breathing
Nanostructures of Anodic circuit connected to a low-cost mechanical ventilation system. Results indicate that the designed
Aluminium Oxide (AAO). air flow sensor/transducer is a low-cost complete medical device for mechanical ventilators that is
Mathematics 2022, 10, 2403. able to provide all the ventilation parameters by an equivalent electrical signal to directly display
https://doi.org/10.3390/ the following factors: air flow, pressure and volume over time. The evaluation of the designed
math10142403 sensor transducer was performed according to sundry transducer parameters such as geometrical
Academic Editor: James M. Buick parameters, material parameters and adaptive coefficients in the main transduction algorithm; in
effect, the variety of the described results were achieved by the faster response time and robustness
Received: 8 April 2022
proportionated by transducers of nanostructures based on Anodic Aluminum Oxide (AAO), which
Accepted: 19 June 2022
enhanced the designed sensor/transducer (ST) during operation in intricate geographic places, such
Published: 8 July 2022
as the Andes mountains of Peru.
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in Keywords: air flow medical sensor; emergency air flow sensor; low-cost air flow sensor; nanostructures;
published maps and institutional affil- COVID-19
iations.

MSC: 65K10

Copyright: © 2022 by the authors.

Licensee MDPI, Basel, Switzerland.
1. Introduction
This article is an open access article
distributed under the terms and
Monitoring has become indispensable for mechanical ventilators in order to provide
conditions of the Creative Commons feedback to medical staff regarding treatment. Moreover, ventilation parameters can be
Attribution (CC BY) license (https:// used for internal control tasks such as adaptive feedback which improves the equipment
creativecommons.org/licenses/by/ performance and response to disturbances. Thus, appropriated values of physical variables
4.0/). and artificial breathing could be achieved by monitoring the ventilation curves in real

Mathematics 2022, 10, 2403. https://doi.org/10.3390/math10142403 213 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 2403

time [1]. In this sense, a flow sensor is a device used for air flow measurement, and its
working concept is commonly based on an induced pressure drop across an orifice which
can be measured and related to the air flow value in a breathing circuit over time [2].
Design considerations for air flow measurement in industrial applications are well
established in the steady-state condition [3]; nevertheless, the air flow provided by the
developed low-cost mechanical ventilator has a particular periodic waveform in a transient
condition. Typical medical air flow sensors measure the pressure drop using static pressure
taps connected to a transducer. However, this represents a risk of leaking polluted air
from the patient, which is a harmful risk given the current pandemic context. To counter
this, modern medical air flow sensors are connected to electrical transducers to generate
electrical values analogous to values of static pressure [4]. Having a sensor connected to
an electrical transducer is quite important for the measured variables, as their electrical
equivalent values can be stored in the main control system of the ventilator. On the other
hand, any kind of virus can be reduced, something that cannot be achieved with typical
airflow connectors. Furthermore, air flow sensors cannot provide air pressure and volume
and other parameters, which require monitoring, are measured using an additional sensor
and calculated at the processing stage.
In this study, a novel air flow sensor/transducer was proposed as an integrated design
that is able to measure the air flow in a breathing circuit in the time domain and estimate
the air pressure and volume parameters with the mathematical model developed and
calibrated experimentally for a specific low-cost mechanical ventilator based on cams [1].
This proposed medical device directly provides the equivalent electrical signals of air
pressure, volume and flow after a transduction stage required due to the nonlinearity of
the air flow and pressure drop [5].
During the COVID-19 pandemic, many low-cost mechanical ventilation systems based
on the compression of an airbag were developed [6], but they were required to measure
the generated air flow rate to the patient [7]. However, air flow sensors were not easy to
resource in Peruvian markets due to the international supply shortage during this period
of the pandemic. Moreover, the design and fabrication for this medical application requires
certain considerations, such as those proposed in this research.

2. Materials and Methods

Broadly, the methodology consists of design, fabrication, experimental study and
experimental validation stages. The design was an iterative process, thus in this paper
only the final version which was improved according to the experimental results is shown.
Figure 1 shows an overview of the research flow. In the design stage, sensor architecture and
functions were established. The transduction algorithm correlates the theoretical pressure
difference and ventilation curves shown in the display. A Computational Fluid Dynamics
(CFD) analysis supports an exhaustive study of air airflow measurement in different flow
conditions as well as with different orifice plate dimensions. A sensor prototype and
equipment were specially designed for the experimental study. The experimental data
obtained were used for the mathematical characterization of the sensor in dynamic [8],
steady and real working conditions. Design improvement and model calibration were
carried out based on the CFD simulation and experimental characterization [9].
Finally, the design was improved and fabricated for the experimental validation
where the device measurement, connected to a mechanical ventilator, was compared to the
measurement of a Fluke air flow meter.

2.1. Device Description

Figure 2 shows the sensor/transducer concept. It consists of an orifice plate sensor,
a pressure transducer and a controller. The crossing air flow through the orifice plate
produces a pressure difference related to the air flow value. Then, the pressure transducer
converts the pressure signals into their electric equivalent signals which are sent to the

214
Mathematics 2022, 10, 2403

controller for data treatment in order to determine and display the pressure, volume and
flow over time using an electronic screen [10].

Transduction algorithm design Design improvement

Sensor transducer design CFD Analysis Computational simulation and Experimental validation
Experimental characterization
- Steady state
- Dynamic response
Prototype and experimental - Response to a Mechanical ventilator
Model Calibration
study

METHOD
EXPERIMENTAL STUDY AND DESIGN IMPROVEMENT

Figure 1. Research overview.

Figure 2. Scheme of the designed air flow sensor transducer.

2.2. Device Design

Figure 3 shows the integrated air flow sensor/transducer developed for this particular
transient flow measurement. It consists of a compact air flow sensor based on an orifice plate
and a transduction stage of the measured pressure difference over time. The transduction
stage and pressure difference measurement were integrated in this device in order to
prevent leaks or contaminated air flow. The device measures the air flow through it and
calculates (using a microcontroller Arduino UNO) the breathing variables transduced as
electrical values that are depicted on the electronic screen and sent to the main control
system of the mechanical ventilator.

215
Mathematics 2022, 10, 2403

Figure 3. Design of an integrated air flow sensor/transducer for a breathing circuit.

Figure 4 shows the compact air flow sensor developed for the integrated design, which
is composed of an upstream and downstream pipe welded to an orifice plate. Dimensions
are established based on the analytic and experimental results, which guarantee the pressure
difference measurement over time for this particular medical application with a mechanical
ventilation system OxygenIP.PE, whose nominal working conditions correspond to a
transient flow [11].

Figure 4. Compact air flow sensor developed for the air flow sensor/transducer.

2.3. Mathematical Modelling

The design of the flow sensor was based on the mathematical model analysis as a
consequence of the theoretical equations, which describe the flow behavior through an
orifice plate. Moreover, the analysis was enhanced due to the correlation with polynomial
models that provided adaptive coefficients for the optimal measurements, for which the
process is described in more detail in the following paragraphs.
In Figure 5, a scheme to explain the theoretical model of the flow inside the airflow
circuit is depicted. In which f is the air flow from a source, the section area of the hose is
a, and the flow reduces its section area b while crossing the orifice plate. Thus, a pressure
difference can be measured between points P1 and P2, moreover v1 and v2 represent the
fluid flow speed of the air flow in the direction depicted by the red arrow [11,12].

Figure 5. Scheme of the orifice plate theoretical model.

216
Mathematics 2022, 10, 2403

In order to find the pressure, airflow and volume equations, the Stokes equations were
analyzed using Equation (1), in which, “f ” is the function as the consequence of airflow
and “μ” and “D” are the average velocity and dispersion coefficient, respectively [13,14].

∂f ∂f ∂f2
+μ =D 2 (1)
∂t ∂x ∂x
This expression can be reduced to the Bernoulli model, and with the analysis of the
energy balance, Equation (2) can be obtained [13], in which ρ is the fluid flow density, p1
and p2 are the initial and final referential pressure through the fluid flow path, and v1 and
v2 are the fluid flow speed.
1 1
ρv 2 + p1 = ρv2 2 + p2 (2)
2 1 2
By conducting a static pressure analysis, it is possible to obtain Equation (3), where  is
the volumetric flow rate [13], A1 and A2 are the cross sections area, and the initial and the
final static pressure are p1 and p2 , which are proportionate to the difference of pressure ΔP.
0
1
1 2ΔP
Φ = A2 1   2  (3)
2
 1− A
A
2
1

From here, it is further possible to reduce to the equation by considering the “discharge
coefficient Cd ” so that Equation (4) is obtained [13].
0
1
1 2ΔP
Φ = Cd A2 1   2  (4)
2
 1− A
A
2
1

Finally, a theoretical model can be derived from the expected breathing curve as shown
by Equation (5) [13], in which Ri has the geometrical and material information of the fluid
flow sensor. Equation (5) proposes the relation among the volumetric flow rate Φ with the
pressure difference of ΔP, and τ is the response time.
 
ΔP  t

Φ= 1 − eτ (5)
Ri

2.4. Computational Fluid Dynamics (CFD) Analysis

Sensor Case Setup
The geometry was modelled after the actual state bench design proposal (with an
extended outlet to overcome reverse flow issues). Following Bridgeman’s steps [15], the
sensor case was modelled as axisymmetric and the pressure sensing ports’ geometries were
disregarded to save computational resources and time. Because of this, better and more
structured meshes were obtained. Five geometries were generated in total (one for each
orifice diameter). The meshes were then generated with an element size of 0.00025 m, all
with good or excellent mesh metrics results, according to Ansys [16]. Figure 6 shows the
model for the 10 mm orifice plate of the sensor case, in which A and B keep the distance for
the static pressure orifices.
The main equations that govern this analysis are provided by the Navier–Stokes
equations, mainly the x-y momentum and continuity equations. The energy equation was
not included in the analysis since heat transfer was not one of the concerns of this study.
The SST k-ε turbulence model was selected for this study because it combines the following
2 separate models: the k-ε and k-ε turbulence models. The first model is better at capturing
near-the-wall turbulence effects, and the second model is superior at capturing turbulence
effects away from the walls.

217
Mathematics 2022, 10, 2403

Figure 6. Sensor case geometry for 10 mm orifice plate.

The following initial and border conditions were considered:

 Air assumed as incompressible fluid [17];
 Density = 1.18 kg/m3 ;
 Atmospheric pressure at the outlet (101,325 Pa);
 Dynamic Viscosity = 1.7984 × 10−5 kg/(m-s);
 No slip condition at the walls.
A mesh sensitivity analysis was performed for 3 different meshes. The first one
contained 6962 elements; the second, 27,843; and the third, 43,045. The populations corre-
sponded to meshes with element sizes of 0.0005 m, 0.00025 m, and 0.0002 m, respectively.
These 3 different meshes were run as steady state cases for the 4 mm orifice and 100 L/min
airflow. The results were then graphed.
Figure 7 shows that the results began to converge towards a single pressure difference
value. If the mesh refining continued, a single pressure difference value would be reached.
Since the geometry was simple, a highly populated mesh was not necessary. This is shown
in Figure 6, where almost double the number of elements only resulted in an increase of
0.31% of the pressure difference, as well as a considerable increase of almost 6 times the
required number of iterations before the residuals reached convergence. Therefore, to save
computational resources and time, but at the same time still obtain acceptable results, a
mesh with an element size of 0.00025 m was selected for the simulations. The mesh used
for the 10 mm orifice model is shown in Figure 8.

Figure 7. Mesh Sensitivity Analysis.

Figure 8. Mesh for 10 mm orifice model.

218
Mathematics 2022, 10, 2403

2.5. Transduction Stage Design

After studying the theoretical model with the Navier–Stokes equations and analyzing
the sensor case via simulation methodologies, experiments were conducted to develop the
experimental model of the designed sensor/transducer. Equations (6) and (7) represent
a non-linear system, where u(t) is the input signal (for example pressure from cam over
pivot), y(t) is the response (such as for example airflow crossing airbag to a connected
hose), x(t) is an internal variable which allows the connection between input signals and
response, f and h represent the non-linear characteristics of the system such as functions,
while θ represents physical parameters of the system and t the time. The function f contains
the mechanic parameters, fluid parameters and electrical parameters for the correlation
between the theoretic and experimental model of the designed sensor [1,13,18].

dx (t)
= f ( x ( t ), u ( t ), θ ) (6)
dt
y ( t ) = h ( x ( t ), u ( t ), θ ) (7)
It is possible to propose the mathematical model of the system (sensor/transducer)
with Equations (6) and (7) described above. Moreover, it is necessary to find the optimal
signal (input variable) that provides a predicted response. Hence, Equation (8) depends on
the trajectory solution as set point Rs and r(k i ) is the set point of sample time k i .

Rs T = (111 . . . 11111111)r (k i ) (8)

Equation (9) has a dependence of matrices Rs and ΔU. The optimal predicted solution
for a desired value in the system is obtained based on the cost function “J” among optimal
predicted response with this “trajectory solution Y” such as the optimal predicted solution
ΔU, which is described using Equation (9). The procedure to calculate ΔU is given by Equa-
tions (10)–(14), furthermore, R is a diagonal matrix to tune the desired optimal response
matrix.
J = ( Rs − Y ) T ( Rs − Y ) + ΔU T RΔU (9)
On the other hand, the “Optimal predicted response Y” has a dependence of “matrices
F, ϕ and the state vector X(ki)” as shown by Equation (10)

Y = FX (k i ) + ϕΔU (10)

Moreover, using the costing function “J” the optimal estimation was analyzed, as it is
described by Equation (11) after replacing Equation (10) in Equation (9).
 
J = ( Rs − FX (k i )) T ( Rs − FX (k i )) − 2ΔU T ϕ T ( Rs − FX (k i )) + ΔU T ϕ T ϕ + R ΔU (11)

Looking for the minimal error, Equation (11) was derived as the dependence of “ΔU”
in Equation (12).
∂J  
= −2ϕ T ( Rs − FX (k i )) + 2 ϕ T ϕ + R ΔU (12)
∂ΔU
Therefore, the minimal value of “J” was achieved using Equation (13).

∂J
=0 (13)
∂ΔU
This means that the optimal “ΔU” is given by Equation (14), which helps to obtain the
optimal and predicted “Y”.
  −1
ΔU = ϕ T ϕ + R ϕ T ( Rs − FX (k i )) (14)

219
Mathematics 2022, 10, 2403

On the other hand, the matrix ϕ can help to obtain information of the material sensor
regarding the thin film, which is subject to the air pressure. Moreover, the geometry
(nanostructures) of the thin film can be stored over the matrix. Therefore, Equation (15)
provides the correlation between the optimal predictive solution of the transduction with
the matrix adaptive coefficients, which keep information of the material and geometrical
characteristic of the designed sensor.
Furthermore, the periodical responses from the sensor/transducer are provided by
Equation (15) though a transfer function analysis, thereby making it possible to identify the
first order behavior in the steady state. Therefore, Cs represents the answer variable for the
first order, Ns is the input excitation signal, t is the response time and K is the proportional
gain.
Cs K
= (15)
Ns ts + 1
Equation (15) is reduced in the time domain and Equation (16) is obtained, which
keeps the model of Equation (5) in a steady state. In this context, the excitation signal is ΔP
that looks for the response variable fluid flow Φ, and R is the constant derived from the
first order solution for Equation (14) in Equation (10). Hence, it was possible to correlate
stability parameters with the geometrical and material properties of the designed sensor.
 
ΔP
Φ= (16)
R

For this reason, the transducer adapts the signal received from the sensor according to
the value measured of the physical variable to another equivalent (for example, the pressure
difference transduced to its equivalent as an electrical signal, which is from “cmH2O” to
“mV”). In this context, Figure 9 represents the static curve given in the section “A, B, C” for
the coordinate system “X1, Y1”. However, the static curve can help to identify the linear
response regions, by working in linear regions or with non-linear mathematical models.

Figure 9. Scheme to represent the sensor/transducer data processing.

It should be noted that depending on the response time of the system (the system is
considered as a research plant), it can lose information in the linearization process, due to
its non-linear behavior, which is the reason why many authors suggest working in linear
zones despite such risks (losing information in linearization process). In addition, the
process of discretizing the signal information may also be lost, such as the dependence on
the response time of the system and the sampling time (important parameter to discretize a
signal).
Based on the conditions described in the previous paragraph, a transducer is neces-
sary to transform the pressure difference into its electrical equivalent (in mV) in order to
obtain the value for the air flow and also the displaced air volume. Theoretically, using
Navier–Stokes models, the following relationship of the three physical variables is achieved:
“pressure, flow and volume”, where the numerical solutions of the differential equations
are also prone to loss of information as a dependence of the methodologies of numerical
approximations in the solutions. However, the disadvantage of a completely theoretical

220
Mathematics 2022, 10, 2403

correlation is that the modeling of disturbances is frequently not achieved, which is solved
by either approximation modeling or by correlation with experimental data.
Therefore, the transients of the plant’s dynamic system (while it is first-order) for the
sections represented by d0 , d1 ; d2 , d3 ; d4 , d . . . and with similar slopes only in the linear
trends “AB” or “BC” are located in the coordinates “X2, Y2”. In such sections, the response
times of the plant (the designed sensor/transducer) can be obtained.
Moreover, for each section, d1 , d2 ; d3 d4 , and “d . . . , henceforth” the characteristics
of overshoots, settling time, damping and parameters that indicate the stability of the
designed system can be maintained.
The static and dynamic curves are algebraically interpreted as models of differential
equations (polynomial models). Consequently, the correlation is obtained between the theo-
retical modelling and modelling based on the experimental data analysis results in the final
model of the designed “sensor/transducer” system, that implies a response “Experimental
Model (EM) and Theoretical Model (TM)” from linear (linearized) or completely non-linear
(NLM) models.
Indeed, the system formed by the sensor/transducer represented by “ST” receives a
mechanical signal “Ms1” (fluid mechanical signal for this case, which is pressure difference)
and transforms it into the electrical equivalent “Es” or a response signal “Ms2” (fluid
mechanics variable as volume or airflow equivalent signal). This conversion (transduction)
requires characterization and calibration steps, which are depicted by the scheme sum-
marized in Figure 10. From the static measurement data analysis, is possible to recognize
the linear and nonlinear operations in the ST system. This information is quite important
according to the design of the algorithm, which provides the final information of the ST
measurement, such as the breathing variables of pressure, volume, and airflow.

Figure 10. Scheme of a sensor/transducer system.

In the mathematical model of the sensor/transducer as a system, the transducer is

composed by electronic devices that convert the fluid-mechanic variable of “pressure
difference” to its electrical equivalent, as is the case for the variables “airflow and volume”.
Furthermore, the characterization of each physical variable to be measured is necessary,
with electro/mechanical devices that are able to measure the desired physical variables
being monitored, which is showed in Figure 11.

Figure 11. Setup of the designed sensor/transducer system.

221
Mathematics 2022, 10, 2403

3. Experimental Section
3.1. Orifice Plate Diameter Study
The working concept of the orifice plate is to generate a pressure difference which
is related to the flow passing through it [1]. In this sense, different diameters could be
considered; however, an optimum range of diameters presents a balance between the
produced pressure difference sufficiently high to be used on the transduction stage, and
sufficiently low to not to produce significant air flow or pressure drop in the breathing
circuit.
Figure 12 (by views of subfigures a–c) shows the variable orifice-plate flow sensor it is
manufactured by two pipes joined together with a ferrule clamp to have a quick release
mechanism. Downstream of the ferrule clamp, an orifice-plate is welded to the flange. In
this study, 5 parts were fabricated with different diameters: 4, 6, 8, 10, 12 mm. Distances of
the pressure taps to the orifice plate were established according to the design considerations
of an orifice plate air flow sensor [3]. However, the distance of the downstream pressure
tap was not able to be less than half radius because of welding limitations for the geometry.

(a) (b) (c)

Figure 12. Variable Orifice-plate flow sensor.

3.1.1. Steady State and Dynamic Response

The variable orifice plate air flow sensor was tested for the air flow measurement in
a steady state condition. Figure 13a shows the experimental setup designed for this test,
where the variable orifice plate was installed in the air flow test bench, an air compressor
providing the air flow through the variable orifice plate sensor was included, a pressure
difference transducer based on microphones connected to an Arduino UNO microcontroller
board was integrated and an algorithm was designed and programmed for this transduction
stage. Figure 13b shows a schematic representation of the experimental setup and the
interaction of its components.

3.1.2. Measurement in a Mechanical Ventilator

The variable orifice plate of the air flow sensor was installed in a breathing circuit
connected to a mechanical ventilation system based on cams “OxygenIP.PE” [1]. This
experiment corresponded to the real working conditions of the air flow sensor/transducer
which was designed. A medical gas flow meter Fluke VT650 was connected to the circuit
to obtain the real air flow over time, which is showed by the Figure 14a.
The artificial lung Fluke Accu Lung (a precision test lung) was used in order to simulate
the ventilation parameters, which is showed by the Figure 14b as part of the experiments
setup. This device is able to simulate 3 lung stages, from a healthy lung up to a damaged
lung. A healthy lung corresponds to a compliance of 50 mL/cm H2O and a resistance of
5 cm H2O-s/L, while a damaged lung corresponds to a compliance of 10 mL/cm H2O and
a resistance of 50 cm H2O-s/L. In this sense, for each orifice diameter inside the sensor, the
experiment was run four times, twice for each lung parameters and twice for the minimum
and maximum cams, XS and XL, respectively.

222
Mathematics 2022, 10, 2403

Figure 13. Schematic representation of experimental setup for steady state air flow.

Figure 14. Scheme of experimental setup for the sensor transducer connected to a Mechanical
Ventilator.

223
Mathematics 2022, 10, 2403

3.2. Experimental Validation of the Sensor/Transducer Device

In the Figure 15, the designed sensor connected to the mechanical ventilator (Appendix A
contains more details) is depicted; Figure 15a shows the designed sensor is shown to be
connected to the air flow calibrator; and Figure 15b shows the air flow calibrator that needs to
be connected to the personal computer in order to validate the expected measurements, and
all the setup is showed by Figure 15c.

(a)

(b)
(c)

Figure 15. Setup of the sensor/transducer device connected to OxygenIP.PE Mechanical ventilator.

4. Results
4.1. CFD Simulation Results
4.1.1. Steady State Results
The following results were obtained for the steady state simulations for a 100 L/min
airflow in a 10 mm diameter orifice. Figure 16a shows how the velocity increased as the
area reduced (orifice throat), which is explained by the flow’s intention of conserving
the flow rate (mass conservation and incompressibility assumption). On the other hand,
Figure 16b shows how the static pressure reduced, which occurred in response to the
velocity’s increase. This is explained by the flow’s intention of conserving its momentum.
It must also be noted that the behavior of these contours matched the results presented by
Karthik [19].

4.1.2. Profile of Mechanical Ventilator Curves

A certain (expected) proportionality between the airflow and difference pressure
curves is shown in Figure 17a. It is widely known that a higher pressure drop corresponds
to a more abundant flow. This is explained because in turbulent flows the pressure loss is
proportional to the flow velocity squared. This diameter alternative presents a maximum
pressure drop of 274.28 Pa for a corresponding airflow of 67.74 L/min. Next, the steady and
transient state analysis results were compared and plotted. Figure 17b shows an excellent
similarity between the transient and steady state curves for the plotted points. Although a
margin is observed between both curves, a good estimate of the transient behavior of the
sensor can be modeled using steady state approximations. This behavior was expected,

224
Mathematics 2022, 10, 2403

since Funk et al. [20] concluded in their work that steady state approximations can be used
in transient situations with excellent or valid results for most engineering applications.

(a)

(b)

Figure 16. (a) Velocity Contours for a 100 L/min airflow, 10 mm diameter orifice; (b) Static pressure
contours for a 100 L/min airflow, 10 mm diameter orifice.

(a) (b)

Figure 17. (a) Airflow profile and DP profile over time; (b) Steady vs. transient state results for a
10 mm diameter orifice.

225
Mathematics 2022, 10, 2403

4.2. Experimental Results

4.2.1. Steady State
In the following paragraphs, the results of experimental tests to verify the results of
the proposed ST in steady state are provided. The analyses were made by comparisons
with a Fluke instrumentation equipment, because the steady state information helped to
obtain an understanding of the response of the designed ST under disturbances, such as
unexpected overshoots or electromagnetic noise. In that context, the tests were run for the
4 mm diameter sensor/transducer and a steady-state inlet pressure of 5 mBar, 10 mBar and
20 mBar provided by an air compressor. A graphical comparison between the airflow plots
is shown in Figure 18. The “Fluke” curve provides information of the airflow measured by
the “Fluke” analyzer, and the “ST” curve shows the airflow as a response of the designed
sensor/transducer ST.

Figure 18. Flow and Flow ST for diameter 4 mm (Inlet pressure: 5, 10, 20 mBar).

The pressure difference curves (in mV) are plotted together in Figure 19. The “ST” curve
provides information related to the pressure measured by the sensor/transducer in mV.

Figure 19. Pressure measured by ST in mV, for diameter 4 mm (Inlet pressure: 5, 10, 20 mBar).

226
Mathematics 2022, 10, 2403

The pressure difference curves are plotted together next in Figure 20. Each curve
provides information of the pressure measured by the sensor/transducer in Pascal.

Figure 20. Pressure measured by ST (Pa), for diameter 4 mm (Inlet pressure: 5, 10, 20 mBar).

The tests were run for the 8 mm diameter sensor/transducer and a steady-state inlet
pressure of 10 mBar, 20 mBar and 30 mBar was provided by an air compressor, for which
the plotted airflow curves are shown in Figure 21. The “Fluke” curve provides information
on the airflow measured by the “Fluke” analyzer, and the “ST” curve shows the airflow as
the response of the designed sensor/transducer ST.

Figure 21. Flow and Flow ST for diameter 8 mm (Inlet pressure: 5, 10, 30 mBar).

The transduced pressure difference curves in mV are shown in Figure 22. Each curve
provides information on the pressure measured by the sensor/transducer in mV.

227
Mathematics 2022, 10, 2403

Figure 22. Pressure measured by ST (mV), for diameter 8 mm (Inlet pressure: 5, 10, 20 mBar).

On the other hand, the pressure difference curves are shown in Figure 23. Each curve
provides information on the pressure measured by the sensor/transducer in Pascal.

Figure 23. Pressure measured by ST (Pa), for diameter 8 mm (Inlet pressure: 5, 10, 20 mBar).

4.2.2. Open and Close Loop

Figure 24 shows the behavior of the airflow obtained by the designed sensor transducer
(ST curve) in comparison with the airflow reference pattern (Fluke curve measured by the
Fluke equipment). It is possible to understand that the ST system can measure changes in
airflow when the tester prototype is closed or open to evaluate the steady state behavior of
the designed ST. Nevertheless, it was possible to develop better responses of the system
such as the dependence of the diameter of the ST, and the best responses (less error percent
between ST and Fluke measurements) were achieved with orifice diameters of 8, 10 and
12 mm.

228
Mathematics 2022, 10, 2403

Figure 24. Airflow comparison in open loop and closed loop for the ST evaluation.

Figure 25 shows the pressure difference measured by the designed ST for a closed
loop (upper subfigure) and open loop (lower subfigure). The ST system can detect changes
or tendencies to maintain a steady state as a consequence of an open or closed loop of the
tester prototype; therefore, it was possible to measure this pressure difference with the
designed ST.

Figure 25. Pressure difference comparison in open loop and closed loop for the ST evaluation.

4.2.3. Dynamic Response

Figure 26 presents the dynamical response of the system ST to a pressure excitation
signal given by 10, 20, 30 mBar obtained through the airflow regulated by the air compressor
of the tester prototype, with an orifice diameter of 4 mm. It was found that ST cannot

229
Mathematics 2022, 10, 2403

provide a faster response in the presence of transient changes (red curves for airflow,
pressure difference and volume).

Figure 26. Flow, pressure difference and volume by dynamic analysis.

This can be explained by the geometrical characteristics of the ST and its nonlinear
behavior in diameters from 4 mm to 6 mm; however, the ST system can provide a better
response in the presence of disturbances or transient changes when the orifice diameter is
between 8, 10 and 12 mm. The reference corresponds to the measurements of the Fluke
equipment (blue curve for airflow, pressure difference and volume).
For the 8 mm diameter sensor/transducer and the mechanical ventilator excitation
signals, the ventilation curves are shown in Figure 27. In every subfigure, the medical gas
flow meter Fluke VT 650 proportionated the blue color curves in low frequency speed and
red color for high frequency speed, which provide information on the volume, pressure and
airflow (subfigures I–III), and the ST curves provide the volume, pressure, airflow the green
color curves in low frequency speed and yellow color for high frequency speed (subfigures
I–III). During the 2 min of measurement, the ST demonstrated a faster and robust sensor
behavior (for 8 mm of diameter) as a consequence of the dynamical evaluation made by a
low-cost mechanical ventilator based on cams OxygenIP.PE.

Figure 27. The dynamical behavior of the sensor/transducer connected to a mechanical ventilator
low cost based on cams OxygenIP.PE.

230
Mathematics 2022, 10, 2403

Figure 28 shows the airflow curves for the 8 mm diameter sensor/transducer and
the mechanical ventilator excitation signals. In every subfigure, the Fluke curves (blue
color) provide information on the airflow measured by the medical gas flow meter Fluke
VT 650 and the ST curves (red color) represent the airflow measured by the designed
sensor/transducer.

Figure 28. The dynamical behavior of the sensor/transducer connected to an air compressor in
airflow range of work between 75 L/min and 105 L/min.

During the 1 min of measurement, the ST shows a faster and robust sensor behavior
as a consequence of the dynamical evaluation made using an air compressor in the range
of work from 75 L/min to 105 L/min.

5. Discussion
5.1. Data Interpretation Analysis
In order to interpret the experimental data with the designed algorithm, it was neces-
sary to establish a correlation with the theoretical model to enhance the adaptive parameters
of the polynomial model analyzed in Section 2.5.
Therefore, the experimental analysis of the designed polynomial model was correlated
with Equation (18) in the theoretical model, as this equation gives the pressure difference “ΔP”
as dependence on the airflow “Φ”, geometrical parameters “A2 , A1 ” and flow density “ρ”.
    2 
Φ 2 ρ  A2
ΔP = 1− (17)
A2 2 A1

The theoretical model only produces a quadratic relationship between the airflow with
the pressure difference, while “A2 ” tends to decrease, the pressure difference and airflow
tend to increase.
Since the airflow for the breathing analysis had an “increasing and decreasing behav-
ior”, to obtain an appropriate measurement with the designed ST, the static and dynamic
behavior of the sensor was studied using an adaptive analysis of the polynomial model.
Therefore, the static behavior was provided as a dependence of “airflow, pressure
difference, and diameter”. Figure 29 depicts a flow and two points crossing its axis, for
which P1, V1, Y1 and P2, V2, Y2 are the pressure, flow speed, and position at points 1 and
2, respectively.

231
Mathematics 2022, 10, 2403

Figure 29. Schematic representation of airflow transmission through points 1 and 2.

The static curve gives the relation among two variables in order to identify a linear
and nonlinear range of work; therefore, it is possible to understand its dynamic behavior.
The dynamical analysis of the airflow attempts to obtain the linear answer of the
system so that the ST can make estimations of the physical variables volume, and pressure
as a consequence of the pressure difference, which was achieved by the correlation between
the experimental and theoretical analyses described in previous chapters. Therefore, the
curves “A, B, and C” represent the airflow curves that were expected to be achieved in the
characterization of the interpretation data with the ST, which is depicted by the Figure 30.

Figure 30. Schematic representation of the airflow dynamic and static behavior.

Two types of studies were performed, one for the static responses of the ST system,
and another for the dynamic ones. The first was conducted using a tester prototype with a
compressor as an air source, which is depicted in Figure 31 and was used for the static and
dynamic tests. The pressure difference measured by the sensor was calibrated in order to
provide the same response as the air flow meter Fluke. Finally, the other group of static

232
Mathematics 2022, 10, 2403

and dynamic tests was performed with the mechanical ventilator as the air source. Hence,
the pressure difference registered provided information to understand the dynamics of the
breathing variables (pressure, volume and airflow) in order to understand the behavior of
the sensor/transducer as part of the ventilation circuit.

Figure 31. Block diagram testing transducers.

The designed ST depends on the diameter and pressure difference; for this reason, it
was necessary to choose which transducer to use to obtain the electrical equivalent of the
pressure difference. Finally, the evaluated alternatives were as follows: transducer MAS1,
the integrated circuit MDPDX10P, the transducer constructed with two small microphones
sourced from the ARDUINO company, and a transducer based on nanostructures. Different
alternatives are shown in Figure 31.
The measured pressure difference by the designed ST was central to obtaining the
airflow after a transduction process, which was obtained by MDPX and MAS1, making the
electrical equivalent conversion from air pressure difference signal to airflow signal in Volts.
The electrical transduction was performed with two microphones measuring decibels and
correlating with pressure by converting the response to volts. The electrical equivalence
values help to analyze the statistics of the physical variables “air pressure, volume and
airflow” as a consequence of the designed algorithms by polynomial (as they are depicted
by “Algor Ans1”, “Algor Ans2”, and “Algor Ans3” in Figure 31) correlations between the
transduction signals. However, in order to enhance the robustness and response time of
the designed ST, we modified the samples that receive the pressure difference by samples
based on nanostructures of AAO, as is depicted in S1, S2 in Figure 31, which also send the
transduction information through the algorithm depicted by “Algor Ans4”.
The best estimation of the breathing physical variables “airflow, pressure, and volume”
was chosen because of the pressure difference transduction techniques that were studied in
the previous chapters. These microphone-based transducers obtained estimated variables
with a maximum error of 1% in a steady state, while the MAS1-based transducer estimated
variables with a 0.95% error in a steady state. In turn, the MDPX-based transducer presented
a steady-state error of around 2%. However, the cost of the microphone-based transducer
was around 70% cheaper than the most expensive transducer, the MAS1. For this reason,
the microphone-based transducer was selected as the main transducer for the designed ST.
When performing the main analysis of the transducer algorithm, we sought to identify
the adaptive coefficients from the experimental data obtained by the transducer and the
polynomial identification estimate using pressure difference “DP”, the pressure “P”, the

233
Mathematics 2022, 10, 2403

airflow “F”, and the volume “V”. If the adaptive estimation was predicted according to the
reference variable with an error of less than 1% in the steady state, then the ST provided
information on the breathing variables, which is depicted in Figure 32.

Figure 32. Flow diagram of the ST measurement and data storing.

Therefore, it was necessary to define the input variables in order to correlate the
estimated responses under the dependencies of the references variable. Figure 33 shows
the two input variables of the system from the data interpretation analysis, which are given
by the pivot displacement in sexagesimal degrees (red curve color curve in subfigure III of
Figure 33). The pressure difference in its electrical equivalent (mV) was considered as a
second input or excitation variable (blue color curve in subfigure I of Figure 33). Moreover,
the pressure difference measured by the nanosensor based on nanostructures, as shown in
subfigure II, revealed that the nanostructure sensor maintains a rapid response time and
robustness under disturbances of the pivot (green color curve in subfigure II of Figure 33).

Figure 33. Motor angle displacement and microphones answer.

As a consequence of the input excitation signals, it was possible to achieve and

identify the adaptive coefficients which were evaluated in high and low operation because
the frequency range of the requirements is between 15 Hz and 20 Hz.

234
Mathematics 2022, 10, 2403

Therefore, Figure 34 shows the result of the ST as a consequence of the electrical

transduction from the pressure difference to the airflow values in millivolts for cams of
extra small, medium, and extra-large sizes. The obtained data are presented in Figure 34 in
subfigures I–III, in which blue indicates high frequency and red curves low frequency.

Figure 34. Microphones in high and low operation.

On the other hand, it is apparent that for the ST response as a consequence of the
electrical transduction from the pressure difference to the airflow values in millivolts for
cams of extra small, medium, and extra-large sizes, the information achieved is shown in
Figure 34 for the subfigures I–III, in which the blue color indicates high frequency and red
represents low frequency. With this information, it is possible to identify a linear range of
work.
The ventilator works as an intermediatory for the signal received from the rotor
position sensor (RS1(AS5047) and RS2) used to measure the rotor speed of the motor and
control it with a driver (DR1) that has its own control system with its own actuator and
sensor with an electrical current to the motor (M).
This is the reason why a controller (as central control unit) also requires the signal
from a rotor position sensor to measure the angular displacement of the pivot to provide
an estimation of the air volume, and finally, the airflow sensor (FS1), which has its own
transduction algorithm to provide information to medical doctors through the touchscreen
panel, as depicted in Figure 35.

Figure 35. Block diagram of the control system for online operation.

235
Mathematics 2022, 10, 2403

The ST achieved good performance for the diameter of 8 mm due to its optimal
response based on appropriated magnitudes and tracking for the mechanical ventilator
curves. In the “data interpretation analysis” the ST with an orifice diameter of 10 mm
was used. Despite the nonlinear response, the airflow was obtained because of the good
performance of the ST algorithm in transducing the pressure difference signal to the airflow
variable for a larger range of work (low and high speed of the ventilator).
The range of work of the test was limited to a high frequency of 20 Hz and a low
frequency of 15 Hz. The transduction was achieved with the microphone sensors used to
obtain the pressure difference as is shown in Figure 36I, where the blue curve provides
information about the system working at high speeds and the red curve for low-speed
operation. Figure 36I shows that at high speeds it was possible to obtain an increased
amplitude (50 percent). Both signals were achieved in millivolts due to the transducer
(small microphones) having greater sensitivity at higher speeds.

Figure 36. Microphones answer in high and low operation and small and big cam.

Furthermore, it was necessary to define the input variables to correlate the estimated
responses under the dependencies of the reference variable (analyzed in previous chapters).
Figure 36II shows the input variables of the system from the data interpretation analysis,
which are given by the motor displacement in sexagesimal degrees, with the blue curve
indicating high speed and red curve representing low speed. It is necessary to remember
that motor displacement causes pivot displacement and it is enough to use only one sensor
to achieve the angle displacement during the polynomial correlation variable analysis.
Therefore, the main variable obtained from the ST is the airflow rate, given by the
transduction operation supported by the predictive estimations for the other breathing
variables “volume and pressure” obtained by the polynomial analysis (described in chapters
lines above) from the airflow variable.
Figure 36III–V presents the airflow rate, volume, and pressure, respectively, in which
the blue color curves are the airflow, volume, and pressure in high-frequency operation,
and the red color curves are the airflow, volume, and pressure in low-frequency operation,
which were obtained as a consequence of the predictive/adaptive algorithm; however,
it is necessary to keep in mind the importance of the previous data analysis during the
characterization and calibration of the ST to obtain the presented results.
It was necessary to develop a multivariable control algorithm to control the respiratory
variables (air flow, pressure and volume), hence the pressure difference transduced in
volume, air flow and pressure as respiratory variables required a multivariable control
algorithm, which was able to control the rotor speed of the mechanical ventilator motor,
the air pressure difference and the air flow, as the implicit function, pressure and volume.

236
Mathematics 2022, 10, 2403

5.2. Rotor Speed Control of the Mechanical Ventilator Motor

The rotor speed control of the mechanical ventilator motor was achieved by an adap-
tative predictive model, for which a system identification of the motor physical parameters
was necessary. However, unexpected disturbances caused by vibrations during the cam
movement can reduce the performance of the control. Therefore, it was necessary to design
a PID (Proportional, Integral and Derivative) control as part of the identification system to
achieve the physical parameters of the mechanical ventilator motor.
Hence, the control inside the identification of the motor parameters was obtained
by the classic PID controller equation in the Laplace (S) domain, in which KPp is the
proportional constant, KIp is the integral constant, and KDp is the derivative constant.

KI p
PID (S) = KPP + + KD p S (18)
S
The plant (mechanical ventilator motor) was analyzed in two sections, namely the elec-
trical section and the mechanical section which are described in the following paragraphs.

Rotor Control Position of the Mechanical Ventilator Motor as a Function of Electrical

Current
The rotor control position of the mechanical ventilator motor is depicted by the block
diagrams in Figure 37. Inp(S) is the electrical current signal as the input excitation signal,
Yans(s) is the position response that is necessary to obtain the speed of the motor, Cont(S)
is the controller during the identification system, Plant(S) is the transfer function for the
plant that is the electrical motor composed by its electrical subsystem and its mechanical
subsystem, and Sens(S) is the transfer function for the position sensor.

Figure 37. Block diagram for the rotor position control of the mechanical ventilator.

Using the algebra analysis from the block diagram depicted above, the following is
obtained:
Inp(S) − Sens(S)Y (S)Cont(S) MS(S) = Y (S) (19)
From which following equation is achieved:

Yans(S) Cont(S) Plant(S)

= (20)
Inp(S) 1 + Cont(S) Plant(S)Sens(S)

In Equation (20), the transfer function for the rotor position sensor is generalized, but
in spite of that, the response time of the position sensor is very short in comparison to the
speed of the motor. Hence, its transfer function was reduced as Sens(S) = Ks. Therefore, the
characteristic equation is given by Equation (21), in which m is the mass of the rotor motor,
ki is the electrical current coefficient of the motor and ky is the displacement coefficient of
the motor.   
KI p Ki
KPP + + KD p S Ks + 1 = 0 (21)
S m S2 + K y

237
Mathematics 2022, 10, 2403

The reduction of the equation above is given by Equation (22).

 
KD p S2 + KPp S + KI p K I KS + S(mS2 + Ky ) = 0 (22)

Equation (22) is organized in Equation (23) as a polynomial in descending order.

$ % $ %
KD p Ki Ks 2 KPp Ki Ks + Ky KI p Ki Ks
S3 + S + S+ =0 (23)
m m m
The controller’s parameters need to be obtained by different methodologies such as a
stability analysis and a comparison with the theoretical model of dynamic systems given
by Equation (24) [21], in which ω0 is the natural frequency for the system, is the damping
effect, and α is the coefficient as the auxiliary connector between Equations (23) and (24).

S 3 + ω0 ( 2 + α ) S 2 + ω0 2 ( 1 + 2 α ) S + α ω0 3 = 0 (24)

Therefore, the controller parameters KPp and KI p , can be obtained by the compari-
son of the coefficients from Equations (23) and (24), from which Equations (25)–(27) are
proposed as functions of Ki , Ks , KD p , m, ω0 and .
$ %
KD p Ki Ks
= ω0 ( 2 + α ) (25)
m
$ %
KPp Ki Ks + Ky
= ω0 2 ( 1 + 2 α ) (26)
m
KI p Ki Ks
= α ω0 3 (27)
m
Moreover, α is obtained from Equation (27) and represented by Equation (28)

KI p Ki Ks
α= (28)
mω0 3

The controller parameter KI p is obtained by analyzing Equations (25)–(28), and pre-

sented in Equation (29)
2 mω0 3
KI p = KD p ω0 2 − (29)
Ki K s
In a similar context, the controller parameter KPp is obtained with Equations (25)–(29),
and presented in Equation (30).
$ %
2 Ki Ks KD p ω0 + mω0 2 1 − 4 2 − Ky
KPp = (30)
Ki K s

Therefore, Equations (29) and (30) are the integral and proportional controller param-
eters that determine the controller parameter KD p , which can be analyzed as a reference
parameter (analyzed by stability of the dynamic system) to find the PID control for the
motor of the mechanical ventilator. However, the identification parameters of the motor
require a controller with slow reactions to provide sufficient time to identify the parameters
of the system, and the controller selected for the identification of the physical parameters
of the mechanical ventilator motor was a PI controller, hence the derivative parameter KD p
becomes null, and Equation (30) can be obtained from Equation (29), thereby providing the
integral parameter of the PI controller.

2 mω0 3
KI p = − (31)
Ki K s

238
Mathematics 2022, 10, 2403

Finally, Equation (32) is the proportional parameter of the PI controller obtained from
Equation (30) when the derivative parameter is null.
$ %
mω0 2 1 − 4 2 − Ky
KPp = (32)
Ki K s

5.3. Model Predictive Control Analysis

While the above results indicate good control, they lack a fast response which could be
obtained using a predictive model, such as an Optimal Predictive Control model described
in more detail in the following (Appendix B).
The optimal excitation signal in order to find the optimal response is given by [1,13]
−1
ΔS = (φT φ + A) φT ( AS − LX (k i )) (33)

Therefore, using the last equation, the optimal control response for the pressure
difference can be achieved, as well as the rotor speed and implicitly the volume, airflow
and pressure. Moreover, the optimal response control can be enhanced by the specific
weight matrix “W” that must be included in Equation (A5). Equation (33) has similarities
with Equation (A9) but the difference is given by the adaptive matrix coefficient “φ”,
because for this analysis the matrix depends only on the requirement control strategies
while maintaining its dependence on “ϕ” (the geometrical and material coefficients of the
designed ST).
As a consequence, the main control algorithm was designed using the adaptive cascade
model, which is depicted in Figure 38 in which “U1” is the input variable (pressure
difference measured by the designed ST) and the internal response variable is the air flow
“U2”, which is the input variable used to obtain the pressure “Y1” and volume “Y2”,
moreover the internal control variable of the rotor speed has a correlation with the desired
airflow required by the mechanical ventilator.

Figure 38. Control adaptive cascade model for the ventilator control system.

To evaluate the performance of the designed ST in comparison with its enhancement

based on AAO nanostructures, experiments of the mechanical ventilator controlled by the
adaptive cascade algorithm in Huancayo (3250 m over sea level) were conducted. It can
be seen in Figure 39I that the curves as well as the controlled air flow (implicit) measured
by the designed ST (blue colour curve) and the air flow measured by the designed ST

239
Mathematics 2022, 10, 2403

enhanced by nanostructures (red colour curve) were obtained with a steady state error of
1.4% of the ST and 0.8% for the ST enhanced by nanostructures.

Figure 39. Control of respiratory variables made in Huancayo, Peru.

In Figure 39II the curves of the controlled volume (implicit) and measured by the
designed ST (blue colour curve) as well as the volume measured by the designed ST
enhanced by nanostructures (red colour curve) are presented, with a steady state error of
1.5% of the ST and 0.7% of the ST enhanced by nanostructures.
Furthermore, in Figure 39III the curves of the controlled pressure (implicit) and mea-
sured by the designed ST (blue colour curve) as well as the pressure measured by the
designed ST enhanced by nanostructures (red colour curve) are shown, with a steady state
error of 1.8% for the ST and 0.8% for the ST enhanced by nanostructures.
Therefore, the designed ST of nanostructures (TiO2 nanotubes) based on Anodic Alu-
minium Oxide (AAO) provided better results in robustness, response time and resolution,
which supported measurements from the approximated range of operation between 0.985
atm (Lima) and 1 atm (Huancayo).

6. Conclusions
A sensor/transducer (ST) was designed to measure the pressure difference through an
orifice plate and to obtain the airflow rate, volume and pressure over time to register the
performance of mechanical ventilators. The geometrical characteristics of the proposed sensor
based on an orifice plate were researched in order to identify the appropriate parameters to
measure the physical variables. The diameter has an important influence on the performance
of the sensor; in this sense, a diameter of 8 mm provides the optimal response, in terms of
magnitudes and tracking, for the mechanical ventilator curves without significantly increasing
the pressure inlet for the breathing circuit which is provided by the mechanical ventilator. A
diameter in the range of 4 mm to 6 mm can achieve the pressure difference of airflow, but with
a nonlinear response and for the range between 10 mm and 12 mm, the ST can measure the
physical variables in linear responses for breathing values (it can be used to measure airflow
shared by several patients connected to one mechanical ventilator as a future application).
Therefore, it was suggested to use the ST with a diameter of 8 mm due to its linear response and
robustness under disturbances.
The mathematical model of the system was designed and evaluated. Consequently, an
algorithm was also designed to obtain the airflow, volume and pressure from the pressure

240
Mathematics 2022, 10, 2403

difference, which were obtained by the transduction process of the transducer joined to
the designed sensor. Moreover, the ST system was analyzed in dynamical and transient
behavior for ranges of work that depended on geometrical parameters and physical values
of airflow, pressure, and volume of mechanical ventilators for artificial human breathing,
This is a novel proposed sensor design such as other new proposals [22–24] because of the
multiple variable correlation also variables considered like disturbances (temperature and
vibration) and considering the effect of nanostructures in this objective.
For a given orifice diameter, a correlation was observed between the pressure dif-
ference and the airflow. This correlation was not linear, but instead showed a parabolic
tendency. Secondly, similar results and tendencies were observed between the steady-state
and transient-state simulations. Therefore, it can be concluded that a steady-state approach
(and its corresponding assumptions) can be considered to validly describe the transient
behavior of this orifice plate sensor.
For a given flow, the smaller the orifice diameter, the larger the pressure difference
induced in the orifice plate. This could happen because, for a given flow, a smaller orifice
diameter requires the flow to pass through the opening. This increases the flow velocity,
but at the cost of reducing the flow pressure at the outlet, resulting in greater pressure
differences in the orifice plate. Therefore, a bigger inlet pressure would be necessary to
help the flow overcome the pressure loss in the pipes. It can also be concluded that a plate
with a smaller orifice diameter would be the best for this sensor. Taking this into account,
lower pressure drops would occur in the plate, which would not considerably affect the
flow, while at the same time being large enough to be detected by the ST and still being in
the measuring range.
The designed ST enhanced by nanostructures of anodic aluminium oxide can provide
a faster and robust response for geographic conditions where the atmospheric pressure
level is different, as was the case in this research (tests in Lima and Huancayo). The
nanostructure samples used for the designed sensors have the potential to fix adaptive
coefficients to improve the operating work and enhance the response system (breathing
variables) in case of disturbances.

Author Contributions: Conceptualization, J.A.C.C., C.G.R.R. and J.J.J.d.C.y.F.; methodology, J.A.C.C.

and C.G.R.R.; software, J.A.C.C., J.H.L.J., C.G.R.R., B.J.M.A.C., S.C.A. and J.J.J.d.C.y.F.; validation,
J.A.C.C. and A.T.H.; formal analysis, J.A.C.C., C.G.R.R. and S.C.A.; investigation, J.A.C.C., C.G.R.R.
and A.M.G.A.; resources, J.A.C.C., C.G.R.R. and S.C.A.; data curation, J.A.C.C. and J.H.L.J.; writing—
original draft preparation, J.A.C.C. and C.G.R.R.; writing—review and editing, J.A.C.C., C.G.R.R.,
A.M.G.A. and J.J.J.d.C.y.F.; visualization, J.A.C.C. and C.G.R.R.; Transduction mathematical and
algorithm design and nanostructure design J.A.C.C. Supervision, J.A.C.C., J.H.L.J., C.G.R.R. and
J.J.J.d.C.y.F.; project administration, J.J.J.d.C.y.F. and A.M.G.A.; funding acquisition, J.J.J.d.C.y.F. All
authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by CONCYTEC-FONDECYT based on the contest “PROYECTOS
ESPECIALES: Modalidad-Escalamiento de proyectos COVID-19” [74169], to Pontificia Universidad
Católica del Perú, to Protofy and MODASA.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Acknowledgments: The authors would like to express their special gratitude to the Protofy team,
who shared their knowledge and experience of the OxyGEN project development. Furthermore,
the authors would like to acknowledge the OxygenIP.PE team, including Andrea Portal, Juan José
Leyton and Romel Rosales who devised and assisted in the design of the components inside the
mechanical ventilator, Broni Huamaní and Giovani Berrospi who assisted in the transduction stage,
experimental tests and sensors assembly, Joaquin Gonzales who developed the electronic display
for displaying the ventilation parameters for medical staff and Cristian Rimac who assisted in the
schematic representations. Furthermore, we express our gratitude to the Alexandr Hinostroza,
because of his medical analysis suggestions during the design and analysis of the experimental data.

241
Mathematics 2022, 10, 2403

Gratitude is also expressed to the researchers of the Metrology Laboratory PUCP, due to their support
in the experimental analysis, as well as acknowledgement to the Mechanical Department of PUCP
due to the support given to use measurement equipment for the experiments.
Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or
in the decision to publish the results.

Appendix A
This appendix describes the special equipment designed for the validation and mathe-
matical characterization of the air flow sensor in different working conditions of air flow
and the low-cost mechanical ventilator based on cams where the air flow sensor was
installed.

Appendix A.1. Steady State Air Flow Test Bench

In order to determine the optimal orifice diameter for the medical flow sensor in
stationary conditions, a test bench was specially designed for this experiment (Figure A1).
The holders of the sensor can be seen in the diagonal view (Figure A1a) and the side view
(Figure A1b) of the test bench, which are quite important due to keep the correct position
of the sensor during the experiments. It consists of a pipe where an air flow is induced by
an air compressor, a fixed orifice plate to the pipe for flow measurement and a variable
orifice-plate flow sensor. The fixed orifice plate was calibrated with a medical gas flow
meter Fluke VT650 in a validation laboratory implemented in the COVID19-pandemic
context at the Pontificia Universidad Católica del Perú. This fixed orifice-plate can help
to determine the real flow through the variable orifice-plate flow sensor. In this sense, the
orifice diameter can be changed and the relation between real flow and pressure difference
can be obtained.

(a) (b)

Figure A1. Air flow test bench designed for model validation.

Appendix A.2. Mechanical Ventilation System Oxygen IP.PE

OxygenIP.PE is a low-cost functional mechanical ventilation prototype based on
cams [1] OxyGEN-IP, which was developed by Protofy [6]. A redesign was carried out
according to the special requirements and available technology in Peru; however, the
working concept (compression mechanism) established by Protofy was maintained due
to the satisfactory results [25]. Figure A2 shows the OxygenIP.PE mechanical ventilator.
Its working concept consists of an air bag compression mechanism based on cams to
produce a pressure increment inside the cam and then air flow to the patient through a
breathing circuit. Moreover, it can work with five different cams depending on the air
volume requirement for a particular patient.

242
Mathematics 2022, 10, 2403

Figure A2. Mechanical ventilation system OxygenIP.PE, where the novel air low sensor/transducer
is implemented.

Appendix B
In this research, it was necessary to work with strategies of Model Predictive Control
(MPC) to join both subsystems (mechanical and electrical) for the main algorithm, for which
an internal identification system was used, while adaptive control coefficients/weights
searched for the right control. Therefore, the nonlinear function “g” and internal variables
“z(t)” due to excitation “v(t)” as a function of time “t” were important.

dx (t)
= g ( z ( t ) v ( t ), θ) (A1)
dt
On the other hand, “m(t)” correlated with “z(t)” and “v(t)” through a nonlinear function
“n”.
m ( t ) = n ( z ( t ), v ( t ), θ) (A2)
Additionally, “AS ” represents trajectory and the input excitation “O” is included
depending of the sample time ki.

AsT = (111 . . . 11111)O(k i ) (A3)

So, the costing function “J” was analyzed to achieve the optimal desired response [21].

J = ( AS − Y ) T ( AS − Y ) + ΔS T AΔS (A4)

In which the expected response is

Y = LX (k i ) + φΔS (A5)

Otherwise,
 
J = ( AS − LX (k i )) T AS − LX (k i )) − 2ΔS T φT ( AS − LX (k i ) + ΔU T (φT φ + A)ΔS (A6)

where “L” and “φ” are matrices that contain all the physical parameters of the system (as
joining matrices above for the identified result) [21].

∂J 
= −2φT ( AS − LX (k i )) + 2(φT φ + R ΔS (A7)
∂ΔS
So
∂J
=0 (A8)
∂ΔS

243
Mathematics 2022, 10, 2403

From which the optimal excitation signal in order to find the optimal response is given
by [1,21]
−1
ΔS = (φT φ + A) φT ( AS − LX (k i )) (A9)

References
1. Calderón Ch, J.A.; Rincón, C.; Agreda, M.; Jiménez de Cisneros, J.J. Design and Analysis of a Mechanical Ventilation System
Based on Cams. Heliyon 2021, 7, e08195. [CrossRef] [PubMed]
2. Freescale Semiconductor Literature Distribution Center, Inc. Ventilator/Respirator Hardware and Software Design Specification;
Freescale Semiconductor Literature Distribution Center, Inc.: Denver, CO, USA, 2011; pp. 11–20.
3. Brzeski, P.; Lazarek, M.; Perlikowski, P. Experimental study of the novel tuned mass damper with inerter which enables changes
of inertance. J. Sound Vib. 2017, 404, 47–57. [CrossRef]
4. Fowles, G.; Boyes, W. Measurement of Flow, 4th ed.; Elsevier: Amsterdam, The Netherlands, 2010; ISBN 978075068308.
5. Biselli, P.J.C.; Nóbrega, R.S.; Soriano, F.G. Nonlinear flow sensor calibration with an accurate syringe. Sensors 2018, 18, 2163.
[CrossRef] [PubMed]
6. OxyGEN Project. Available online: https://www.oxygen.protofy.xyz/download (accessed on 23 November 2020).
7. Chang, J.; Acosta, A.; Aspiazu, J.B.; Reategui, J.; Rojas, C.; Cook, J.; Nole, R.; Giampietri, L.; Pérez-Buitrago, S.; Casado, F.L.; et al.
Masi: A mechanical ventilator based on a manual resuscitator with telemedicine capabilities for patients with ARDS during the
COVID-19 crisis. HardwareX 2021, 9, e00187. [CrossRef] [PubMed]
8. Jamróz, P. Interaction between the Standard and the Measurement Instrument during the Flow Velocity Sensor Calibration
Process. Processes 2021, 9, 1792. [CrossRef]
9. Włodarczak, S.; Ochowiak, M.; Doligalski, M.; Kwapisz, B.; Krupińska, A.; Mrugalski, M.; Matuszak, M. Flow Rate Control by
Means of Flow Meter and PLC Controller. Sensors 2021, 21, 6153. [CrossRef] [PubMed]
10. Koirala, N.; McLennan, G. Mathematical Models for Blood Flow Quantification in Dialysis Access Using Angiography: A
Comparative Study. Diagnostics 2021, 11, 1771. [CrossRef] [PubMed]
11. Bisgaard, J.; Tajsoleiman, T.; Muldbak, M.; Rydal, T.; Rasmussen, T.; Huusom, J.K.; Gernaey, K.V. Automated Compartment Model
Development Based on Data from Flow-Following Sensor Devices. Processes 2021, 9, 1651. [CrossRef]
12. Hammer, A.; Roland, W.; Zacher, M.; Praher, B.; Hannesschläger, G.; Löw-Baselli, B.; Steinbichler, G. In Situ Detection of Interfacial
Flow Instabilities in Polymer Co-Extrusion Using Optical Coherence Tomography and Ultrasonic Techniques. Polymers 2021, 13,
2880. [CrossRef] [PubMed]
13. Landau, L.D.; Leiftshitz, E.M. Fluid Mechanics, Institute of Physical Problems; U.S.S.R. Academy of Sciences: Moscow, Russia, 1987.
14. Aldoghaither, A.; Liu, D.-Y.; Laleg-Kirati, T.M. Modulating Functions Based Algorithm for the Estimation of the Coefficients and Differen-
tiation Order for a Space-Fractional Advection-Dispersion Equation; Society for Industrial and Applied Mathematics: Philadelphia, PA,
USA, 2015.
15. Bridgeman, D.; Tsow, F.; Xian, X.; Forzani, E. A new differential pressure flow meter for measurement of human breath flow:
Simulation and experimental investigation. AIChE J. 2016, 62, 956–964. [CrossRef] [PubMed]
16. Launder, B.E. MAN—ANSYS Fluent User’ s Guide Releasde 15.0. Knowl. Creat. Diffus. Util. 2013, 15317, 724–746.
17. Paz, C.; Suárez, E.; Concheiro, M.; Porteiro, J. CFD transient simulation of a breathing cycle in an oral-nasal extrathoracic model. J.
Appl. Fluid Mech. 2017, 10, 777–784. [CrossRef]
18. CLÍNIC BARCELONA. Available online: https://www.clinicbarcelona.org/noticias/el-dispositivo-de-ventilacion-de-
emergencia-desarrollado-por-clinic-germans-trias-i-pujol-y-ub-con-protofy-xyz-recibe-la-aprobacion-de-la-aemps-para-
hacer-un-estudio-clinico (accessed on 15 May 2020).
19. Karthik, G.S.Y.; Kumar, K.J.; Seshadri, V. Prediction of Performance Characteristics of Orifice Plate Assembly for Non-Standard
Conditions Using CFD. Int. J. Eng. Tech. Res. 2015, 3, 2321–2869.
20. Funk, J.E.; Wood, D.J.; Chao, S.P. The transient response of orifices and very short lines. J. Fluids Eng. Trans. ASME 1972, 94,
483–489. [CrossRef]
21. Calderón, J.A.; Barriga, E.B.; Mas, R.; Chirinos, L.; Barrantes, E.; Alencastre, J.; Tafur, J.C.; Melgarejo, O.; Lozano, J.H.; Heinrich, B.;
et al. Magnetic Bearing Proposal Design for a General Unbalanced Rotor System enhanced because of using sensors/actuators
based in nanostructures. E3S Web Conf. 2019, 95, 01002. [CrossRef]
22. Хaлилoв, И.A.; Керимoв, C.Х.; Бaгирoвa, C.A.; Гaджиевa, φ., III. Cинтез кулaчкoвoгo мехaнизмa с учетoм услoвий
передaчи сил и кoнтaктнoй прoчнoсти. (Khalilov, I.A.; Kerimov, S.K.; Bagirova, S.A.; Gadzhieva, F.S. Synthesis of a Cam
Mechanism Taking into Account the Conditions for the Transfer of Forces and Contact Strength). 2017. Available online:
http://web.iyte.edu.tr/~{}gokhankiper/ISMMS/Khalilov.pdf (accessed on 6 April 2022).
23. Liang, L.; Qin, K.; El-Baz, A.S.; Roussel, T.J.; Sethu, P.; Giridharan, G.A.; Wang, Y. A Flow Sensor-Based Suction-Index Control
Strategy for Rotary Left Ventricular Assist Devices. Sensors 2021, 21, 6890. [CrossRef] [PubMed]
24. Algarni, M. Optimization of Nano-Additive Characteristics to Improve the Efficiency of a Shell and Tube Thermal Energy Storage
System Using a Hybrid Procedure: DOE, ANN, MCDM, MOO, and CFD Modeling. Mathematics 2021, 9, 3235. [CrossRef]
25. Melaibari, A. Free Vibration of FG-CNTRCs Nano-Plates/Shells with Temperature-Dependent Properties. Mathematics 2022, 10,
583. [CrossRef]

244
 mathematics

Article
Mechanical Model and FEM Simulations for Efforts on Biceps
and Triceps Muscles under Vertical Load: Mathematical
Formulation of Results
Emilio Lechosa Urquijo 1 , Fernando Blaya Haro 2 , Juan David Cano-Moreno 2, *, Roberto D’Amato 2
and Juan Antonio Juanes Méndez 1

1 Facultad de Medicina, Departamento Anatomía Humana, Universidad de Salamanca,

37007 Salamanca, Spain; elechosa@gmail.com (E.L.U.); jajm@usal.es (J.A.J.M.)
2 Escuela Técnica Superior de Ingeniería y Diseño Industrial (ETSIDI), Departamento de Ingeniería Mecánica,
Química y Diseño Industrial, Universidad Politécnica de Madrid (UPM), Ronda de Valencia 3,
28012 Madrid, Spain; fernando.blaya@upm.es (F.B.H.); r.damato@upm.es (R.D.)
* Correspondence: juandavid.cano@upm.es

Abstract: Although isometric contractions in human muscles have been analyzed several times,
there are no FEA models that allow us to use the same modeled joint (the elbow under our case) in
different conditions. Most elbow joints use 3D elements for meshing. Representing the muscles in the
joint is quite useful when the study is focused on the muscle itself, knowing stress distribution on
muscle, and checking damage in muscle in a detailed manner (tendon–muscle insertion, for example).
However, this technique is not useful for studying muscle behavior at different positions of the joint.
Citation: Lechosa Urquijo, E.;
This study, based on the mechanical model of the elbow joint, proposes a methodology for modelling
Blaya Haro, F.; Cano-Moreno, J.D.; muscles that will be studied in different positions by meshing them with 1D elements. Furthermore,
D’Amato, R.; Juanes Méndez, J.A. the methodology allows us to calculate biceps and triceps efforts under load for different angles of
Mechanical Model and FEM elbow joint aperture. The simulation results have been mathematically modelled to obtain general
Simulations for Efforts on Biceps and formulations for these efforts, depending on the load and the aperture angle.
Triceps Muscles under Vertical Load:
Mathematical Formulation of Results. Keywords: FEM analysis; biomechanics; effort mathematical models; NMR reconstruction; reverse
Mathematics 2022, 10, 2441. https:// engineering; elbow joint
doi.org/10.3390/math10142441

Academic Editors: Higinio Rubio MSC: 37M05

Alonso, Alejandro Bustos Caballero,
Jesus Meneses Alonso and
Enrique Soriano-Heras
1. Introduction
Received: 21 June 2022
Over the years, the finite element method (FEM or FEA model) has become the
Accepted: 11 July 2022
preferred tool [1,2] by researchers for the study of analyzes of practically any type. The
Published: 13 July 2022
main reason for this preference is the accuracy [3,4] in the solution and the flexibility to
Publisher’s Note: MDPI stays neutral simulate almost any physical event that may be explained by mathematical equations.
with regard to jurisdictional claims in One of the fields in which the finite element method has been used on numerous
published maps and institutional affil- occasions has been biomechanics [5–11], with applications to specific studies such as joint
iations. analysis, prosthesis design and muscle behavior analysis.
Li et al. [5] used the FEM method to run a simulation of fluid structure interactions
coupling the lattice Boltzmann method and the FEM, and this new study can be used on
blood flow and the heart, or urine interaction with bladder. Della Rosa et al. used the FEM
Copyright: © 2022 by the authors.
method to validate the new fixator with application in fractures [6]. Zhang et al. developed
Licensee MDPI, Basel, Switzerland.
This article is an open access article
a FEM analysis of a lumbar spine implanted with a lordotic cage [7] at L3-L4, and the
distributed under the terms and
results were validated by in vitro testing. The stress distribution on the cage and the range
conditions of the Creative Commons of motion (ROM) of L3-L4 were used to assess the stability of the implant. Denozière and
Attribution (CC BY) license (https:// Ku, in their study, proposed a three-dimensional model of a two-level ligamentous lumbar
creativecommons.org/licenses/by/ segment [8], simulated by static analyses with the FEM software to predict that mobility
4.0/). after arthrodesis at the upper level was reduced in all rotational degrees of freedom by an

Mathematics 2022, 10, 2441. https://doi.org/10.3390/math10142441 245 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 2441

average of approximately 44%, relative to healthy normal discs. Samani et al., proposed a
breast biomechanical model using a FEM formulation [9], specially focused on the modeling
of breast tissue deformation which takes place in breast imaging procedures.
The study proposed by Martinez [12] is focused on the characterization of the muscle
and its response under different conditions. Muscle length changes and the internal stress
distribution for different electrical stimulations were studied. Weiss et al. [13] described
strategies for addressing technical aspects of the 3D computational modelling of ligaments
with FEM analysis. Islan et al. [14], studied the behavior of the glenohumeral joint under
different postures that represent the routine of a violinist from an ergonomic point of view
(RULA method) and using a 3D FEM model to study this movement for a high number
of cycles. Sachenkov et al. [15] studied the movement of the femur in the hip, using 1D
models for the muscles. Martins et al. [16] presented a study with FEM techniques applied
to the pelvic floor, using the Hill model for muscles as a 1D element, extrapolating it to 2D
and 3D models, working with isometric and isotonic contractions. Tang et al. [17] studied
muscle fatigue with a 3D muscle model in different activation situations. Syomin et al. [18]
performed a numerical simulation of contraction of the left ventricle approximated by the
axisymmetric body.
Perreault and Heckman evaluated the ability of the Hill model for muscles to describe
muscle force responses for naturally and electrically stimulated muscles [19]. Alonso et al. [20]
proposed a static physiological optimization model instead of the dynamic optimization model.
For this purpose, they used a modeling of muscles by means of 1D elements along with the line
of action of the same. Holzbaur et al. proposed a different biomechanical model representing
the upper extremity, including shoulder, forearm and arm [21]. The muscles are replaced by a
1D element with a force generation parameter based on experimental data. Park et al. [22]
proposed a torque estimation method at the joint of an index finger in the human hand while
pinching, the study is supported by electromyography and Hill muscle model. Soechting
and Flander used a simplified muscle model [23] to predict torques during arm movement,
comparing their prediction with the EMG results measured in 20 subjects. The muscle model
is also based on the Hill muscle model. Zajac developed a mathematical model to analyze
the muscle tendon, ‘musculotendon’ actuator [24] based only on the ration between tendon
length and muscle fiber length, both at rest.
In the study and FEM analysis of human joints, bones are considered rigid elements,
their geometry does not vary when the joint moves, the same happens with cartilages,
and they can be approximated to rigid, nondeformable parts [25]. On the other hand,
when meshing a muscle in 3D elements into a joint, this mesh is only valid for a certain
position, and if there is need to study a different position in the joint, the muscles should be
remeshed once again. This problem disappears by using a 1D element, rod-type, to mesh
the muscles in the joint. Ligaments can be modeled as springs, 1D elements with two nodes.
According to this, a model meshed with 3D elements for rigid bodies and 1D element for
muscles elements, can adopt different positions just by moving the elements of the rigid
bodies to the new positions.
In the case of 1D elements with just two nodes, if those nodes move in the space,
the element will move as well. As ligaments are defined from two different nodes of two
different bones, if the bones move to another position, the ligament will adopt the new
position. This allows us to reuse the same meshed model in several configurations. For this
reason, the final objective of this study is discovering the equation that would allow one to
create a new finite element type that could be used in biomechanics analysis where muscles
forces are involved in “isometric contractions”. This study focuses on the analysis of human
muscle behavior in order to define a new equation that could predict the muscle force that
would appear in a certain muscle under known conditions. The human joint under study
will be the elbow. The proposed FEM analysis is focused only on the biceps and triceps
muscles and the muscle elbow joint considered in this study are (Figure 1):
• Long biceps
• Short biceps

246
Mathematics 2022, 10, 2441

• Brachii
• Long supinator
• Long triceps
• Medium triceps
• Lateral triceps

Figure 1. Flexors (left) and extensors (right) elbow muscles identification.

Since muscles can only work under traction (they do not transmit compression loads),
different models must be proposed for a separate study of the triceps and biceps muscles.
Several analyses, on different aperture angles of the elbow, will be solved looking for the
effort needed on referred muscles, biceps or triceps, to balance 150 N applied on the wrist
in an “isometric contraction”. The scope of the study includes the following aspects:
• Analyzing the elbow under load at different angles;
• Obtaining efforts in different muscles involved in the joint for different positions;
• Defining mathematical models to predict efforts in muscles depending on:
• Joint angle or muscle length;
• Applied load on the wrist.

2. Materials and Methods

The proposed methodology will allow one to calculate elbow muscle effort. Further-
more, a post-processing of the results is proposed in order to obtain the formulation of
efforts in the studied muscles when the load applied changes. These efforts are referred
to isometric load cases. An isometric load involves muscle contraction against resistance
in which the length of the muscle remains constant. The research can be divided into two
main parts:
1. Finite element models of the mechanism of elbow joint. The 3D anatomical model
developed is based on two diagnostic imaging techniques. CT for the bones and
MRI for the muscles. From this three-dimensional model, different FEM models were
generated, for biceps and triceps muscles studies.
2. Mathematical models predicting efforts in muscles studied. From FEM models, isomet-
ric muscle efforts will be obtained for a vertical load applied on wrist, 150 N including
forearm mass, at different angles of aperture in the joint, from 44 to 164 degrees. These
efforts will be extrapolated for any load applied on the wrist.
The first part is based on previous studies [25]. This model is shown in Figure 2, and it
will be described in the more detail in next section.

247
Mathematics 2022, 10, 2441

Figure 2. FEA model, muscles identification, flexors (left) and extensors (right).

2.1. Finite Element Model

For the creation of the FEM model, a 3D CAD model of the elbow joint is necessary,
from which the mesh of the FEM model will be created.
Said CAD model is obtained starting from the images obtained by means of CT and
MRI, of the elements of the joint (bones and muscles). Thanks to the use of software for
generating 3D CAD models from point clouds, Geomagic Desing X, we can generate the
CAD model of solid bodies.
The FEA model was implemented in Siemens FEMAP 10.2 as preprocessor and post-
processor, and NX Nastran 7.1, as solver. FEA software was described in detail in previous
research [25]. Figure 3 shows the two developed FEA models for the biceps and triceps
FEM analysis.

Figure 3. FEA model, biceps study (left) and triceps study (right), 74◦ aperture angle.

2.1.1. FEA Model Creation Meshing Process, Element Types

Different types of elements have been used for different components:
• Bones. They are meshed using 3D elements (tetrahedral type, 2nd order, 10 nodes).
Teo et al. [26] use 3D elements for meshing neck and head parts of the joints in their
study. In the same way, Donahue et al. [27] use 3D elements with axisymmetric
condition, to model knee parts involved in the study.
• Cartilages. They are meshed using 3D elements (tetrahedral type, 2nd order, 10 nodes).
Donahue et al. [27] and also Abidin et al. [28] model cartilages using 3D elements.
• Bone cartilage junction. Since the cartilage and bone form a solidary union between
both parts, during the meshing process, a coincidence of the nodes of the bone and
the cartilage in the contact area has been carried out. By ensuring that the bone and
cartilage nodes coincide in that area, it is possible to ensure that the transmission of
efforts between both parts is correct.

248
Mathematics 2022, 10, 2441

• Contact between cartilages. It has been modeled using rigid elements that join the
nodes of both cartilage meshed parts involved in the joint, as explained in previous
studies [25], using an RBE2 element. In the RBE2 elements used, the degrees of rotation
freedom have been released so that only forces are transmitted in the X, Y or Z axes.
• Muscles. These are meshed using 1D elements (the type is discussed in this study).
Previous studies have considered muscles in this manner, as 1D element acting along
the imaginary axis of the muscle. Alonso et al. [20] model muscles in the leg to
study human march in the same manner, 1D elements aligned with the axis of the
muscles; Sachenkov et al. [29] use the same approximation for muscles, Parekh [30]
approximates muscles to a 1D FEM element aligned with the muscle axis. This is a
simplification, but it is used by several authors because fusiform muscles only have
two insertion points and low pennation angles. Thus, fibers are aligned so with this
imaginary axis which is formed by joining both insertion points.
• Tendons. These are meshed in combination with muscles, as a unique element muscu-
lotendon (MTU). The same 1D element is considered for tendons.
• Tendon insertion in bone. It has been modeled by adding a rigid element (RBE2),
joining the end node of the 1D MTU element with several nodes in the bone, and thus
distributing the reaction force in the area of the insertion of the tendons, as explained
in previous research [25]. Of the restrictions that an element with these characteristics
can apply, only the displacements in the X, Y or Z axes have been applied, freeing the
turns around them. In this way, and since the MTU only transmits axial forces, we
eliminate any possible bending moments that may affect the insertion of the tendon in
the bone.
The RBE2 elements constitute constraint equations between the nodes to which the
element is connected, they are characterized by having a node, called independent, and
whose degrees of freedom are independent. The rest of the nodes connected to the RBE2
element are called dependent, and in them it must be fulfilled that the displacements of the
independent node must be the same as in the dependent nodes.
That is why they are commonly known as rigid elements, since the distance between
the component nodes remains unchanged.
In Figure 4, an example of an RBE2 element can be found, with an independent node
(i) and four dependent nodes (j1 . . . j4).

Figure 4. RBE2 element example: 1 Independent node and 4 dependent nodes.

The force balancing equation on an RBE2 is such that the reactions that appear on
dependent nodes of the element must balance the force applied in the independent node:

∑ { Reactions f orces} = Findependent (1)

dependent

In this way, using an element of this type to simulate tendon insertion in the bone, the
stress that appears in the 1D element that simulates MTU in the insertion zone is distributed,
eliminating stress concentrations.

249
Mathematics 2022, 10, 2441

Figure 5 shows the application of this insert model.

Figure 5. Biceps tendon insertion using RBE2 element.

2.1.2. FEA Model Creation Muscles Modeling

Given that the proposed study is focused on knowing the forces developed by each of
the muscles of the elbow joint, in different load situations, the modeling of the muscle must
be such that it allows one to obtain these values.
As mentioned above, the objective is to be able to use a 1D element that allows us to
reuse the mesh created to solve the FEM simulation without the need to re-mesh the entire
model for each position of the joint.
The behavior of the element used in our study to simulate the muscle in the joint must
be such that:
• It works only in traction.
• The force of the element used must be generated along the axis that joins the two
nodes that define it; it must not have transverse forces.
• It must not present deformations or displacements that alter the isometric character of
the contraction studied. Therefore, the deformations have to be low.
• The 1D-type elements existing in FEM analysis tools are described below, also indicat-
ing the reasons that lead us to discard them as valid for our purpose.
• Constant spring stiffness. In this case, the results are not valid, since a spring delivers
the force along a curve according to the equation F = K·x, where K is its stiffness
(N/mm) and x is its elongation (mm). The use of this element would mean that a
muscle would deliver a single force value for a given length, which is not true. A
muscle at a given length is capable of delivering a force between 0 and the maximum
force that muscle fibers are capable of developing.
• Variable spring stiffness. For the same reason as before, although in this case the
maximum force could adjust according to the length of the muscle, it does not serve
the purpose of the proposed study. The muscle can deliver a force between 0 and
the maximum force for a given length. In the case of using a spring element, there is
also the drawback of not always being able to balance the joint and the present study
would no longer be isometric.
• Beam element. With this element, it is possible to obtain the necessary force in the
joint to balance it. Beam elements are characterized in that they present axial, shear,
and bending forces, and the muscles only transmit axial tensile forces.
• Rod element. The use of this element meets the study requirements; it only transmits
axial forces, and we can decide on a rigidity in it such that the joint does not present
large displacements and therefore it is no longer considered an isometric contraction.

250
Mathematics 2022, 10, 2441

In search of the element that best reflects the behavior of the muscle within the joint,
the ROD element, given that it only presents axial forces, is the one that best suits the
needs of the proposed mathematical study. However, the said element can work under
tension and/or compression, which would alter the results. That is why two different
FEM models are generated for the study of the biceps or triceps muscles. One in which
the triceps muscles are re-moved, and another in which the biceps muscles are removed.
In this way, and considering the direction of the applied load, it can be ensured that the
analyzed muscles work only under traction.
Several tests have been carried out to define the characteristics of the ROD element that
will be used, maintaining the objective of minimizing its stiffness and keeping the element
deformations at low values so as not to lose the isometric load condition in the analyzes.
For these reasons, the elements used in the FEA models are defined as:
Rod element (1D element);
Constant cross section: 100 mm2 ;
Material parameters:
Young modulus: 2000 MPa;
Poisson coefficient: 0.43.
This type of element transmits only axial forces. Its constitutive equation is shown below.
    
F K 0 ε
= x · x (2)
M 0 KT αx

where
• F is the axial force in the element;
• M is the moment along the axis of the element;
• Kx is the axial stiffness of the element;
• KT is the rotational stiffness along the axis of the element;
• εx is the length variation of the element;
• αx is the angle rotated along the element axis;
To ensure that no torsional stresses are produced in the element, it is sufficient to
assign a value of KT = 0.

2.1.3. Material Properties

The meshing process of the 3D bodies that comprise the model, bones and cartilage,
has been carried out automatically using the meshing tool that includes Femap. The
geometric characteristics of the elements used are common to both bones and cartilage;
they are:
• Second-order tetrahedral elements, 10 nodes;
• Nominal element size 1.5 mm.
The materials considered for each body have the following characteristics, taken from
the CES Edu Pack [31].
The use of an isotropic and linear material for cartilage has previously been used in
those cases in which the duration of the applied load is less than the time constant of the
viscoelastic properties of cartilage, around 1500 s. In this case, given that it is considered an
isometric contraction, in which there is no displacement in the joint and the application time
is considered well below the 1500 s indicated, it is possible to consider it as isotropic, linear,
and elastic with the properties indicated in Table 1. Donahue et al. [27] considered isotropic,
linear and elastic tissue for cartilages in the knee joint in their study of tibio-femoral contact
in the human knee joint, based on the time required to apply the load compared to the
time constant for viscoelastic properties of cartilages, for the same reason. Abidin et al. [28]
consider cartilage as an isotropic, linear and elastic material in their study on a human knee
joint model and verification.

251
Mathematics 2022, 10, 2441

Table 1. Tissue properties for bones and cartilages.

Young’s Modulus Density

Poisson
(Mpa) (kg/m3 )
Bone 17,200 0.41 1790
Cartilage 8 0.49 1150

The use of an isotropic material for bone modelling is known and has previously
been used, Teo et al. [26], meshed head and spine using 3D elements with bone tissue as
isotropic, linear, and elastic.

2.1.4. Mesh Resume

The summary of the FEM mesh is as follows:
• Number of elements: 548,139;
• Number of nodes: 821,755;
• Number of second-order tetrahedral elements: 546,035;
• Number of 1D ROD elements: 4 (for biceps analysis) 3 (for triceps analysis);
• Number of rigid elements RBE2: 2100.
A study of the quality of the mesh, based on the Jacobian value of the elements, shows
us how good its quality is for our objective. Figure 6 shows the elements that are below the
limit value of 0.6 (good quality for the considered analysis).

Figure 6. Mesh quality. Jacobian criteria for evaluating mesh quality. Value represented is
(1-measured value on element/ideal value on element). Maximum threshold defined as 0.6.

2.1.5. Boundary Conditions

For the boundary conditions, the scapula is considered fixed in the area (pink color on
the below image) where the muscles of the scapula would be inserted. These muscles (see
Figure 7) are:
1. Levator scapula muscle;
2. Rhomboid minor muscle;
3. Rhomboid major muscle.

252
Mathematics 2022, 10, 2441

Figure 7. Boundary condition, muscles fixing scapula (left), constrain in FEA model, (right),
scapula fixed.

2.1.6. Load Application

In order to be able to validate results of the analysis, the load considered must be easy
to replicate in a controlled environment; that is why we have defined the load as the mass
of a dumbbell of 15 kg (approximately, 150 N) for the biceps analysis. In the case of triceps
loading, the simulation is more similar to the exercise done using a guided load machine.
The application point of this load is the wrist (the hand is not modeled) and although
this can affect slightly to the results, due to the difference in distance of the load applied in
the joint, this can be solved varying the load as needed when applied in the hand.
All models have the same load applied, at the same node, and in the same vertical
direction (see Figure 8).

Figure 8. Applied load (150 N) in wrist, vertical direction, biceps analysis (left) and triceps
analysis (right).

The load is applied in vertical direction, positive axis (ascending) when analyzing the
triceps and vertical direction, negative axis (descending) when analyzing the biceps. All
load cases are considered as static load cases, isometric contraction of the muscle, and, to
avoid compression loads on muscles, two different models have been developed; one of
them includes only biceps muscles, with descending load, and the other one includes only
triceps muscles, with ascending load.
In the model, gravity force has also been considered. Thus, the mass of the arm is
considered in the analysis.
To do so, the density of elements has been modified so that the values of the upper
arm and forearm mass are [32]:
• Forearm mass: 2.5 kg;
• Upper arm mass: 4.3 kg.

253
Mathematics 2022, 10, 2441

All load cases are considered as static load cases; due to the isometric contraction of
the muscle, there must be no displacement in the joint.

2.1.7. Design of Experiment

The FEM model is considered linear and static. No non-linearity has been considered,
neither in the materials nor in the restrictions of the model. That is why the general equation
of the calculation is so important and can be written as follows:

[ R] = [K ]·[U ] (3)

where
[R] = global loading matrix;
[K] = global stiffness matrix;
[U] = global displacement matrix.
Equation (3) can be solved directly using the Gauss elimination method, there is no
iterative process, and the solution is unique.
The resolution of the problem is carried out through several linear static analyses, each
one of them representing a different load condition in the joint, and the result of them is
the necessary effort in each muscle of the joint to balance the applied force on the wrist.
As mentioned above, there are two different models, thus guaranteeing that the
muscles work only in traction, to study the biceps and the triceps muscles.
In each of the models, the configuration of the joint will be modified by varying its
angle between 44 and 164◦ , at intervals of 10◦ .
The applied load will always be 150 N.
As a reference for the reading of the results, the fully extended arm corresponds to
164 degrees and the fully flexed arm corresponds to 44 degrees (see Figure 9).

Figure 9. FEA model, minimum (left) and maximum (right) studied angles.

2.1.8. FEA model Solution

The FEA model has been solved on a workstation with the following characteristics.
• Windows 10, 64-bit, operating system;
• Intel(R) Core (TM) i9-10900X CPU @ 3.70 GHz Processor (20 cores);
• 128 GB RAM memory;
• 1 TB SSD disk drive.
The solution time for any configuration is approximately 20 min.
As contacts in cartilages have been replaced by rigid elements, RBE2 type with free
rotation on both nodes, the solution time has been reduced from 67 h to 20 min.

254
Mathematics 2022, 10, 2441

As described in a previous study [23] the results, comparing the model with contacts
and the model with rigid elements instead, are similar in both cases, and therefore the use
of rigid elements improves the efficiency of the analysis.

2.2. Mathematical Approach

In this work, several mathematical models are derived from the results of the FEM
simulation. First, there are grade 6 polynomial curves, which represents the efforts of the
five muscles involved for the simulated cases, described in previous section, for a vertical
force of 150 N applied on wrist. Muscle effort is defined as the total force that appeared
in the 1D element used for each muscle in FEM models. These models make it possible
to obtain this effort for each muscle by following an equation for each muscle, with the
following general formulation:

i =6
Musclee f f ort = ∑ ai · αi (4)
i =0

Once the forces needed in muscles have been calculated to balance the joint, with
different angles and with a load of 150 N applied at the wrist, a mathematical approach
will be proposed for obtaining these efforts for other loads, due to the mechanism type
considered in the elbow joint, the applied forces and torque analysis done in the joint. The
value of the force exponent is 1 in all cases, so there is a linear relation between the force
applied and the muscle reaction force [33]. According to this, the greater the force applied
to the wrist is, the greater the muscle force must be, and both values are proportional.
Figure 10 and the following equations justify this simplification in the analysis, where:
• Ws means the weight of the forearm;
• Ws  means the equivalent force for Ws applied at the wrist. The moment created by
Ws or Ws  is the same referred to O point;
• Wq means the load applied in the wrist;
• α means the angle between biceps axis and the normal to the radius longitudinal axis;
• β means the angle between the vertical and the longitudinal axis of the radius;
• a, b and c are geometrical dimensions.

Figure 10. Forces scheme in elbow joint.

255
Mathematics 2022, 10, 2441

The left image considers the proper mass of the arm located in its center of gravity.
Balancing equation on that forces scheme, referred to O point scheme, is:
$ %
Ws · a + Wq ·b · sin β
Fm = (5)
c· cos α
Instead of using Ws , we can use Ws  , applied on the wrist (right image). This Ws 
should comply:
Ws · a·sin β = Ws  ·b·sin β (6)
So, the balancing equation would be the following:

$ % b·sin β
Fm = Wq + Ws  (7)
c·cos α
As seen in Equation (7), the relation between Fm and Wq + Ws  is linear, and constant
for a certain angle only depending on b and c and α and β angles.
According to this, the forces in the muscles when the load is different from the 150 N
applied in this study (including the Ws  in this 150 N) can be obtained by the relation
between the external load applied and the reaction in the muscle.

Wq + Ws  Fext c·cosα 150 N

= = = constant(α, β) = (8)
Fm Rmusc b·sinβ Fm (150 N)

In summarizing, forces in muscles depend on:

• Muscle length (or elbow angle);
• Force applied to wrist.
By applying these mathematical approaches, several polynomial equations have been
obtained, one per muscle, which establishes the relation between force in muscle as:

Fm = Fm (W, α) (9)

where:
• Fm is the force in the muscle (N);
• α is the angle of the elbow;
• W = Wq + Ws  , is the load applied on the wrist (N).
In all mathematical models of muscle efforts, the effort could be obtained for any
angle, α, and for any vertical load, f v . If the forearm mass, m f a , is considered and g is the
gravity acceleration (9.81 m/s2 ) Ws  can be expressed as follows:
a a 
Ws  = Ws · = m f a · · g (10)
b b
Thus, for obtaining the effort:
 a  
E f f ort( f v ) = Fm f v + m f a · · g , α (11)
b
General equations will be obtained describing the effort in each muscle depending on
the angle of the joint and the load applied. These equations and their surface representations
will be the results of this research. Both of them, the equations and surfaces, have been
obtained using the curve fitting tool from MATLAB. It has been noted that these effort
models for the five involve muscles are only used to study isometric muscular contractions,
when muscle length does not vary.

256
Mathematics 2022, 10, 2441

3. Results
The model analyzed includes all muscles in the elbow joint, but the post-processing
results are only applied to the biceps and triceps muscles.

3.1. FEM Simulations, 3D Results

Figure 11 shows the results of running the design of the experiment described here-
inbefore for the 13 cases of elbow angle for the biceps model. The figure represents the
Von Mises stress in the involved bones due to the action of the 150 N as a vertical force
(descendent) on the wrist. Each of the images represents the stress map resulting from the
study of an isometric contraction in the elbow joint at the angle indicated in the image for a
load of 150 N applied to the wrist, vertically and downward.

44° 54° 64° 74° 84° Units. MPa

94° 104° 114° 124° 134°

144° 154° 164°

Figure 11. Different elbow angle configuration: bone Von Mises stress distribution under biceps
loading (150 N on wrist) cases. Figures are included in more quality in Appendix A.

Similarly, Figure 12 shows the Von Mises stresses on the bones due to the FEM
simulation of 150 N of vertical load (ascendant) on the wrist. Each of the images represents
the stress map resulting from the study of an isometric contraction in the elbow joint at the
angle indicated in the image for a load of 150 N applied to the wrist, vertically and upward.

257
Mathematics 2022, 10, 2441

44° 54° 64° 74° 84°

Units. MPa
94° 104° 114° 124° 134°

144° 154° 164°

Figure 12. Different elbow angle configuration: bone Von Mises stress distribution under triceps
loading (150 N on wrist) cases. Figures are included in more quality in Appendix A.

3.2. Numerical Axial Efforts on Main Different Muscles

This subsection collects axial efforts on biceps and triceps involved muscles for the
26 FEM simulations (13 for biceps and 13 for triceps). These efforts correspond to the
reaction obtained in the muscle itself, and length corresponds to the length of the muscle
(the whole muscle length including tendons) measured in the FEA model, in that certain
elbow angle. Table 2 has these results if the form of force and the corresponding length for
this applying force for all muscles studied:
• Biceps muscles (long and short biceps);
• Triceps muscles (lateral, medial and long triceps).

258
Mathematics 2022, 10, 2441

Table 2. Axial efforts (N) in main simulated muscles, 150 N applied on wrist. Isometric contractions
at each angle value considered.

Long Biceps Short Biceps Lateral Triceps Medial Triceps Long Triceps
Elbow Angle Force Length Force Length Force Length Force Length Force Length
(◦ ) (N) (mm) (N) (mm) (N) (mm) (N) (mm) (N) (mm)
44 244 373 96 310 148 217 70 211 269 284
54 330 375 136 312 128 216 144 209 312 282
64 341 383 152 320 212 215 115 208 340 281
74 341 391 161 328 208 215 119 208 347 281
84 340 399 168 336 235 214 150 207 340 280
94 333 407 170 344 263 214 186 207 347 279
104 302 415 181 351 233 213 164 207 331 279
114 301 422 159 359 233 213 176 206 307 278
124 274 429 143 365 213 211 170 204 271 276
134 261 435 95 371 177 207 150 201 226 273
144 185 439 87 375 131 203 120 197 175 269
154 121 442 46 379 79 199 82 192 119 264
164 52 445 1 381 24 194 40 188 58 260

3.3. Mathematical Models for Muscle Efforts

3.3.1. Efforts for 150 N of Vertical Load
The obtained R-square values are 99.3%, 98.2%, 99.9%, 93.6%, and 97.2%, respectively
(see Table 3). Showing results in graph form, one graph for biceps muscles and the other
for triceps muscles, a tendency can be observed in both of them. Figure 13 shows the efforts
of biceps muscles at each angle (dots) and the mathematical curves (dash point line) that
approximate them. The effort in muscle (y-axis on the graph) is related to the elbow angle
(x-axis on the graph). The red color is associated with long biceps and the blue color is
associated with short biceps. These curves represent the total effort of each muscle involved.
Thus, long biceps, for a constant vertical load on the wrist, including forearm mass, reach
the maximum effort (340 N) between 60–90◦ , while for short biceps the maximum effort
(170 N) is achieved for the range of 90–110◦ .

Table 3. The equations developed as polynomial curves of grade 6th, based on the general formulation
of Equation (4), has the following constant values for each muscle involved.

Muscle a0 a1 a2 a3 a4 a5 a6 SSE R2 RMSE

−2323 136.6 −2.7 0.02579 −0.0001193 2.116 × 10−7
Long biceps (−3730, (55.39, (−4.477, (0.007278, (−0.0002116, (3.433 × 10−8 , 0 694.5 0.993 10.76
−916.1) 217.8) −0.9232) 0.04431) −2.692 × 10−5 ) 3.889 × 10−7 )
−409 24.12 −0.4361 0.004284 −2.182 × 10−5 4.28 × 10−8
Short biceps (−1729, (−52.09, (−2.104, (−0.01309, (−0.0001085, (−1.236 × 10−7 , 0 367.3 0.982 7.82
911.5) 100.3) 1.232) 0.02166) 6.486 × 10−5 ) 2.092 × 10−7 )
−1550 110.4 −2.702 0.03508 −0.0002503 9.18 × 10−7 −1.354 × 10−9
Long Triceps (−3477, (−23.88, (−6.433, (−0.01795, (−0.0006584, (−7 × 10−7 , (−3.945 × 10−9 , 102.3 0.999 4.13
377.3) 244.7) 1.028) 0.0881) 0.0001577) 2.536 × 10−6 ) 1.236 × 10−9 )
−6448 442 −11.9 0.1638 −0.001217 4.643 × 10−6 −7.144 × 10−9
Medium (−1.38 × 104 , (−70.71, (−26.14, (−0.0386, (−0.002775, (−1.534 × 10−6 , (−1.703 × 10−8 , 1491 0.936 15.76
Triceps 908.4) 954.7) 2.344) 0.3662) 0.0003408) 1.082 × 10−5 ) 2.743 × 10−9 )
4661 −311.3 8.397 −0.1143 0.0008424 −3.214 × 10−6 4.975 × 10−9
Lateral (−3110, (−852.8, (−6.645, (−0.3281, (−0.0008029, (−9.738 × 10−6 , (−5.468 × 10−9 , 1664 0.972 16.65
Triceps
1.243 × 104 ) 230.3) 23.44) 0.09946) 0.002488) 3.309 × 10−6 ) 1.542 × 10−8 )

Analogously, Figure 14 shows the efforts of the triceps muscles (dots) and the mathematical
curves (dash point line) which approximate them. The effort in the muscle (y-axis in the graph)
is related to the elbow angle (x-axis in the graph). In the figure, the red color is associated with
medium triceps, the green color is associated with long triceps, and the blue color is associated
with the lateral triceps. These curves can be used to know efforts on the three muscles involved
in the triceps FEM model. Additionally, as an example, if a muscle had limited effort for any

259
Mathematics 2022, 10, 2441

reason, we could know which positions of the elbow joint are not recommendable to use.
Therefore, for an isometric contraction with 150 N as a vertical load on the wrist (including
the forearm mass), for a maximum effort of 200 N in the lateral triceps, it is necessary to avoid
placing the elbow joint between 65◦ and 130 degrees approximately.

Figure 13. Calculated (dots) and tendency polynomial equation (dash point line) comparison, for
biceps forces (y-axis). Isometric contractions on the elbow at different positions, from 44 to 164 degrees
(x-axis) and 150 N are loaded on the wrist.

Figure 14. Comparison of the (continuous line) calculated and (dash point line) comparison, for tri-
ceps forces (y-axis). Isometric contractions on the elbow at different positions, from 44 to 164 degrees
(x-axis) and 150 N are loaded on the wrist.

260
Mathematics 2022, 10, 2441

3.3.2. General Formulation

Applying the mathematical approach described above, equations of efforts for the five
muscles involved will be presented, for any wrist load and elbow angle.
The approximation is done according to a fourth-grade polynomial with the
general equation:

f (α, w) = p00 + p10·α + p01·w + p11·w·α + p20·αˆ2 + p21·w·αˆ2 + p30·αˆ3

(12)
+ p31·w·αˆ3 + p40·αˆ4
where α is the angle of the elbow and w is the load applied to the wrist.
All coefficients have been calculated for a 95% confidence bound.
The surface of Figure 15 shows the tendency of the effort appearing in the short biceps
muscle as a result of applying a vertical load on the wrist. As an example, for a constant
vertical load of 750 N (including forearm mass), we could limit the short biceps effort to
200 N if we put the elbow joint angle between 150◦ to 164◦ .

’ͲͲൌ͹͹ǤͻͳሺെͳͺͶǤ͵ǡ͵ͶͲǤʹሻ ‘‘†‡••‘ˆˆ‹–ǣ
’ͳͲൌെ͵Ǥ͸ͳͺሺെͳͶǤ͹͹ǡ͹Ǥͷ͵Ͷሻ ǣͻǤͺʹʹ‡൅ͲͶ
’ͲͳൌെͲǤͺ͵Ͷ͵ሺെͳǤͳͺ͵ǡെͲǤͶͺͷͷሻ Ǧ•“—ƒ”‡ǣͲǤͻͻͳ͸
’ʹͲൌͲǤͲ͸ͳ͵ͷሺെͲǤͳͲ͹ͻǡͲǤʹ͵Ͳ͸ሻ †Œ—•–‡† Ǧ•“—ƒ”‡ǣ
’ͳͳൌͲǤͲͶ͸ͷͶሺͲǤͲ͵ͶͻͶǡͲǤͲͷͺͳ͵ሻ ͲǤͻͻͳ͵
’͵ͲൌെͶǤʹͶͷൈ ͳͲ−ͲͶ ሺെͲǤͲͲͳͷʹ͸ǡ ǣʹʹǤͻͺ
ͲǤͲͲͲ͸͹͹Ͷሻ 
’ʹͳൌെʹǤͺͲ͵ൈͳͲെͲͶሺെͲǤͲͲͲ͵ͻͺͺǡ
െͲǤͲͲͲͳ͸ͳͻሻ
’ͶͲ ൌ ͳǤͲʹʹൈ ͳͲെͲ͸ ሺെͳǤͷͻͷൈ
ͳͲെͲ͸ǡ͵Ǥ͸ͶൈͳͲെͲ͸ሻ
’͵ͳ ൌ ͳǤ͸ͻͷൈ ͳͲെͲ͹ ሺെʹǤͲͺͳൈ
ͳͲെͲ͹ǡͷǤͶ͹ͳൈͳͲെͲ͹ሻ

Figure 15. General formulation of short biceps efforts in function of wrist load (W) and elbow angle
(α) and surface reconstruction of the proposed polynomial.

In the range of the example of the previous figure, 150–164◦ , we could find an effort
on long biceps bigger than in short biceps, reaching around 600 N. Figures 16–19 collect the
results of the surface reconstruction for the rest of studied muscles.
’ͲͲൌെʹͲͳͻሺെʹͶ͸ͳǡെͳͷ͹ͺሻ ‘‘†‡••‘ˆˆ‹–ǣ
’ͳͲൌͻͶǤͺͳሺ͹͸ǤͲ͵ǡͳͳ͵Ǥ͸ሻ ǣʹǤ͹ͺ͸‡൅Ͳͷ
’ͲͳൌͲǤͲͶ͵ͺ͸ሺെͲǤͷͶ͵ͷǡͲǤ͸͵ͳʹሻ Ǧ•“—ƒ”‡ǣͲǤͻͻ͵ͺ
’ʹͲൌെͳǤͷͶͶሺെͳǤͺʹͻǡെͳǤʹͷͻሻ †Œ—•–‡†Ǧ•“—ƒ”‡ǣͲǤͻͻ͵ͷ
’ͳͳൌͲǤͲͷ͸Ͳ͸ሺͲǤͲ͵͸ͷ͵ǡͲǤͲ͹ͷͷͺሻ ǣ͵ͺǤ͹
’͵ͲൌͲǤͲͳͲͶ͹ሺͲǤͲͲͺ͸ͳǡͲǤͲͳʹ͵ʹሻ 
’ʹͳ ൌ െͲǤͲͲͲ͵Ͷ͸͵ ሺെͲǤͲͲͲͷͶͷͺǡ
െͲǤͲͲͲͳͶ͸ͺሻ
’ͶͲ ൌ െʹǤͷͳ͸ൈ ͳͲെͲͷ ሺെʹǤͻͷ͹ൈ
ͳͲെͲͷǡെʹǤͲ͹ͷൈͳͲെͲͷሻ
’͵ͳൌͳǤͲͺʹൈͳͲെͲ͹ሺെͷǤʹ͹͹ൈͳͲെͲ͹ǡ
͹ǤͶͶʹൈͳͲെͲ͹ሻ

Figure 16. General formulation of long biceps efforts based on wrist load (W) and elbow angle (α)
and surface reconstruction of the proposed polynomial.

261
Mathematics 2022, 10, 2441

’ͲͲൌͳͲͺͳሺ͸ͲͲǤ͵ǡͳͷ͸ͳሻ ‘‘†‡••‘ˆˆ‹–ǣ
’ͳͲൌെͷͲǤ͹ሺെ͹ͳǤͳǡെ͵ͲǤ͵ሻ ǣ͵Ǥʹ͹͵‡൅Ͳͷ
’ͲͳൌെͲǤ͸ͲͶʹሺെͳǤʹͶͳǡͲǤͲ͵ʹͶሻ Ǧ•“—ƒ”‡ǣͲǤͻͺͶͳ
’ʹͲൌͲǤͺʹ͸͵ሺͲǤͷͳ͹ͳǡͳǤͳ͵͸ሻ †Œ—•–‡†Ǧ•“—ƒ”‡ǣͲǤͻͺ͵Ͷ
’ͳͳൌͲǤͲ͵ͺ͹͸ሺͲǤͲͳ͹ͷͻǡͲǤͲͷͻͻʹሻ ǣͶͳǤͻͷ
’͵Ͳ ൌ െͲǤͲͲͷ͸Ͳͳ ሺെͲǤͲͲ͹͸ͳ͵ǡ
െͲǤͲͲ͵ͷͻሻ
’ʹͳ ൌ െͻǤ͸ʹൈ ͳͲെͲͷ ሺെͲǤͲͲͲ͵ͳʹͶǡ
ͲǤͲͲͲͳʹሻ
’ͶͲ ൌͳǤ͵Ͷ͸ൈ ͳͲെͲͷ ሺͺǤ͸ͺͺൈ ͳͲെͲ͸ǡ
ͳǤͺʹͶൈͳͲെͲͷሻ
’͵ͳൌെ͸ǤͻͳͷൈͳͲെͲ͹ሺെͳǤ͵ͺͳൈͳͲെͲ͸ǡ
െʹǤʹ͸ʹൈͳͲെͲͻሻ

Figure 17. General formulation of lateral triceps efforts based on wrist load on the wrist (W) and
elbow angle (α) and surface reconstruction of the proposed polynomial.

’ͲͲൌെͳ͵Ǥͷ͸ሺെͳ͸ͳǤͻǡͳ͵ͶǤͺሻ ‘‘†‡••‘ˆˆ‹–ǣ
’ͳͲൌͲǤ͸͵͸͸ሺെͷǤ͸͸Ͷǡ͸Ǥͻ͵͹ሻ ǣ͵Ǥͳʹʹ‡൅ͲͶ
’ͲͳൌെͲǤʹ͸ͷ͸ሺെͲǤͶ͸ʹʹǡെͲǤͲ͸ͻͲͳሻ Ǧ•“—ƒ”‡ǣͲǤͻͻͻ͵
’ʹͲൌെͲǤͲͳͲ͵ͺሺെͲǤͳͲͷͻǡͲǤͲͺͷͳʹሻ †Œ—•–‡†Ǧ•“—ƒ”‡ǣͲǤͻͻͻʹ
’ͳͳൌͲǤͲ͸͹Ͳ͹ሺͲǤͲ͸Ͳͷ͵ǡͲǤͲ͹͵͸ሻ ǣͳʹǤͻ͸
’͵Ͳ ൌ ͹ǤͲ͵ͷൈ ͳͲെͲͷ ሺെͲǤͲͲͲͷͷͳǡ
ͲǤͲͲͲ͸ͻͳ͹ሻ
’ʹͳ ൌ െͲǤͲͲͲͶ͹Ͷ͵ ሺെͲǤͲͲͲͷͶͳǡ
െͲǤͲͲͲͶͲ͹ͷሻ
’ͶͲൌെͳǤ͸ͻͳൈͳͲെͲ͹ሺെͳǤ͸ͶͷൈͳͲെͲ͸ǡ
ͳǤ͵Ͳ͸ൈͳͲെͲ͸ሻ
’͵ͳ ൌ ͷǤͶ͵ͻൈ ͳͲെͲ͹ ሺ͵Ǥ͵ͳൈ ͳͲെͲ͹ǡ
͹Ǥͷ͸ͺൈͳͲെͲ͹ሻ

Figure 18. General formulation of long triceps efforts based on wrist load (W) and elbow angle (α)
and surface reconstruction of the proposed polynomial.

’ͲͲൌെͳͳͺǤͻሺെ͸ͶͺǤ͸ǡͶͳͲǤͻሻ ‘‘†‡••‘ˆˆ‹–ǣ
’ͳͲൌͷǤͷ͹͸ሺെͳ͸ǤͻʹǡʹͺǤͲ͹ሻ ǣ͵Ǥͻͺ‡൅Ͳͷ
’ͲͳൌͲǤʹͷͳ͵ሺെͲǤͶͷͲ͸ǡͲǤͻͷ͵͵ሻ Ǧ•“—ƒ”‡ǣͲǤͻ͸Ͳ͵
’ʹͲൌെͲǤͲͻͲͺͺሺെͲǤͶ͵ͳͺǡͲǤʹͷͲͳሻ †Œ—•–‡†Ǧ•“—ƒ”‡ǣͲǤͻͷͺ͸
’ͳͳൌͲǤͲͲͲʹͺͶ͵ሺെͲǤͲʹ͵ͲͷǡͲǤͲʹ͵͸ʹሻ ǣͶ͸Ǥʹ͸
’͵ͲൌͲǤͲͲͲ͸ͳ͸ሺെͲǤͲͲͳ͸Ͳ͵ǡͲǤͲͲʹͺ͵ͷሻ
’ʹͳ ൌ ͲǤͲͲͲʹʹ͵͸ ሺെͳǤͶͺൈ ͳͲെͲͷǡ
ͲǤͲͲͲͶ͸ʹሻ
’ͶͲൌെͳǤͶͺͳൈͳͲെͲ͸ሺെ͸Ǥ͹ͶͻൈͳͲെͲ͸ǡ
͵Ǥ͹ͺ͹ൈͳͲെͲ͸ሻ
’͵ͳൌെͳǤ͵͹ͷൈͳͲെͲ͸ሺെʹǤͳ͵ͷൈͳͲെͲ͸ǡ
െ͸ǤͳͷʹൈͳͲെͲ͹ሻ

Figure 19. General formulation of the medium triceps efforts based on wrist load (W) and elbow
angle (α) and surface reconstruction of the proposed polynomial.

4. Discussion
In this study, an analytical calculation method is proposed to obtain the forces in the
flexor muscles (long and short biceps) and extensors (long, lateral, and middle triceps) in
any configuration of the opening angle of the elbow joint and to any load applied to the
wrist. The final objective of this study is to allow the creation of a new type of finite element

262
Mathematics 2022, 10, 2441

that can be included in a musculoskeletal model that can be analyzed in different positions
and configurations without the need to repeat the model meshing process.
There are several studies along the same lines. Therefore, Martinez [12] is focused
on the analysis of a certain muscle in a certain load situation, analyzing what happens
in the muscle itself. Islan et al. [14], are focused on the muscles involved in the shoulder
joint, under static condition predicting the fatigue and cumulative damage to the muscles.
Alonso et al. [20] study the forces in the muscles involved in the human march under certain
load conditions, weight and step frequency. Park et al. [22] develop a predictive method for
muscle forces based on electromyography techniques. Teo et al. [26] are focused on the effect
of a collision on the neck muscles. As can be seen, none of them addresses the problem from
our perspective. The use of the finite element method is widely recognized as a calculation
tool for structural, static, and/or dynamic problems. However, characterization of the
muscle, so that it is easy to include it in an FEM model has not been addressed until now.
The main drawbacks presented by the studies carried out so far, in order to include
the muscle in a valid finite element model to analyze the behavior of a muscle within a
joint in various positions of this, are as follows:
• Studies have focused on one joint position, obtaining results related to muscle effort
that are difficult to extrapolate to different positions. Ramírez [12] obtains the results
for the anterior tibial muscle under study conditions, but moving the results to other
load situations should be difficult to solve.
• The equations obtained to characterize the muscle are difficult to implement in a FEM
model. Sachenkov et al. [29] use 1D elements to simulate muscles, but there is no single
equation for any muscle to implement it in another analysis. The same happens with
Alonso et al. [20], where the result is based on a static and physiological optimization
that integrates the forces of any Hill MTU unit.
• Several of these studies use 3D elements to represent the muscle, which allows a
better understanding of the internal behavior of the muscle itself (distribution of
stresses and strains, deformations, etc.) but does not allow the created mesh to be
reused in other positions of the joint. Therefore, Martínez [12] studies the anterior tibia
muscle in a certain position, and if the muscle length changes, a new mesh should
be developed. Another example is shown in research by Islan et al. [14], where they
analyze the shoulder muscles of a violin player holding the instrument, and so, if the
shoulder position changes, the mesh will do as well, remeshing the muscles into their
new position.
With this work, and in a first approximation, we have obtained the equations that
govern the behavior of the flexor and extensor muscles of the elbow, only in isometric
contractions and referring to this behavior to its response in the form of force provided,
throughout the flexion path of the joint for any load applied to the wrist.
For example, to calculate the estimated force that must appear in each of the joint
muscles to support a load applied to the wrist of 235 N with the elbow forming an angle of
119 degrees, we need to substitute the values of α and w in Equation (12), depending on
which muscle we want to analyze.
Changing the indicated values in Equation (12) would obtain the following estimated
forces in the muscles.
Vertical downward load, strength in Biceps muscles.
• Long biceps: 477 N;
• Short biceps: 245 N.
Upward vertical load, strength in the triceps muscles.
• Long triceps: 438 N;
• Middle triceps: 231;
• Side triceps: 304.
In addition to what has been said, the equations obtained to characterize the muscles
in their response are easy to implement mathematically, which will undoubtedly help in

263
Mathematics 2022, 10, 2441

the future in the possibility of creating a new type of finite element that responds to these
behavioral equations obtained.
With the creation of this new type of finite element to represent muscles, a very
interesting possibility is opened to analyze the behavior of the joint in different positions,
with variable loads, different load cycles, etc. notably reducing the study work since
the meshing process of the model is eliminated by being able to reuse the mesh for the
desired positions.
Regarding the analysis of the results obtained, it is worth mentioning the following
important aspects of the study.
Applying the proposed methodology and using the finite element method, we can
obtain the forces that must appear in the muscles involved in the elbow joint to keep it in
balance, in any position between 44 and 164 degrees of opening, with a load applied to the
wrist of 150 N (Table 2, Figures 13 and 14).
Through this same analysis it is possible to obtain the tensions that appear in the
insertions of the tendons of the joint, which allows us to foresee damage before certain load
situations (Figures 11 and 12).
This same analysis allows us to obtain a map of tensions in the rigid tissues of the joint,
bones, and cartilage. This allows a detailed study of damage to the elements mentioned
(Figures 11 and 12).
There are no previous studies that carry out studies similar to the proposed study,
so a direct comparison with what has already been done is not possible. However, the
trend observed in muscle effort during joint movement is similar to what other studies
have published, [34], where the results of the torque in elbow vs. elbow angle shape are
consistent with our results and [35] where the relation between the torque of the elbow and
angle of the elbow is similar to our expressions, although these studies measure the moment
in the joint (or what is the same, the sum of the moments generated by each muscle) and
not the independent force in each muscle the results presented in these investigations are
similar to those presented in this study in terms of the evolution of the moment in the joint,
or of the force in the muscle in our case, along the angle of the joint.
The correlation also presents high R values, which confirms the validity of the analysis
(Figures 13 and 14).
Through the application of classical mechanics and the balance of forces and moments
in the joint, we can extrapolate the results obtained for a specific load, 150 N in our case, to
different loads applied to the joint (Figures 15–19).
Finally, it should be mentioned that the equations that predict the force in each of the
muscles in the joint are obtained as a function of the opening angle of the elbow and the
applied load.

5. Conclusions
The use of a simplified model of the muscle, as a 1D element with axial forces only
(ROD type element), into a FEA model allows one to calculate the forces that must be
generated in the muscles involved in the joint.
It is also possible to predict the force required in each muscle for different load situations.
On the basis of the same analysis, the stresses that appear in other elements of the joint,
bones, tendon insertions, or cartilage, can be obtained. With the obtained characteristic
equations, different lines of research are opened focused on a better characterization of
the muscles. It seems necessary to evolve the equations obtained in this study to obtain
those that give the force of the muscle as a function of the speed of contraction. Or, based
on studies conducted to assess muscle fatigue, they could be modified to predict it in a
moving joint.
The evolution of this study should follow the following stages:
• Validation of the theoretical model obtained through testing with individuals with
whom it is possible to measure the forces in cases of biceps and/or triceps load that,
by means of comparison with this study, allows for knowing its validity.

264
Mathematics 2022, 10, 2441

• Modification of the equations obtained in this study in order to obtain their evolution
for concentric or eccentric contractions.
• Creation of a new type of 1D element for application in finite element models that
allows the behavior of a muscle to be characterized according to the equations obtained
for any type of contraction, isometric, concentric, or eccentric.
Once these phases are completed, the work could be extrapolated to other types of
muscle (flat, or peniform) in order to characterize more muscles that allow us to study new
joints according to the proposed methodology.
One limitation of the presented study is the approximation of the muscle to the 1D
element. This implies that following this methodology cannot obtain information regarding
what happens inside the muscle.
Another one is that the analysis is performed for a certain size of elbow joint, and so the
FEA model created is valid for this joint size; in any other case (children, older people, etc.),
the FEA model should be resized, and the same methodology should be applied.

Author Contributions: Conceptualization, F.B.H. and J.A.J.M.; data curation, E.L.U., J.D.C.-M. and
R.D.; formal analysis, E.L.U., F.B.H. and R.D.; investigation, E.L.U., F.B.H. and J.A.J.M.; methodology,
F.B.H. and R.D.; project administration, F.B.H.; supervision, F.B.H., R.D. and J.A.J.M.; writing—original
draft, E.L.U. and J.D.C.-M.; writing—review and editing, R.D. All authors have read and agreed to the
published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
In this appendix, the sub-Figures included in Figures 11 and 12 are going to be collected
with higher quality.

44°

Figure A1. Cont.

265
Mathematics 2022, 10, 2441

54°

64°

Figure A1. Cont.

266
Mathematics 2022, 10, 2441

74°

84°

Figure A1. Cont.

267
Mathematics 2022, 10, 2441

94°

104°

Figure A1. Cont.

268
Mathematics 2022, 10, 2441

114°

124°

Figure A1. Cont.

269
Mathematics 2022, 10, 2441

134°

144°

Figure A1. Cont.

270
Mathematics 2022, 10, 2441

154°

164°

Figure A1. Different Elbow Angle Configuration: Bone Von Mises Stress Distribution Triceps Loading
Case (150 N on wrist).

271
Mathematics 2022, 10, 2441

44°

54°

Figure A2. Cont.

272
Mathematics 2022, 10, 2441

64°

74°

Figure A2. Cont.

273
Mathematics 2022, 10, 2441

84°

94°

Figure A2. Cont.

274
Mathematics 2022, 10, 2441

104°

114°

Figure A2. Cont.

275
Mathematics 2022, 10, 2441

124°

134°

Figure A2. Cont.

276
Mathematics 2022, 10, 2441

144°

Figure A2. Different Elbow Angle Configuration: Bone Von Mises Stress Distribution Triceps Loading
Case (150 N on wrist).

References
1. Strouboulis, T.; Babuška, I.; Whiteman, J.R. The Finite Element Method and Its Reliability; Clarendon Press: Oxford, UK, 2001;
ISBN 978-0-19-850276-0.
2. Babuska, I.; Whiteman, J.; Strouboulis, T. Finite Elements: An Introduction to the Method and Error Estimation; OUP: Oxford, UK,
2010; ISBN 978-0-19-850669-0.
3. Pawełko, P.; Jastrz˛ebski, D.; Parus, A.; Jastrz˛ebska, J. A new measurement system to determine stiffness distribution in machine
tool workspace. Arch. Civ. Mech. Eng. 2021, 21, 49. [CrossRef]
4. Wang, D.; Zhang, S.; Wang, L.; Liu, Y. Developing a Ball Screw Drive System of High-Speed Machine Tool Considering Dynamics.
IEEE Trans. Ind. Electron. 2021, 69, 4966–4976. [CrossRef]
5. Li, Z.; Oger, G.; Le Touzé, D. A partitioned framework for coupling LBM and FEM through an implicit IBM allowing non-
conforming time-steps: Application to fluid-structure interaction in biomechanics. J. Comput. Phys. 2021, 449, 110786. [CrossRef]
6. Della Rosa, N.; Bertozzi, N.; Adani, R. Biomechanics of external fixator of distal radius fracture, a new approach: Mutifix Wrist.
Musculoskelet. Surg. 2020, 106, 89–97. [CrossRef] [PubMed]
7. Zhang, N.-Z.; Xiong, Q.-S.; Yao, J.; Liu, B.-L.; Zhang, M.; Cheng, C.-K. Biomechanical changes at the adjacent segments induced
by a lordotic porous interbody fusion cage. Comput. Biol. Med. 2022, 143, 105320. [CrossRef] [PubMed]
8. Denozière, G.; Ku, D.N. Biomechanical Comparison between Fusion of Two Vertebrae and Implantation of an Artificial Interver-
tebral Disc. J. Biomech. 2006, 39, 766–775. [CrossRef] [PubMed]
9. Samani, A.; Bishop, J.; Yaffe, M.J.; Plewes, D.B. Biomechanical 3-D finite element modeling of the human breast using MRI data.
IEEE Trans. Med. Imaging 2001, 20, 271–279. [CrossRef]
10. Jaecques, S.; Van Oosterwyck, H.; Muraru, L.; Van Cleynenbreugel, T.; De Smet, E.; Wevers, M.; Naert, I.; Sloten, J.V. Individualised,
micro CT-based finite element modelling as a tool for biomechanical analysis related to tissue engineering of bone. Biomaterials
2003, 25, 1683–1696. [CrossRef]
11. Renner, S.M.; Natarajan, R.N.; Patwardhan, A.G.; Havey, R.M.; Voronov, L.I.; Guo, B.Y.; Andersson, G.B.; An, H.S. Novel model to
analyze the effect of a large compressive follower pre-load on range of motions in a lumbar spine. J. Biomech. 2007, 40, 1326–1332.
[CrossRef]
12. Martínez, A.M.R. Modelado y simulación del tejido músculo-esquelético. Validación Experimental con el Músculo Tibial Anterior
de Rata. Ph.D. Thesis, Universidad de Zaragoza, Zaragoza, Spain, 2011. Available online: http://purl.org/dc/dcmitype/Text.
(accessed on 5 November 2021).
13. Weiss, J.A.; Gardiner, J.C.; Ellis, B.J.; Lujan, T.J.; Phatak, N.S. Three-dimensional finite element modeling of ligaments: Technical
aspects. Med. Eng. Phys. 2005, 27, 845–861. [CrossRef]
14. Islan, M.; Carvajal, J.; Pedro, P.S.; D’Amato, R.; Juanes, J.A.; Soriano, E. Linear Approximation of the Behavior of the Rotator
Cuff under Fatigue Conditions. Violinist Case Study. In Proceedings of the ACM 5th International Conference on Technological
Ecosystems for Enhancing Multiculturality, Cádiz, Spain, 18–20 October 2017; p. 58.

277
Mathematics 2022, 10, 2441

15. Sachenkov, O.A.; Hasanov, R.F.; Andreev, P.S.; Konoplev, Y.G. Numerical Study of Stress-Strain State of Pelvis at the Proximal
Femur Rotation Osteotomy. Russ. J. Biomech. 2016, 20, 220–232. [CrossRef]
16. Martins, J.A.C.; Pato, M.P.M.; Pires, E.B. A finite element model of skeletal muscles. Virtual Phys. Prototyp. 2006, 1, 159–170.
[CrossRef]
17. Tang, C.; Tsui, C.; Stojanovic, B.; Kojic, M. Finite element modelling of skeletal muscles coupled with fatigue. Int. J. Mech. Sci.
2007, 49, 1179–1191. [CrossRef]
18. Syomin, F.A.; Tsaturyan, A.K. Mechanical model of the left ventricle of the heart approximated by axisymmetric geometry. Russ.
J. Numer. Anal. Math. Model. 2017, 32, 327–337. [CrossRef]
19. Perreault, E.J.; Sandercock, T.G.; Heckman, C.J. Hill Muscle Model Performance during Natural Activation and Electrical
Stimulation. In Proceedings of the 23rd Annual International Conference of the IEEE Engineering in Medicine and Biology
Society, Istanbul, Turkey, 25–28 October 2001; Volume 2, pp. 1248–1251.
20. Alonso, F.J.; Galán-Marín, G.; Salgado, D.R.; Pàmies Vilà, R.; Font Llagunes, J.M. Cálculo de Esfuerzos Musculares en la Marcha
Humana Mediante Optimización Estática-Fisiológica. In Proceedings of the XVIII Congreso Nacional de Ingeniería Mecánica,
Ciudad Real, Spain, 3–5 November 2010; pp. 1–9.
21. Holzbaur, K.R.S.; Murray, W.M.; Delp, S.L. A Model of the Upper Extremity for Simulating Musculoskeletal Surgery and
Analyzing Neuromuscular Control. Ann. Biomed. Eng. 2005, 33, 829–840. [CrossRef]
22. Park, W.-I.; Lee, H.-D.; Kim, J. Estimation of isometric joint torque from muscle activation and length in intrinsic hand muscle.
In Proceedings of the 2008 International Conference on Control, Automation and Systems, Seoul, Korea, 14–17 October 2008;
pp. 2489–2493. [CrossRef]
23. Soechting, J.F.; Flanders, M. Evaluating an Integrated Musculoskeletal Model of the Human Arm. J. Biomech. Eng. 1997,
119, 93–102. [CrossRef]
24. Zajac, F.E. Muscle and Tendon: Properties, Models, Scaling, and Application to Biomechanics and Motor Control. Crit. Rev.
Biomed Eng. 1989, 17, 359–411.
25. Lechosa Urquijo, E.; Blaya Haro, F.; D’Amato, R.; Juanes Méndez, J.A. Finite Element Model of an Elbow under Load, Muscle
Effort Analysis When Modeled Using 1D Rod Element. In Proceedings of the Eighth International Conference on Technological
Ecosystems for Enhancing Multiculturality, Salamanca, Spain, 21–23 October 2020; Association for Computing Machinery: New
York, NY, USA, 21 October 2020; pp. 475–482.
26. Teo, E.C.; Zhang, Q.H.; Qiu, T.X. Finite Element Analysis of Head-Neck Kinematics Under Rear-End Impact Conditions. In
Proceedings of the 2006 International Conference on Biomedical and Pharmaceutical Engineering, Singapore, 11–14 December
2006; pp. 206–209.
27. Donahue, T.L.H.; Hull, M.L.; Rashid, M.M.; Jacobs, C.R. A Finite Element Model of the Human Knee Joint for the Study of
Tibio-Femoral Contact. J. Biomech. Eng. 2002, 124, 273–280. [CrossRef]
28. Abidin, N.A.Z.; Kadir, M.R.A.; Ramlee, M.H. Three Dimensional Finite Element Modelling and Analysis of Human Knee
Joint-Model Verification. J. Phys. Conf. Ser. 2019, 1372, 012068. [CrossRef]
29. Sachenkov, O.A.; Hasanov, R.; Andreev, P.; Konoplev, Y. Determination of Muscle Effort at the Proximal Femur Rotation Osteotomy.
IOP Conf. Series: Mater. Sci. Eng. 2016, 158, 012079. [CrossRef]
30. Jesal, N. Parekh Using Finite Element Methods to Study Anterior Cruciate Ligament Injuries: Understanding the Role of ACL
Modulus and Tibial Surface Geometry on ACL Loading. Ph.D. Thesis, The University of Michigan, Ann Arbor, MI, USA, 2013.
31. CES EduPack Bulletin: January 2017. Available online: https://www.grantadesign.com/newsletters/ces-edupack-bulletin-ces-
edupack-2017-new-products-database-symposia-deadlines-shared-resources-webinars-and-more/ (accessed on 20 June 2022).
32. Bruno, S.; José, M.; Filomena, S.; Vítor, C.; Demétrio, M.; Karolina, B. The Conceptual Design of a Mechatronic System to Handle
Bedridden Elderly Individuals. Sensors 2016, 16, 725. [CrossRef] [PubMed]
33. Arcila Arango, J.C.; Cardona Nieto, D.; Giraldo, J.C. Abordaje Físico-Matemático Del Gesto Articular. Available online: https:
//www.efdeportes.com/efd171/abordaje-fisico-matematico-del-gesto-articular.htm (accessed on 5 November 2021).
34. Loss, J.F.; Candotti, C.T. Comparative Study between Two Elbow Flexion Exercises Using the Estimated Resultant Muscle Force.
Braz. J. Phys. Ther. 2008, 12, 502–510. [CrossRef]
35. Murray, W.M.; Delp, S.L.; Buchanan, T.S. Variation of Muscle Moment Arms with Elbow and Forearm Position. J. Biomech. 1995,
28, 513–525. [CrossRef]

278
 mathematics

Article
Miura-Ori Inspired Smooth Sheet Attachments for
Zipper-Coupled Tubes
Dylan C. Webb 1 , Elissa Reynolds 1 , Denise M. Halverson 1, * and Larry L. Howell 2

1 Department of Mathematics, Brigham Young University, Provo, UT 84602, USA; dylancw2@byu.edu (D.C.W.);
elissar2@byu.edu (E.R.)
2 Department of Mechanical Engineering, Brigham Young University, Provo, UT 84602, USA; lhowell@byu.edu
* Correspondence: halverson@math.byu.edu

Abstract: Zipper-coupled tubes are a broadly applicable, deployable mechanism with an angular
surface that can be smoothed by attaching an additional smooth sheet pattern. The existing design
for the smooth sheet attachment, however, leaves small gaps that can only be covered by adding
flaps that unfold separately, limiting applicability in situations requiring a seamless surface and
simultaneous deployment. We provide a novel construction of the smooth sheet attachment that
unfolds simultaneously with zipper-coupled tubes to cover the entire surface without requiring
additional actuation and without inhibiting the tubes’ motion up to an ideal, unfolded state of
stability. Furthermore, we highlight the mathematics underlying the design and motion of the new
smooth sheet pattern, thereby demonstrating its rigid-foldability and compatibility with asymmetric
zipper-coupled tubes.

Keywords: zipper-coupled tubes; Miura-ori pattern; deployable mechanism; origami inspired design;
Citation: Webb, D.C.; Reynolds, E.; smooth sheet attachment
Halverson, D.M.; Howell, L.L.
Miura-Ori Inspired Smooth Sheet MSC: 74-10; 51E24
Attachments for Zipper-Coupled
Tubes. Mathematics 2022, 10, 2643.
https://doi.org/10.3390/
math10152643 1. Introduction
Academic Editors: Higinio Rubio Origami is the basis for many deployable mechanisms, including self-scaling, modular
Alonso, Alejandro Bustos, Jesus robots [1], satellite reflectarray antennas that pack efficiently [2], and multimodal biomed-
Meneses Alonso and Enrique ical devices that actuate electromagnetically [3]. Zipper-coupled tubes are multistable
Soriano-Heras origami structures that fold up compactly and unfold bidirectionally to fill space and resist
compression [4]. An asymmetric generalization of zipper-coupled tubes with smooth sheet
Received: 29 June 2022
attachments was introduced previously [5]. Together, these origami-based mechanisms
Accepted: 24 July 2022
form a deployable device with a smooth surface that is advantageous in applications, such
Published: 28 July 2022
as prefab architecture, when drivability and walkability are important, and in smooth
Publisher’s Note: MDPI stays neutral medical devices, when sharp edges could harm the body. The smooth sheet attachment we
with regard to jurisdictional claims in presented previously, however, does not fully cover the surface of zipper-coupled tubes
published maps and institutional affil- without additional flaps that actuate separately [5]. We rectify this problem by offering an
iations. alternative, Miura-ori inspired [6–8], construction of a smooth sheet attachment that fully
covers the surface of, and deploys simultaneously with, asymmetric zipper-coupled tubes.
In this paper we briefly review the design of asymmetric zipper-coupled tubes and
then demonstrate how to (1) construct a Miura-ori inspired smooth sheet attachment
Copyright: © 2022 by the authors.
without gaps that attaches to the mountain folds of asymmetric zipper-coupled tubes and
Licensee MDPI, Basel, Switzerland.
This article is an open access article
(2) handle design variations in the symmetric case. We present a mathematically robust
distributed under the terms and
design method, decomposing the attachment into pairs of compatible, tessellating cells
conditions of the Creative Commons inspired by the Miura-ori pattern and then defining the cells’ vertices throughout the
Attribution (CC BY) license (https:// folding motion, thereby confirming rigid-foldability of the smooth sheet attachment and
creativecommons.org/licenses/by/ highlighting the mathematical processes involved in mechanism design. Our Miura-ori-
4.0/). based smooth sheet attachment expands the utility of asymmetric zipper-coupled tubes;

Mathematics 2022, 10, 2643. https://doi.org/10.3390/math10152643 279 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 2643

combined, these origami-inspired mechanisms are ideal for applications requiring a rigidly
deployable structure with a smooth surface.

2. Zipper-Coupled Tubes Review

Zipper-coupled tubes are an origami pattern consisting of two or more deployable
tubes coupled together in a zipper fashion that makes their motion compatible. Designed by
Filipov et al., this structure is remarkable for its ability to deploy from a flat, stowable form
into a stable, space-filling structure with only one degree of freedom [4]. In this section
we review briefly the construction of asymmetric zipper-coupled tubes from two tube
segments that are each, in turn, composed of two degree-four vertex cells; the interested
reader may refer to our previous work for a thorough treatment of the construction [5].
Throughout this paper we use the notational convention that an arbitrary vector x has
unit length direction vector x̂ with length x, and hence

x = xx̂

Thus, the notation x̂ always represents a unit length vector associated with a vector denoted
x, where x = xx̂.
The building blocks for zipper-coupled tubes are the compatible degree-four vertex
cells C1 and C2 illustrated in Figures 1 and 2. When combined, they form a single tube seg-
ment [5]. To couple correctly and satisfy rigid and flat-foldability [9], the design constraints
of these degree-four vertex cells include:

α3 = π − α1 α1 + α3 = α2 + α4 α1 < α2
α4 = π − α2 d sin a4 = b sin a1 α1 + α2 ≤ π

If α1 + α2 < π, the zipper-coupled tubes are called asymmetric, having a characteristic

tilt and a customizable unfolding motion. In the special case that α1 + α2 = π and c = a,
the two cells in Figure 1 are congruent, symmetric Miura-ori cells and the resulting zipper-
coupled tubes are those constructed by Filipov et al. [4].

Figure 1. (a) Vectors defining the basic asymmetric degree-four vertex cell, C1 . (b) Vectors defining
the complementary degree-four vertex cell, C2 . (Adapted from [5]).

280
Mathematics 2022, 10, 2643

Figure 2. (a) Vertices of the basic asymmetric degree-four vertex cell, C1 . (b) Vertices of the comple-
mentary degree-four vertex cell, C2 . The points Xi∗ identify with Xi . (Adapted from [5]).

Like a Miura-ori cell, the basic degree-four vertex cell has one degree of freedom in
its motion. Let the basic and complementary cells lie flat in the xz-plane when unfolded,
as depicted in Figures 1 and 2, and let the angles between the xz-plane and panels σ1 and σ4 ,
respectively, be equal as σ1 and σ4 fold toward each other (see Figure 1a). Call this motion
parameter γ. Then, placing F0 at the origin and fixing F1 on the negative z-axis in R3 (see
Figure 2a), the motion of the basic cell is determined by the following vector paths:

â(γ) = [0, 0, −1] T

b̂(γ) = [− sin α1 cos γ, sin α1 sin γ, − cos α1 ] T
T
a x c 3 + b x a y c 3 + by a x bx sin2 γ + ay by cos2 γ + k2 sin2 γ cos2 γ
ĉ(γ) = , ,−
k cos γ k sin γ a2x sin2 γ + a2y cos2 γ + k2 sin2 γ cos2 γ
d̂(γ) = [sin α4 cos γ, sin α4 sin γ, − cos α4 ] T

where

a x = sin α1 cos α4 − sin α4 cos α1 bx = sin α1 cos α3 − sin α4 cos α2

ay = sin α1 cos α4 + sin α4 cos α1 by = sin α1 cos α3 + sin α4 cos α2
k = 2 sin α1 sin α4 ĉ = c1 , c2 , c3 

The same vectors define the motion of both basic and complementary cells. By combin-
ing one basic cell (C1 ) and one complementary cell (C2 ), we obtain the first tube segment in a
zipper-coupled pair, illustrated in Figure 3a. Its vertices, identified with their corresponding
position vectors, are given by:

E0 (γ) = d(γ) F0 (γ) = 0 G0 (γ) = b(γ)

E1 (γ) = a(γ) + d(γ) F1 (γ) = a(γ) G1 (γ) = a(γ) + b(γ)
E2 (γ) = a(γ) + c(γ) + d(γ) F2 (γ) = a(γ) + c(γ) G2 (γ) = a(γ) + b(γ) + c(γ)
E3 (γ) = c(γ) + d(γ) F3 (γ) = c(γ) G3 (γ) = b(γ) + c(γ)

281
Mathematics 2022, 10, 2643

Figure 3. (a) The vertices in the first origami tube segment in the construction of an asymmetric
zipper-coupled tube segments pair. (b) The vertices in the second origami tube segment in the
construction of an asymmetric zipper-coupled tube segments pair. (Adapted from [5]).

The second tube segment is a copy of the first, but rotated 180◦ about the y-axis and
then shifted by an offset vector so that it attaches to the first tube segment along the vertical
creases (illustrated in Figure 3b). Let C3 and C4 denote the copies of C1 and C2 , respectively,
comprising the second tube segment. To define the motion of the second tube segment, let
x̄ denote the 180◦ rotation of a vector x about the y-axis and define the offset vector:
 
b cos α1 + d cos α4
s( γ ) = 1 + a( γ ) + c( γ )
2a

Then the vertices of the second tube segment are defined by:

E0 (γ) = d̄(γ) + s(γ) F0 (γ) = s(γ) G0 (γ) = b̄(γ) + s(γ)
E1 (γ) = ā(γ) + d̄(γ) + s(γ) F1 (γ) = ā(γ) + s(γ) G1 (γ) = ā(γ) + b̄(γ) + s(γ)
E2 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) F2 (γ) = ā(γ) + c̄(γ) + s(γ) G2 (γ) = ā(γ) + b̄(γ) + c̄(γ) + s(γ)
E3 (γ) = c̄(γ) + d̄(γ) + s(γ) F3 (γ) = c̄(γ) + s(γ) G3 (γ) = b̄(γ) + c̄(γ) + s(γ)

The pair of tube segments, with vertices positioned as indicated above, form a single
component in a pair of zipper-coupled tubes (Figure 3b), which will be denoted as Z0 .
The zipper-coupled tubes can be extended by taking multiple copies of Z0 and attaching
them end-to-end. In particular, for i = 1 . . . n, let

Z i = Z0 + i (d − b)

Then
9
n
Z= Zi
i =0

denotes zipper-coupled tubes with n + 1 components.

Of critical importance, at a certain point in the motion of Z , the upper (or lower)
creases simultaneously become coplanar, as bolded in Figure 4b. This state is the ideal state,
and it occurs at a parameter value:
⎛0 #$ ⎞
1 %2
1
⎜ 2 ay by − a2x + k2 + ay by + a2x + k2 − 4a2y a x bx ⎟
γ0 = cos−1 ⎜
⎝
⎟
⎠
2k2

The value γ0 will be the terminal value for the deployment of Z with a smooth sheet
attachment. Thus, by construction, the smooth sheet attachment lies flat on the surface of
the zipper-coupled tubes when γ = γ0 and folds up with the zipper-coupled tubes until
γ = π/2, at which point the entire structure lies in a plane and has no volume.

282
Mathematics 2022, 10, 2643

Figure 4. (a) The top smooth sheet attachment (S1 is lightly shaded, S4 is darkly shaded, and co-
planar ridges are bolded). (b) Adding the top and bottom smooth sheet attachments to asymmetric
zipper-coupled tubes (co-planar ridges are bolded). (c) The bottom smooth sheet attachment (S3 is
lightly shaded, S2 is darkly shaded, co-planar ridges are bolded).

3. Asymmetric Smooth Sheet Attachment

In this section we design a Miura-ori inspired smooth sheet attachment that actuates
simultaneously with the asymmetric zipper-coupled tubes pattern, folding up flat and
unfolding into a rigid sheet without gaps in the ideal state. This additional pattern broad-
ens the potential applications of the zipper-coupled tubes structure and expands on the
design of the attachment described previously [5]. As with the zipper-coupled tubes, we
decomposed the overall structure into its basic units: four distinct smooth sheet cells that
tesselate. We define the vectors used to construct each cell to describe the motion of the
cell’s vertices and confirm its compatibility with the zipper-coupled tubes structure.
The seamless smooth sheet attachment consists of four distinct cells denoted Si , for
i = 1, 2, 3, 4. In Figure 4a, the lightly shaded cell is S1 and the darkly shaded cell is S4 ,
while in Figure 4c, the lightly shaded cell is half of S3 and the darkly shaded cell is S2 .
Cells S1 and S3 share the same configuration, but like the zipper-coupled tube segments,
one is rotated 180◦ about the y-axis; the same is true of S2 and S4 . The manner in which
the smooth sheet cells attach to Z in the ideal state (γ = γ0 ) is highlighted in Figure 4b.
When the mechanism is fully deployed, cells on the top sheet meet along their jagged edges,
as seen in Figure 4a. The cells on the bottom sheet, however, meet along their straight edges,
as seen in Figure 4c—this is a key feature in defining a design that will cover zipper-coupled
tubes without gaps and deploy without restricting the tubes’ motion up to the ideal state.
Regarding the single component Z0 , the cell Si attaches to Ci . However, whereas the
cells S2 and S4 form a bridge between a pair of creases in Z0 , the cells S1 and S3 form
a bridge between a crease of Z0 and a crease of Z1 (bolded in Figure 4b). Upon careful
examination of the zipper-coupled tubes, we observed that S1 can be designed as though it
bridged two creases of C4 and then moved to bridge creases of the copies of C1 in Z0 and
Z1 ; this strategy eliminates some complexity in defining S1 . In contrast, S4 is designed
directly, transversing the two creases of C4 where it attaches. Designing both cells atop C4
allows us to re-use the same vectors and enables the interested reader to easily compare the
Miura-ori inspired cells and the original smooth sheet cells with gaps [5].

3.1. Design of S1
The first smooth sheet cell, S1 , has two panels that fold toward each other as the
zipper-coupled tubes fold up. We first discuss the design of S1 when built atop C4 —the
desired relation between S1 and C4 is shown in Figure 5a. The vectors that define S1 in this
temporary configuration will be re-used to define S4 , allowing the edges of these two cells
to mesh when placed opposite each other. After construction, the smooth sheet cell S1 will
be moved to attach to C1 , as seen in Figure 5b.

283
Mathematics 2022, 10, 2643

In preparation of the design of S1 , let the displacement between G2 (γ) and E2 (γ) be
described by
q̄(γ) = d̄(γ) − b̄(γ)
and
q̄⊥ (γ) = q̄(γ) − (q̄(γ) · c̄ˆ (γ))c̄ˆ (γ)

Figure 5. (a) Placing S1 on C4 . (b) Attaching S1 to asymmetric zipper-coupled tubes.

As depicted in Figure 6a,b, q̄⊥ (γ) is the component vector of q̄(γ) orthogonal to c̄(γ),
the displacement between E1 (γ) and E2 (γ) (or equivalently, the displacement between
G1 (γ) and G2 (γ)). In the ideal state of the zipper-coupled tubes, the distance between the
ridges of Z on which the smooth sheet will be attached is

Δ = q̄⊥ (γ0 )

Figure 6. (a) Vectors in S1 . (b) Vector components of ū and v̄. (c) Relation between h and w in the
ideal state.

Note that, of necessity and by design, the top edges of both S1 and S4 are parallel
with q̄(γ0 ). To remove gaps between zipper-coupled tubes stacked laterally to Z0 and in
anticipation of the behavior depicted in Figure 4c, we extended the side edges of S1 in the
direction of c̄ by the length ε, as shown in Figures 5a and 6a, where

b sin α1 sin γ0
ε=
2|c2 (γ0 )|

Toward this end, let

ε̄ε (γ) = εc̄ˆ

284
Mathematics 2022, 10, 2643

The smooth sheet cell is determined by the vectors ū, v̄, and w̄ illustrated in Figure 6a.
As highlighted in Figure 6b, let

d sin α1 − b sin α4 + Δ
u1 = (1)
2
b sin α4 − d sin α1 + Δ
v1 = (2)
2
Note that Δ = u1 + v1 . Examining the components of ū and v̄ in the ideal state,
as shown in Figure 6a,b, we see that u2 + v2 = ||q̄(γ0 ) − q̄⊥ (γ0 )||. Since
u2 u1 u v2 u
= = 1 and = 2
u2 + v2 u1 + v1 Δ v1 u1

it follows that
u1
u2 = ||q̄(γ0 ) − q̄⊥ (γ0 )|| (3)
Δ
u2
v2 = v1 (4)
u1

Rodrigues’ rotation formula [10] rotates a vector x by angle θ in a counter-clockwise

direction about a unit vector k, and is given by:

R[k, θ ](x) = x cos θ + (k × x) sin θ + kk, x(1 − cos θ )

Let

û1 (γ) = R[−c̄ˆ (γ), λ(γ)](q̄ˆ ⊥ (γ))

û2 (γ) = −c̄ˆ (γ)

where
 
−1 q̄⊥ (γ)2 + u21 − v21
λ(γ) = cos
2q̄⊥ (γ)u1

Then we define ū(γ) and v̄(γ) as

ū(γ) = u1 û1 (γ) + u2 û2 (γ)

v̄(γ) = ū(γ) − q̄(γ)

In order to satisfy flat-foldability, the direction of the vector w̄ must be the same as the
direction of −c̄. Thus, for some positive constant w:

w̄(γ) = −wc̄ˆ (γ)

The primary concern in choosing the length w was to avoid intersections with the
zipper-coupled tubes during deployment. So that the smooth sheet cells S1 and S4 lined up
correctly, we defined w and h to be the lengths necessary so that in the ideal state, the tips
of the vectors w̄ (−wc̄ˆ ) and hĉ meet, as shown in Figure 6c. In particular, we set
$ %
hĉ(γ0 ) + F3 (γ0 ) = −wc̄ˆ (γ0 ) + E2 (γ0 ) + ε̄ε (γ0 ) − ū(γ0 )

Then, w (and h) can be obtained as follows:

$ %
[h, w] T = [proj(ĉ(γ0 )), proj(c̄ˆ (γ0 ))]−1 proj ā + c̄ + d̄ + s + ε̄ε − ū − c (γ0 ) (5)

285
Mathematics 2022, 10, 2643

where the function proj : R3 → R2 is defined such that

proj[ x, y, z] T = [ x, y] T

Remark 1. Note that the matrix that is inverted in Equation (5) will be singular only in the
symmetric case when c̄ has the same direction as −c.

We have now stipulated all three vectors—ū, v̄, and w̄—which define S1 when it is
attached to C4 . What remains is to move S1 so that it attaches to C1 . The desired placement
is depicted in Figure 5b, where S1 is positioned on top of C1 and bridges Z0 and Z1 ; note
that the zipper-coupled tubes have been rotated about the y-axis in this figure so that S1
lies parallel with the xy-plane (see [5] for further details). Let X [i ] denote the point in Zi
that is a copy of X in Z0 , for i ≥ 1. Then, the vertex F3 ∗ in S1 attaches to F3 in Z0 and the
vertex F3 ∗ [1] in S1 attaches to F3 [1] in Z1 . More particularly, the edge P0 F3∗ attaches to F0 F3
and P0 [1] F3∗ [1] attaches to F0 [1] F3 [1].
Recall that the 180◦ rotation of a vector x̄ about the y-axis is denoted x. Thus, the ver-
tices in S1 when attached to C1 as desired are:

F3 (γ) = c(γ) O0 (γ) = −ε (γ) + u(γ) F3 [1](γ) = c(γ) − b(γ) + d(γ)

P0 (γ) = −ε (γ) O3 (γ) = −ε (γ) + u(γ) − w(γ) P0 [1](γ) = −ε (γ) − b(γ) + d(γ)

3.2. Design of S3
The cell S3 is a rotated copy of S1 attached to C3 , so the vectors that define S3 are
rotated copies of the vectors that define S1 , shifted by s. Let the vertices on smooth sheet
cells attached to the rotated tube segment in Z0 —i.e., the vertices in S2 and S3 —be denoted
with primes. Note that the smooth sheet cell S3 bridges Z0 and Z−1 , so the smooth sheet
cell contains the vertices F3 and F3 [−1]. Thus, the vertices of S3 when attached to C3 are:

F3 (γ) = c̄(γ) + s(γ) O0 (γ) = s(γ) − ε̄ε (γ) + ū(γ) F3 [−1](γ) = c̄(γ) + s(γ) − b̄(γ) + d̄(γ)
P0 (γ) = s(γ) − ε̄ε (γ) O3 (γ) = s(γ) − ε̄ε (γ) + ū(γ) − w̄(γ) P0 [−1](γ) = s(γ) − ε̄ε (γ) − b̄(γ) + d̄(γ)

3.3. Design of S4
The smooth sheet cell S4 attaches on top of C4 and fits together with S1 in the ideal
state, as illustrated in Figure 7. We make the edge of S4 opposite of S1 straight in the ideal
state so that zipper-coupled tubes with smooth sheet attachments can be stacked laterally
without gaps. From another point of view, the edge is made straight in preparation of the
design of S2 , a copy of S4 attached to the bottom of Z0 —the straight edges of S2 and S3
meet in the ideal state, as illustrated in Figure 4c.
The smooth sheet cell S4 has a degree-four vertex folding pattern inspired by the
Miura-ori cell, as shown in Figure 8a. This allows the cell to close the gap on the top of
Z0 in the ideal state and fold up without intersecting the adjacent tube segment. For flat-
foldability of the cell, we require the sum of opposite angles at the interior vertex to be
180◦ (see Kawasaki-Justin theorem [11,12]). The pattern in Figure 8a is described by the
previously defined vectors ū, v̄, and w̄ and the yet-to-be-defined vectors r̄, t̄, h̄, f̄, and ḡ.
In the ideal state, r̄ and t̄ are the projections of b and d, respectively, into the xy-plane;
this is necessary to ensure flat foldability. Thus (see Figures 3 and 8a),

β 2 = ∠ P1 Q1 Q2 = ∠ F3 G1 G2 and β 3 = ∠ P1 O1 O2 = ∠ F3 E1 E2

In particular,
 
d(γ0 ) − s0 a(γ0 )
β 2 = cos−1 −c̄ˆ (γ0 ) ·
d(γ0 ) − s0 a(γ0 )
 
b(γ0 ) − s0 a(γ0 )
β 3 = cos−1 −c̄ˆ (γ0 ) ·
b(γ0 ) − s0 a(γ0 )

286
Mathematics 2022, 10, 2643

Figure 7. (a) Placing S4 on C4 . (b) Attaching S4 to asymmetric zipper-coupled tubes.

Figure 8. (a) Vectors in S4 . (b) Vector components of f̄ and ḡ.

Having derived these angles, we are now ready to define r̄ and t̄ as follows (see
Equations (1)–(4)):
 
u1
r̄(γ) = − − u2 c̄ˆ (γ) + ū(γ)
tan β 3
 
v1
t̄(γ) = − + v2 c̄ˆ (γ) + v̄(γ)
tan β 2

Observing Figure 8a, note that r and t can be expressed simply as:
u1 v1
r= t=
sin β 3 sin β 2

In the ideal state, the crease defined by h̄ must have the same direction as −c to satisfy
flat-foldability. Moreover, so that there are no gaps when S1 and S4 come together in the
ideal state, the length of h̄ should be the value h given by Equation (5), according to the
premise upon which Equation (5) was derived (see also Figure 6c). Thus,

h̄(γ0 ) = − hĉ(γ0 )

For an arbitrary parameter value γ, the unit vectors adjacent to h̄ˆ (γ) that emanate from
the degree-four vertex in the interior of S4 are r̄ˆ (γ) and ˆt̄(γ). Because opposite angles in a
degree-four vertex sum to 180◦ [11,12] and we require a rigid folding, h̄(γ) is determined
by the following system of equations:
 $ %
h̄ˆ (γ) · r̄ˆ (γ) = cos π − cos−1 c̄ˆ (γ) · ˆt̄(γ)
 $ %
h̄ˆ (γ) · ˆt̄(γ) = cos π − cos−1 c̄ˆ (γ) · r̄ˆ (γ)

h̄ˆ (γ) · h̄ˆ (γ) = 1

287
Mathematics 2022, 10, 2643

The first two equations are linear and the third is quadratic. Hence, there are precisely
two solutions: one corresponding to a valley fold assignment and one corresponding
to a mountain fold assignment. The solution corresponding to a mountain fold is the
correct solution.
Because ḡ corresponds to an edge of the panel defined by h̄ and t̄, we can define it in
terms of these vectors. We want the position of vertex Q4 in S4 to equal that of O3 in S1
when in the ideal state. This is equivalent to saying that gh h̄ˆ and − gt ˆt̄ define the same point
when the former is extended from Q1 and the latter is extended from O3 in the ideal state,
as shown in Figure 8b. The reader will recognize that this problem is formulated similarly
to that in the end of Section 3.1, where the lengths h and w were computed using a system
of equations based on two intersecting vectors. We employ the same technique, with the
following system of equations:

gh h̄ˆ (γ0 ) + Q1 (γ0 ) = − gt ˆt̄(γ0 ) + O3 (γ0 )

giving us
   !−1 $ %
[ gh , gt ] T = proj h̄ˆ (γ0 ) , proj ˆt̄(γ0 ) proj −ε + u − w − ā − c̄ − d̄ − s − ε̄ε + ū − w̄ − t̄ (γ0 )

We solve for gh and gt and use these components of projection to define ḡ, using h̄ˆ and
ˆt̄ as a basis:
ḡ(γ) = gh h̄ˆ (γ) + gt ˆt̄(γ)
We define f̄ similarly, solving for f h and f r via the same method and another system of
equations:
f h̄ˆ (γ ) + O (γ ) = − f r̄ˆ (γ ) + O [−1](γ )
h 0 1 0 r 0 3 0

This gives us
   !−1 $ %
[ f h , f r ] T = proj h̄ˆ (γ0 ) , proj r̄ˆ (γ0 ) proj −ε + u − w − ā − b̄ − c̄ − s − ε̄ε + ū − w̄ − r̄ (γ0 )

Thus, we have
f̄(γ) = f h h̄ˆ (γ) + f r r̄ˆ (γ)
Now that we have defined all the essential vectors in S4 , we are ready to attach the
degree-four vertex to C4 in the zipper-coupled tubes structure. In the ideal state, this
smooth sheet cell matches exactly with the edges of S1 , providing a smooth surface devoid
of gaps on the top of asymmetric zipper-coupled tubes (Figure 4a). In particular, the edge
O1 O2 attaches to E1 E2 and Q1 Q2 attaches to G1 G2 . Likewise, the edges O1 O4 and Q1 Q4
align perfectly with adjacent cells in the ideal state and fold up at different rates to avoid
intersections. The vertices in S4 identified with their corresponding position vectors are:

O1 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ) − ū(γ) + w̄(γ) + r̄(γ)
O2 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ)
O4 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ) − ū(γ) + w̄(γ) + r̄(γ) + f̄(γ)

P1 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ) − ū(γ) + w̄(γ)
P2 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ) − ū(γ)
P4 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ) − ū(γ) + w̄(γ) + h̄(γ)

Q1 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ) − ū(γ) + w̄(γ) + t̄(γ)
Q2 (γ) = ā(γ) + b̄(γ) + c̄(γ) + s(γ) + ε̄ε (γ)
Q4 (γ) = ā(γ) + c̄(γ) + d̄(γ) + s(γ) + ε̄ε (γ) − ū(γ) + w̄(γ) + t̄(γ) + ḡ(γ)

288
Mathematics 2022, 10, 2643

3.4. Design of S2
The smooth sheet cell S2 is a rotated copy of S4 that attaches to C2 . The vertices in this
smooth sheet cell are

O1 (γ) = a(γ) + c(γ) + d(γ) + ε (γ) − u(γ) + w(γ) + r(γ)

O2 (γ) = a(γ) + c(γ) + d(γ) + ε (γ)
O4 (γ) = a(γ) + c(γ) + d(γ) + ε (γ) − u(γ) + w(γ) + r(γ) + f(γ)

P1 (γ) = a(γ) + c(γ) + d(γ) + ε (γ) − u(γ) + w(γ)

P2 (γ) = a(γ) + c(γ) + d(γ) + ε (γ) − u(γ)
P4 (γ) = a(γ) + c(γ) + d(γ) + ε (γ) − u(γ) + w(γ) + h(γ)

Q1 ( γ ) = a( γ ) + c( γ ) + d( γ ) + ε ( γ ) − u( γ ) + w( γ ) + t( γ )
Q2 ( γ ) = a( γ ) + b( γ ) + c( γ ) + ε ( γ )
Q4 ( γ ) = a( γ ) + c( γ ) + d( γ ) + ε ( γ ) − u( γ ) + w( γ ) + t( γ ) + g( γ )

We have now completed the details for the smooth sheet attachment in the asymmetric
case; a summary of the edges and vertices in the attachment is given in Tables A1 and A2,
suppressing γ for concision. This attachment folds up flat and actuates with the zipper-
coupled tubes structure to form a smooth surface, leaving no gaps between the various
asymmetric cells we have described. The symmetric case merits more discussion, however,
because there are fewer constraints on the vectors in the Si , allowing for multiple rigidly
foldable patterns given specific design parameters.

4. Symmetric Smooth Sheet Attachment

Unlike their asymmetric counterparts, symmetric zipper-coupled tubes have no tilt
and fold parallel with the x-axis [5]. Furthermore, the center creases in S1 and S4 lie
parallel with the y-axis in the ideal state, as illustrated in Figure 9. There are multiple
valid lengths for these creases, therefore, in a design that covers all the gaps in the surface
while maintaining rigid foldability without prohibitive intersections. In this section we
comment on the diversity in symmetric, Miura-ori inspired smooth sheet cell construction
and recommend values for certain lengths.

Figure 9. S1 and S4 in a symmetric smooth sheet attachment.

The design parameters for the smooth sheet cells Si are uniquely determined for all
cases where α1 + α2 < π. When α1 + α2 = π, however, there is no longer a unique solution
to Equation (5). In particular, c̄(γ0 ) has the same direction as −c(γ0 ) in the symmetric case,
so the matrix inverted in Equation (5) is singular and the values h and w are not uniquely
defined. Similarly, β 2 and β 3 are no longer constrained, and we may define these features
of the design problem advantageously by choosing a solution that minimizes the amount
by which the smooth sheet attachment protrudes from the structure when folded.

289
Mathematics 2022, 10, 2643

In the design of S1 and S3 , let w∗ replace the value of w. Likewise, in the design of
S4 and S2 , let w∗∗ replace the value of w. As highlighted in Figure 9, for the symmetric
case we no longer require that w∗ = w∗∗ . In selecting a value for w∗ , we set it as large as
possible to maximize the surface area of S1 , thus minimizing the amount by which the
edges of S4 can protrude from the zipper-coupled tubes structure. Applying the analysis
given in [5] (see Section 7.1.3), the largest value for w∗ can be shown to be

d sin α1 + b sin α4 − Δ
w∗ = c − −ε
2 tan (α2 − α1 )

The only requirements for h, w∗∗ , β 2 , and β 3 in the symmetric case are

h + w∗∗ = c − ε and β2 = β3

Adjustments to the values h, w∗∗ , β 2 , and β 3 can also assist in minimizing the protru-
sion of S4 from the zipper-coupled tubes. Optimal values can be determined by numerical
methods according to the specific design application. However, care should be taken in
making these adjustments to avoid intersections with the structure underneath.
On a final note, although f̄ and ḡ are determined after defining the previous quantities,
a convenient simplification in their definition: because S4 is symmetric, the vectors f̄ and ḡ
are parallel to h̄ and have equal lengths. Moreover, in the ideal state,

O1 O4 = O2 O4 − O1 O2

Therefore,
 
u1
gt = f r = 0 and gh = f h = (2c − w∗ ) − w∗∗ +
tan β 3

5. Conclusions
We have successfully defined a smooth sheet attachment that folds up with the zipper-
coupled tubes and unfolds to the ideal state without inhibiting their motion to form a flat
surface without any gaps (Figure 10). This pattern is defined for both the asymmetric and
symmetric cases, and we provide access to code which the reader may use to visualize
the origami structures described and print out the corresponding fold patterns: https:
//github.com/dylanwebbc/azct (accessed on 23 July 2022).

Figure 10. Model of asymmetric zipper-coupled tubes with a Miura-ori inspired smooth sheet
attachment at different stages of unfolding; the nature of the deployment of the degree-four vertex
cells is clearly observable on the bottom of the device. Parameters: α1 = 13 π, α2 = 59 π, a = b = c.
Dimensions (inches): 4.9 by 3.3 when folded, 7.4 by 3.5 by 1.5 when unfolded.

290
Mathematics 2022, 10, 2643

Note that the smooth sheet cells protrude from the zipper-coupled tubes structure
when folded. When gaps are tolerable and the folded state must be minimized for trans-
portation, constraining the cells to fold up within the zipper-coupled tubes while max-
imizing surface area in the ideal state results in the smooth sheet attachment described
previously [5]. Thus, the smooth sheet design can be tailored to the situation, much like
zipper-coupled tubes themselves. To inform future applications of these structures, we sug-
gest dynamic and quasi-static analyses. Constructing a device for architectural applications
will likely require the use of thick origami and compliant hinges, and remote self-actuation
via magnetism or heat could be useful in space or medical applications [13–15].
Miura-ori-inspired smooth sheet attachments enhance the utility of zipper-coupled
tubes in various situations. The tubes are useful in architecture because they pack tight
and deploy to a rigid state [4]; our gapless smooth sheet attachments improve existing
designs by increasing drivability and walkability. If a local bridge collapses, for example,
a prefabricated bridge based on zipper-coupled tubes with smooth sheet attachments
can easily be transported on a single vehicle and swiftly deployed on-site to provide
smooth, emergency transit. Space structures are another popular application of origami-
inspired mechanisms—the Miura-ori pattern that the smooth sheet is based on is common
in deployable space array design. Accommodating for material thickness, however, makes
Miura-ori sheets challenging to deploy [16]. In contrast, a thin solar array constructed
from Miura-ori inspired smooth sheet cells can deploy rigidly because it is supported by
zipper-coupled tubes.
The design of smooth sheet attachments without gaps is key to the development of
more versatile zipper-coupled tubes. We have communicated a clear design method for the
origami-based structure, examining the mathematics of its motion in detail. By elucidating
the possibility for further enhancements on the zipper-coupled tubes structure, we hope to
spur many novel and exciting applications beyond those mentioned.

Author Contributions: Conceptualization, D.C.W.; Methodology, D.M.H. and L.L.H.; Software,

D.C.W.; Supervision, D.M.H. and L.L.H.; Validation, D.C.W.; Visualization, D.C.W. and E.R.; Writing—
original draft, D.C.W. and D.M.H.; Writing—review & editing, D.C.W., E.R. and D.M.H. All authors
have read and agreed to the published version of the manuscript.
Funding: This research was funded by a grant from the College of Physical and Mathematical
Sciences at Brigham Young University.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A

Table A1. Defining the edges in Miura-ori based, smooth sheet attachments for a pair of asymmetric
zipper-coupled tubes.

S1 S2 S3 S4

O0 P0 O0 P0 [1] O1 P1 O2 P2 O4 P4 O0 P0 O0 P0 [−1] O1 P1 O2 P2 O4 P4
O3 F3 O3 F3 [1] P1 Q1 P2 Q2 P4 Q4 O3 F3 O3 F3 [−1] P1 Q1 P2 Q2 P4 Q4
O0 O3 P0 F3 O1 O2 O1 O4 P1 P2 O0 O3 P0 F3 O1 O2 O1 O4 P1 P2
P0 [1] F3 [1] P1 P4 Q1 Q2 Q1 Q4 P0 [−1] F3 [−1] P1 P4 Q1 Q2 Q1 Q4

291
Mathematics 2022, 10, 2643

Table A2. Defining the vertices in Miura-ori inspired smooth sheet attachments for a pair of asym-
metric zipper-coupled tubes.

Vertex Position Vertex Position

O0 −ε + u O0 s − ε̄ε + ū
O1 a+c+d+ε −u+w+r O1 ā + c̄ + d̄ + s + ε̄ε − ū + w̄ + r̄
O2 a+c+d+ε O2 ā + c̄ + d̄ + s + ε̄ε
O3 −ε + u − w O3 s − ε̄ε + ū − w̄
O4 a+c+d+ε −u+w+r+f O4 ā + c̄ + d̄ + s + ε̄ε − ū + w̄ + r̄ + f̄
P0 −ε P0 s − ε̄ε
P1 a+c+d+ε−u+w P1 ā + c̄ + d̄ + s + ε̄ε − ū + w̄
P2 a+c+d+ε −u P2 ā + c̄ + d̄ + s + ε̄ε − ū
P4 a+c+d+ε −u+w+h P4 ā + c̄ + d̄ + s + ε̄ε − ū + w̄ + h̄
Q1 a+c+d+ε −u+w+t Q1 ā + c̄ + d̄ + s + ε̄ε − ū + w̄ + t̄
Q2 a+b+c+ε Q2 ā + b̄ + c̄ + s + ε̄ε
Q4 a+c+d+ε −u+w+t+g Q4 ā + c̄ + d̄ + s + ε̄ε − ū + w̄ + t̄ + ḡ
F3 c F3 c̄ + s
F3 [1] c−b+d F3 [−1] c̄ + s − b̄ + d̄
P0 [1] −ε − b + d P0 [−1] s − ε̄ε − b̄ + d̄

References
1. Mena, L.; Muñoz, J.; Monje, C.A.; Balaguer, C. Modular and Self-Scalable Origami Robot: A First Approach. Mathematics 2021, 9,
1324. [CrossRef]
2. Rubio, A.J.; Kaddour, A.S.; Georgakopoulos, S.V.; Ynchausti, C.; Magleby, S.; Howell, L.L. A Deployable Hexagonal Reflectarray
Antenna for Space Applications. In Proceedings of the 2021 United States National Committee of URSI National Radio Science
Meeting (USNC-URSI NRSM), Boulder, CO, USA, 4–9 January 2021; pp. 136–137. [CrossRef]
3. Zhang, F.; Li, S.; Shen, Z.; Cheng, X.; Xue, Z.; Zhang, H.; Song, H.; Bai, K.; Yan, D.; Wang, H.; et al. Rapidly deployable and
morphable 3D mesostructures with applications in multimodal biomedical devices. Proc. Natl. Acad. Sci. USA 2021, 118,
e2026414118. [CrossRef] [PubMed]
4. Filipov, E.T.; Tachi, T.; Paulino, G.H. Origami tubes assembled into stiff, yet reconfigurable structures and metamaterials. Proc.
Natl. Acad. Sci. USA 2015, 112, e1509465112. [CrossRef] [PubMed]
5. Webb, D.C.; Elissa, R.; Halverson, D.M.; Howell, L.L. Deployable Space-Filling Mechanisms: Asymmetric Zipper-coupled Tubes
and Smooth Sheet Attachments. In Proceedings of the ASME 2022 International Design Engineering Technical Conferences and
Computers and Information in Engineering Conference, St. Louis, MI, USA, 14–17 August 2022; Forthcoming.
6. Miura, K. The Science of Miura-Ori: A Review. In Origami 4; Lang, R.J., Ed.; A K Peters/CRC Press: Boca Raton, FL, USA, 2009;
Chapter 4, pp. 87–99.
7. Miura, K. Method of Packaging and Deployment of Large Membranes in Space. Inst. Space Astronaut. Sci. Rep. 1985, 618, 1–9.
8. Miura, K.; Natori, M. 2-D Array Experiment on Board a Space Flyer Unit. Space Sol. Power Rev. 1985, 5, 345–356.
9. Zimmermann, L.; Stanković, T. Rigid and Flat Foldability of a Degree-Four Vertex in Origami. J. Mech. Robot. 2020, 12, 011004.
[CrossRef]
10. Rodrigues, O. Des lois géométriques qui régissent les déplacements d’un système solide dans l’espace, et de la variation des
coordonnées provenant de ces déplacements considérés indépendamment des causes qui peuvent les produire. J. Mathématiques
Pures Appliquées 1840, 5, 380–440. Available online: http://xxx.lanl.gov/abs/http://sites.mathdoc.fr/JMPA/PDF/JMPA_1840_1
_5_A39_0.pdf (accessed on 1 January 2022).
11. Kawasaki, T. On the Relation between Mountain-Creases and Valley-Creases of a Flat Origami. In Proceedings of the 1st International
Meeting on Origami Science and Technology; Huzita, H., Ed.; Universita di Padova: Padova, Italy, 1989; pp. 229–237.
12. Justin, J. Mathematics of origami, part 9. Br. Origami 1986, 118, 28–30.
13. Yellowhorse, A.; Tolman, K.; Howell, L.L. Optimization of Origami-Based Tubes for Lightweight Deployable Structures. In
Proceedings of the ASME 2017 International Design Engineering Technical Conferences and Computers and Information in
Engineering Conference, Cleveland, OH, USA, 6–9 August 2017. [CrossRef]

292
Mathematics 2022, 10, 2643

14. Kim, Y.; Yuk, H.; Zhao, R.; Chester, S.A.; Zhao, X. Printing ferromagnetic domains for untethered fast-transforming soft materials.
Nature 2018, 558, 274–279. [CrossRef] [PubMed]
15. Kuribayashi, K.; Tsuchiya, K.; You, Z.; Tomus, D.; Umemoto, M.; Ito, T.; Sasaki, M. Self-deployable origami stent grafts as a
biomedical application of Ni-rich TiNi shape memory alloy foil. Mater. Sci. Eng. A 2006, 419, 131–137. [CrossRef]
16. Bolanos, D.; Ynchausti, C.; Brown, N.; Pruett, H.; Hunter, J.; Clark, B.; Bateman, T.; Howell, L.L.; Magleby, S.P. Considering
thickness-accommodation, nesting, grounding and deployment in design of Miura-ori based space arrays. Mech. Mach. Theory
2022, 174, 104904. [CrossRef]

293
 mathematics

Article
Jump and Initial-Sensitive Excessive Motion of a Class of
Relative Rotation Systems and Their Control via Delayed
Feedback
Ziyin Cui and Huilin Shang *

School of Mechanical Engineering, Shanghai Institute of Technology, Shanghai 201418, China;

206091149@mail.sit.edu.cn
* Correspondence: shanghuilin@sit.edu.cn

Abstract: Jump and excessive motion are undesirable phenomena in relative rotation systems, causing
a loss of global integrity and reliability of the systems. In this work, a typical relative rotation system
is considered in which jump, excessive motion, and their suppression via delayed feedback are
investigated. The Method of Multiple Scales and the Melnikov method are applied to analyze critical
conditions for bi-stability and initial-sensitive excessive motion, respectively. By introducing the
fractal of basins of attraction and the erosion of the safe basin to depict jump and initial-sensitive
excessive motion, respectively, the point mapping approach is used to present numerical simulations
which are in agreement with the theoretical prediction, showing the validity of the analysis. It is found
that jump between bistable attractors can be due to saddle–node bifurcation, while initial-sensitive
excessive motion can be due to heteroclinic bifurcation. Under a positive coefficient of the gain,
the types of delayed feedback can both be effective in reducing jump and initial-sensitive excessive
motion. The results may provide some reference for the performance improvement of rotors and
Citation: Cui, Z.; Shang, H. Jump
and Initial-Sensitive Excessive
main bearings.
Motion of a Class of Relative Rotation
Systems and Their Control via Keywords: relative rotation; jump; safe basin; fractal; heteroclinic bifurcation; delayed feedback
Delayed Feedback. Mathematics 2022,
10, 2676. https://doi.org/10.3390/ MSC: 58Z05; 37G35; 37G15
math10152676

Academic Editors: Higinio Rubio

Alonso, Alejandro Bustos,
1. Introduction
Jesus Meneses Alonso and
Enrique Soriano-Heras Rotating machinery, widely used in engineering fields such as transmission systems
of vehicles [1] and main bearings of aeroengines [2], plays an important role in the theory
Received: 28 June 2022 of nonlinear dynamics. Due to the nonlinearities of its dynamical systems, it usually
Accepted: 27 July 2022
undergoes complex dynamics, which have a negative impact on its operational stability
Published: 29 July 2022
and reliability [3,4]. Among these, relative rotation systems have been given much attention,
Publisher’s Note: MDPI stays neutral as nonlinear vibration is undesirable. For instance, main shafts can be fatigue damaged
with regard to jurisdictional claims in by a jump [5,6] among multiple attractors, or by a chaotic response [7], and can even be
published maps and institutional affil- fractured due to excessive motion [8]. Thus, comprehensive study of the global dynamics
iations. of relative rotation systems has become a practical issue.
So far, many researchers have studied the responses of relative rotation systems. Wang et al. [9]
obtained the approximate periodic solution of a class of nonlinear systems with relative rotation
by applying the Method of Multiple Scales and proved the uniqueness of the periodic response.
Copyright: © 2022 by the authors.
Xiao et al. [10] verified the existence and uniqueness of periodic response in another type of rel-
Licensee MDPI, Basel, Switzerland.
ative rotation system with a time-varying stiffness. Li et al. [11] proposed a relative rotation
This article is an open access article
system with nonlinear elastic force and nonlinear generalized damping force, and discussed
distributed under the terms and
its periodic solution problem theoretically. Shi et al. [12] numerically studied the chaotic
conditions of the Creative Commons
behaviors of a relative rotation system under parametric excitation by Poincaré map and
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
maximal Lyapunov exponent. By simulating relative torsional vibrations of an imbalanced
4.0/).
shaft under a limited power supply, Verichev [13] found that the interaction of shaft and

Mathematics 2022, 10, 2676. https://doi.org/10.3390/math10152676 295 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 2676

power supply may result in strange attractors such as classical Lorenz and Feigenbaum
attractors. In a nonlinear relative rotation system with a triple-well Mathieu-Duffing oscilla-
tor, Liu et al. [14] obtained the threshold of chaos regarding Smale horseshoe commutation
and exhibited the erosion process of safe basins. A class of nonlinear relative rotational
systems containing two rotors was also built whose chaotic response was presented [15].
By introducing the erosion of the safe basin to describe the safe performance of the spur
gear pair, Zhu et al. [16] classified the multiple meshing states and presented the transition
process from safe to unsafe. To control the complex dynamics of the main transmission
system of a scraper conveyor, Ju et al. [17] analyzed its local bifurcations and proposed a
nonlinear state feedback controller whose effect was studied numerically. Considering the
effectiveness of time-delay feedback on controlling fractal erosion of the safe basin and
chaos in nonlinear dynamical systems [18,19], Zhao et al. [20] applied delayed displacement
feedback in a relative torsional vibration system for reducing its response amplitude. On
this basis, Shang et al. [21] discussed the effect of delayed position feedback on controlling
the erosion of the safe basin and chaos. Although there has been meaningful research on
the nonlinear vibration characteristics of the relative rotation system, the study of excessive
motion and jump among multiple attractors is mainly carried out by numerical simulation
and the mechanism of these complex dynamical behaviors and their control is still not
yet clear.
To this end, we select a class of nonlinear relative rotation systems composed of two
rotors and study the mechanisms behind its jump and initial-sensitive excessive motion,
as well as the effect of delayed feedback on suppressing these phenomena. The paper is
arranged as follows. In Section 2, the dynamical model of a relative rotation system is
constructed and made dimensionless. In Section 3, the mechanism of jump and excessive
motion is analyzed. In Section 4, two control strategies, namely, delay position feedback
and delay velocity feedback, are applied to the original system, respectively, whose control
mechanism is then discussed. Section 5 contains the discussion.

2. Dynamical Model and Unperturbed Dynamics

We consider a typical torsional vibrating system whose simplified diagram is presented
in Figure 1. θ1 and θ2 are the absolute rotating angles of two rotors, respectively, and J1 and
J2 represent their polar moments of inertia. Considering nonlinear torsional stiffness and
nonlinear damping, the governing equation of this vibrating system can be derived by the
momentum theorem as
.. . .
J1 θ 1 + c12 (θ 1 − θ 2 ) + K1 (θ1 − θ2 ) + K3 (θ1 − θ2 )3 + f 12 = Te1 ,
.. . . (1)
J2 θ 2 + c12 (θ 2 − θ 1 ) + K1 (θ2 − θ1 ) + K3 (θ2 − θ1 )3 − f 12 = Te2 ,

where K1 is the coefficient of linear stiffness between the two rotors; K3 is the coefficient of
nonlinear stiffness satisfying K3 < 0; c12 is the coefficient of linear damping; Te1 and Te2 are
external rotational torques loaded on the two rotors, respectively; f 12 is a class of nonlinear
stick-slip frictions [20,21] given by
. . . . 3
f 12 = c0 + c1 (θ 1 − θ 2 ) + c2 (θ 1 − θ 2 ) , (2)

in which c0 , c1 and c2 are the constant of static friction, the coefficient of linear damping
and the coefficient of cubic nonlinear damping, respectively. By denoting x = θ1 − θ2 in
Equation (1), the relative rotation system can be obtained as below:

.. c12 J2 + c12 J1 . K ( J + J2 ) K3 ( J1 + J2 ) 3 J1 + J2 . .3 J T − J1 Te2 − ( J1 + J2 )c0

x+ x+ 1 1 x+ x + (c1 x + c2 x ) = 2 e1 . (3)
J1 J2 J1 J2 J1 J2 J1 J2 J1 J2

296
Mathematics 2022, 10, 2676

Figure 1. Simplified diagram of a two-rotor torsional vibrating system.

The torques Te1 and Te2 are usually harmonic excitations, thus can be represented as

J2 Te1 − J1 Te2 − ( J1 + J2 )c0

= F cos(Ωt) (4)
J1 J2

where F and Ω are the amplitude and the frequency of the excitation, respectively. Intro-
ducing the following variables,

K1 ( J1 + J2 ) K3 (c + c1 )( J1 + J2 ) c2 ( J1 + J2 )ω0 Ω F
ω0 2 = , b = − , μ = 12 ,g = , T = ω0 t, ω = ,f = , (5)
J1 J2 K1 J1 J2 ω0 J1 J2 ω0 ω0 2
the dimensionless form of the relative rotation system (3) is obtained as

dx dy
y= , = −μy − x + bx3 − gy3 + f cos(ωT ). (6)
dT dT
in which the parameters μ, b, g, and ω are positive, and the position x ( T ) and the velocity
y( T ) represent relative rotational angular and relative angular velocity at the moment T,
respectively.
Since the parameters c12 , c0 , c1 , and c2 in the original system (1) are very small, the
relevant terms concerned, μ, g and f in Equation (6), can be considered as perturbed. The
unperturbed system can be written as

dx dy
= y, = − x + bx3 . (7)
dT dT
There will be three equilibria
√
in the unperturbed
√
system (7), namely, the center S1 (0,0)
and two saddle points S2 ( b
b ,0) and S3 (− b
b ,0). The Hamiltonian of Equation (7) is

1 2 1 2 1 4
H ( x, y) = y + x − bx . (8)
2 2 4
Accordingly, the heteroclinic orbits surrounding the center S1 (0,0) can be given by
√ √ √ √
b 2 2b 2
x± (t) = ± tanh( t ), y ± ( t ) = ± sech2 ( t ). (9)
b 2 2b 2
Fixing b = 0.3, the unperturbed orbits are depicted in Figure 2.

Figure 2. Orbits of the unperturbed system (7) for b = 0.3.

297
Mathematics 2022, 10, 2676

3. Complex Dynamics
3.1. Multistability and Jump
The periodic solution near the equilibrium S1 (0,0) is considered. The Method of
Multiple Scales (MMS) [22] is employed to obtain the approximate theoretical solution of
Equation (3). Supposing a small parameter ε (0 < ε  1) to rescale the nondimensional
parameters as follows:
μ, b = ε+
μ = ε+ g, f = ε f+,
b, g = ε+ (10)
Equation (6) becomes

d2 x +dx + 3
μ dx 3
2
+ x = ε(− + bx − g+( ) + f+cos(ωT )). (11)
dT dT dT
Further rescaling parameters and operators in Equation (11) yields
n
∂ d
ω = 1 + εσ, Ti = εi T, x ( T ) = x0 ( T0 , T1 , · · · ) + εx1 ( T0 , T1 , · · · ) + · · · , Di = , = ∑ εi Di (i = 0, 1, 2 · · ·), (12)
∂Ti dT i =0

where σ is the detuning parameter, and separating the coefficients of ε0 and ε1 of the
system (11) one finds
ε0 : D0 2 x0 + x0 = 0, (13)
and

+D0 x0 + +
ε1 : D0 2 x1 + x1 = −2D1 D0 x0 − μ bx0 3 − g+( D0 x0 )3 + f+cos( T0 + σT1 ). (14)

The solution of Equation (13) is

x0 ( T0 , T1 ) = A( T1 )eiT0 + cc (15)

where cc denotes the complex conjugate of the preceding terms. Substituting Equation (15)
into Equation (14) and separating the secular terms yields

+ A 3( g+ + i+
μ b) A2 A i f+eiσT1
D1 A + + + = 0. (16)
2 2 4
In addition, by defining

a( T1 ) iβ(T1 )
A( T1 ) = e , ψ = σT1 − β, (17)
2
separating the real and imaginary parts of Equation (16), and returning the parameters to the
nondimensional parameters of the system (3), the modulation equation can be written as

da μa 3ga3 f sin ψ dψ 3ba3 f cos ψ

=− − + , a = a ( ω − 1) + + . (18)
dT 2 8 2 dT 8 2
The steady-state form of Equation (18) is obtained by setting D1 a = D1 ψ = 0 and
returning the parameters of (18) to the nondimensional parameters as:

3ga3 3ba3
μa + = f sin ψ, −2(ω − 1) a − = f cos ψ. (19)
4 4
Accordingly, the approximate periodic solution can be expressed as x0 = a cos(ωT − ψ). By
eliminating ψ from the relationships in Equation (19), its amplitude a can be solved from
the following equation
2 2
3ga2 3
(μ + ) a2 + (2ω − 2 + ba2 ) a2 = f 2 . (20)
4 4

298
Mathematics 2022, 10, 2676

The corresponding characteristic equation is

μ 9ga2 μ 3ga2 9 3
(λ + + )(λ + + ) + (ω − 1 + ba2 )(ω − 1 + ba2 ) = 0, (21)
2 8 2 8 8 8
meaning that for

27( g2 + b2 ) 4
a + 3( gμ + b(2ω − 2)) a2 + 4(ω − 1)2 + μ2 > 0, (22)
16
the periodic solution is unstable and becomes a saddle. Namely, when the inequation (22)
becomes an equation, saddle-node bifurcation will occur in the system (6). For

μ = 0.03, b = 0.3, g = 0.03, (23)

the response curve of Equation (20) is plotted in Figure 3 where the coexistence of two
stable periodic branches in the same range of the excitation parameters implies bi-stability.
For example, for f = 0.13 and ω within the range (0.65, 0.86), bistable periodic attractors
coexist (see Figure 3a). Fixing ω = 0.80, the increase of the amplitude f can also lead to
saddle-node bifurcation, and thus bi-stability. The numerical results totally match the
theoretical prediction.

(a) (b)

Figure 3. Response curves of system (2). (a) Amplitude a vs. ω when f = 0.13; (b) Amplitude a vs. f
when ω = 0.80.

3.2. Initial-Sensitive Excessive Motion

Excessive motion of the system (6) can be understood as an escape or unbounded
solution [21], for instance, a relative rotating angle in system (6) more than the material of
the structure can bear. This may be induced by the variation of the system parameters or the
initial conditions of the system. In the latter case, excessive motion coexists with bounded,
meaning that a tiny disturbance of initial conditions leads to a change of dynamical be-
havior, such as from bounded vibration to unbounded. This is a typical initial-sensitive
phenomenon, initial-sensitive excessive motion, indicating the loss of global integrity of
dynamic systems [3,8]. Since it is often due to global bifurcation [15,21], we apply the
Melnikov method [23] to detect the conditions for it. Substituting the heteroclinic orbits (9)
into the Melnikov function yields
- +∞ √ √ # √
2 2μ 8 2g
M± (t0 ) = −∞ (−μy( T ) − gy3 ( T ) + f cos(ω ( T + t0 )))y( T )dT = − 3b − 35b2 + 2b πω f csch( 2πω2 ) cos( ωt0 ). (24)

299
Mathematics 2022, 10, 2676

For " √ √ √
2 2πω 2 2μ 8 2g
πω f csch( )≥ + , (25)
b 2 3b 35b2
Namely,
2(35μb+12g)sinh( πω
√ )
2
f ≥ f0 = 3 , (26)
105πωb 2
there will be a value of t0 satisfying M± (t0 ) = 0 and M±  (t0 ) = 0, meaning that the roots
of Equation M± (t0 ) = 0 are simple, enabling the existence of the transverse heteroclinic
orbits, and the system (6) may undergo initial-sensitive motion. According to Equation (26),
f 0 is the amplitude threshold of heteroclinic bifurcation. For parameter values given by
Equation (23), the critical value for heteroclinic bifurcation is f 0 = 0.09. In contrast, it follows
from Inequation (25) that the threshold of the excitation frequency ω for heteroclinic
bifurcation cannot be expressed as an explicit or monotonic increasing function of other
parameters. For example, for f = 0.13 and the values of parameters in Equation (3), it can
be calculated from Equation (25) that for ω < 1.22, heteroclinic bifurcation will occur in the
system (6).
In order to verify the criterion obtained in this subsection, numerical simulations are
carried out by fixing parameter values as in Equation (23). The 4th Runge-Kutta approach
and the point-mapping method [24] are employed to describe the phenomenon’s initial-
sensitive excessive motion. First, some terms such as basin of attraction and safe basin
are introduced briefly. A basin of attraction is defined as the set of initial conditions that
can lead to the same attractor [25]. If the boundary of basin of attraction of an attractor is
fractal and mixed with another, jump among multiple attractors may occur [25]. Safe basin
is defined as the union of basins of attraction for all bounded attractors [26]. Fractal of the
safe basin of the system (6) induces the occurrence of initial-sensitive excessive motion. In
this paper, the basin of attraction is drawn in the sufficiently large space region defined as
−3.0 ≤ x(0) ≤ 3.0, −3.0 ≤ y(0) ≤ 3.0 by generating a 600 × 600 array of starting conditions,
for each of these starting points. The escaping set for infinite time is approximated with
good accuracy by a study with 1000 excited circles. The time step is taken as 0.01. The
white region represents the numerical approximation to the basins of attraction of excessive
motion. The red and blue regions are the basins of attraction for attractors from lower and
higher stable periodic brunches of Figure 3, respectively. Thus, the union of the red region
and the blue is the so-called safe basin.
The evolution of basins of attraction with the frequency ω for f = 0.13 and with the
amplitude f for ω = 0.80 can be observed in Figures 4 and 5. In Figure 4, each the safe
basins, i.e., the union of red region and blue, is fractal-eroded when ω ranges from 0.62
to 0.87 satisfying ω < 1.22 i.e., the condition of heteroclinic bifurcation, which shows the
occurrence of initial-sensitive excessive motion, in agreement with the analytical prediction.
In Figure 4a, there is only a red region, meaning that there is only one periodic attractor
from the lower stable branch. When ω = 0.69 (see Figure 4b), the red region and the blue
one coexists, showing bistability. Specifically, on the left side of the origin, the red region
is fractal and mixed with the white region and the blue, indicating that both jump and
initial-sensitive excessive motion may occur there. As ω increases to 0.81 (see Figure 4c),
most areas of the red region will be eroded by the blue one, showing that most of the safe
initial conditions lead to the higher-amplitude attractor. For ω = 0.87 (see Figure 4d), the
red region disappears, and there is only the blue, which means that the response shifts from
the lower-amplitude branch to the higher-amplitude, in agreement with the theoretical
results of Figure 3a.

300
Mathematics 2022, 10, 2676

(a) (b) (c) (d)

Figure 4. Evolution of basins of attraction with the increase of ω for f = 0.13. (a) ω = 0.62; (b) ω = 0.69;
(c) ω = 0.81; (d) ω = 0.87.

(a) (b) (c) (d)

Figure 5. Evolution of basins of attraction with the increase of f for ω = 0.80. (a) f = 0.03; (b) f = 0.07;
(c) f = 0.12; (d) f = 0.20.

In Figure 5, the safe basin (the union of red and blue regions) is steadily eroded by the
white region with the increase of the amplitude f, implying that initial-sensitive excessive
motion can be induced by the increase of f. In Figure 5a,b, the boundary of safe basins is
smooth. When f is more than 0.09, i.e., the critical threshold solved from Equation (26),
the boundary of safe basin will become fractal (see Figure 5c,d). In Figure 5a–c, the red
region and the blue one are mixed, indicating bi-stability as well as jump. Note that for
f = 0.12, the red and blue regions are evenly mixed in the neighborhood of the origin (see
Figure 5c). For f = 0.20, only the basin of attraction of the attractors in the higher stable
branch of Figure 3b exists, which is seriously eroded by the white region, showing that the
excessive motion is easy to trigger.

4. Time-Delay Feedback to Control Complex Dynamics

In this section, two types of linear delay feedback controllers, i.e., delayed position
feedback and delayed velocity feedback, are applied to the system (1), and their effect
on suppressing the phenomena jump and initial-sensitive excessive motion is discussed.
The corresponding control diagram is shown in Figure 6 where Q is the coefficient of
. . . .
the gain, and τ time delay. P(θ1 , θ2 , θ 1 , θ 2 , θ1τ , θ2τ , θ 1τ , θ 2τ ) represents delayed relative-
angle feedback (θ1 (t − τ ) − θ2 (t − τ )) − (θ1 (t) − θ2 (t)) or delayed relative-angular-velocity
. . . .
feedback (θ 1 (t − τ ) − θ 2 (t − τ )) − (θ 1 (t) − θ 2 (t)), which is an active control transferring
the motion state/velocity of the dynamical system [27]. The corresponding governing
equation is
.. . . . . . .
J1 θ 1 + c12 (θ 1 − θ 2 ) + K1 (θ1 − θ2 ) + K3 (θ1 − θ2 )3 + f 12 − QP(θ1 , θ2 , θ 1 , θ 2 , θ1τ , θ2τ , θ 1τ , θ 2τ ) = Te1 ,
.. . . . . . . (27)
J2 θ 2 + c12 (θ 1 − θ 2 ) + K1 (θ1 − θ2 ) + K3 (θ1 − θ2 )3 + f 12 + QP(θ1 , θ2 , θ 1 , θ 2 , θ1τ , θ2τ , θ 1τ , θ 2τ ) = Te2 .

301
Mathematics 2022, 10, 2676

Figure 6. Diagram of the relative rotating system under delayed feedback control.

Letting
J1 + J2 J + J2
Gp = Q, Gv = 1 Qω0 , τ+ = ω0 τ, (28)
J1 J2 J1 J2
and based on Equations (2)–(6), we obtain the relative rotation system under delayed
position feedback control, i.e.,

dx dy
= y, = −μy − x + bx3 − gy3 + f cos(ωT ) + G p ( x ( T − τ+) − x ( T )), (29)
dT dT
and the following delayed-velocity-feedback controlled system

dx dy
y= , = −μy − x + bx3 − gy3 + f cos(ωT ) + Gv (y( T − τ+) − y). (30)
dT dT
In the delayed-feedback-controlled systems (29) and (30), G p , Gv and τ+ are indepen-
dent parameters. For G = 0 or τ+ = 0, the feedback terms in Equations (29) and (30)
become 0, and the delayed systems (29) and (30) become the uncontrolled non-dimensional
system (6). In this paper, considering the engineering application, we do not consider the
periodic characteristics of τ+ but restrict that 0 ≤ τ+ ≤ 2π. Since there is no signal returned
to the controlled systems (29) and (30) before T = 0, it is supposed that the initial states
of the delayed system when −τ+ ≤ T < 0 satisfy x ( T ) = y( T ) = 0. Then, safe basins of
the delayed-feedback controlled systems can also be projected into the initial-state plane
x (0) − y(0), similar to the uncontrolled system (6).

4.1. Delayed Position Feedback Control

4.1.1. Primary Resonant Response and Stability of Solutions
As discussed in Section 3.1, the Method of Multiple Scales is applied to obtain the
periodic response of the delayed system. Rescaling the gain as G p = ε G +p , and following
the similar procedure as in Section 3.1, one can finally obtain the slow flow equation:

da μa 3ga3 f sin ψ aG p sin τ+ dψ 3ba3 f cos ψ aG p (1 − cos τ+)

=− − + − , a = ( ω − 1) a + + − . (31)
dT 2 8 2 2 dT 8 2 2
One can get the amplitude and the phases of the response by equating the right sides
of Equation (31) to zero, i.e.,

3ga3 3ba3
μa + + aG p sin τ+ = f sin ψ, −2(ω − 1) a − + aG p (1 − cos τ+) = f cos ψ. (32)
4 4
Accordingly, the approximate periodic solution can be expressed as x0 = a cos(ωT − ψ).
Its amplitude a can be solved from the following equation
2 2
3ga2 3
(μ + G p sin τ+ + ) a2 + (2ω − 2 + ba2 + G p cos τ+ − G p ) a2 = f 2 . (33)
4 4

302
Mathematics 2022, 10, 2676

The stability of the solutions, thus obtained, can be ascertained by computing the
eigenvalues from the corresponding characteristic equation:
3ga2
λ2 + λ(μ + G p sin τ+ + 2 )
(μ+ G p sin τ+)2 G p cos τ+− G p 2
(34)
27( g2 +b2 ) a4 3a2
+ 64 + 4 + (ω −1+ 2 ) + 4 ( gG p sin τ
+ + bG p cos τ+ + b(2ω − 2 − G p ) + gμ) = 0.

Assuming λ = iv in Equation (34) and separating the real and imaginary parts,
we have
3ga2
v(μ + G p sin τ+ + )=0 (35)
2
and

27( g2 +b2 ) a4 (μ+ G p sin τ+)2 G p cos τ+− G p 2 3a2

− v2 + 64 + 4 + (ω − 1 + 2 ) + 4 ( gG p sin τ
+ + bG p cos τ+ + b(2ω − 2 − G p ) + gμ) = 0. (36)

As we know, when the delay is small, the delayed feedback can be expanded in a
Taylor series so that the controlled system (29) will become

dx dy
= y, = −(μ + G p τ+)y − x + bx3 − gy3 + f cos(ωT ). (37)
dT dT
For G p > 0, it is much like increasing the damping of the system (6), thus reducing the
3ga2
periodic vibration [20,27]. Letting G p > 0, one has μ + G p sin τ+ + 2 > 0 for 0 ≤ τ+ ≤ π.
Thus, v = 0 in Equation (35); if

27( g2 +b2 ) a4 3a2 G p cos τ+− G p 2 (μ+ G p sin τ+)2

64 + 4 ( gG p sin τ
+ + bG p cos τ+ + b(2ω − 2 − G p ) + gμ) + (ω − 1 + 2 ) + 4 = 0, (38)
saddle-node bifurcation will occur in the periodic branch. For G p > 0, when τ+ is long
enough to satisfy

Gp G p cos τ+ 2 (μ + G p sin τ+)2

( gG p sin τ+ + bG p cos τ+ + b(2ω − 2 − G p ) + gμ)2 − 3( g2 + b2 )((ω − 1 − + ) + ) < 0, (39)
2 2 4
there will be no positive roots for Equation (38), meaning under a positive gain, the delayed
position feedback can be used to reduce saddle-node bifurcation and thus bi-stability. For
example, given Gp = 0.3 and ω = 0.8, one can calculate from Equation (39) that there is
no saddle-node bifurcation for τ+ > 0.63. In contrast, it is hard to express the range of
τ+ for saddle-node bifurcation as a function of f. Fixing the value of f and combining
Equations (33), (38) and (39), one can obtain the condition for τ+ to reduce saddle-node
bifurcation. For instance, given f = 0.13, one can calculate that the condition of τ+ to reduce
saddle-node bifurcation is τ+ > 0.27.
Based on the above theoretical analysis, the results are verified by numerical simulation
as shown in Figure 7. Figure 7a shows the frequency-response for f = 0.13, Gp = 0.3 and
different values of time delay satisfying τ+ << π. Obviously, in Figure 7a, the amplitude of
the response reduces with the increase in time delay. Note that for τ+ = 0.35 (longer than
the theoretical values 0.27), there is no saddle-node bifurcation on the periodic branches,
showing that bi-stability and jump will not occur. It can be seen from Figure 7a that
the increase of time delay in the position feedback can reduce saddle-node bifurcation,
and thus bi-stability and jump. This can also be verified by the evolution of basins of
attraction with the increase of time delay (see Figure 8). It follows from Figure 8 that as τ+
increases, the basin of attraction of the higher-amplitude periodic attractor (the blue region)
will be eroded by the basin of attraction of the lower-amplitude one (see the red region).
Meanwhile, the safe basin (red and bule regions) becomes smooth. For ω = 0.80, one can
see from Figure 7b that even though the saddle-node bifurcation still exists, increase of time
delay can help to reduce the range of f where bistable attractors coexist. The corresponding

303
Mathematics 2022, 10, 2676

sequences of basin of attraction with the increase of f and τ+ will be discussed in detail in
the next subsection.

(a) (b)

Figure 7. Variation of the response amplitude of the system (29) with the excitation for Gp = 0.3.
(a) f = 0.13; (b) ω = 0.80.

(a) (b) (c) (d)

Figure 8. Evolution of basins of attraction with the increase of τ when f = 0.13 and ω = 0.78. (a) τ = 0;
(b) τ = 0.05; (c) τ = 0.15; (d) τ = 0.35.

4.1.2. Heteroclinic Bifurcation

To employ the Melnikov method to analyze the critical condition for heteroclinic
bifurcation in a delayed-feedback-controlled system, it is necessary to consider the delayed
position feedback as a perturbed term, namely, the values of the delay and the gain should
satisfy that the stability of the equilibrium of the uncontrolled linear system cannot be
changed by the delayed feedback. Since the equilibrium S1 (0,0) is stable for μ > 0, the
delay should not exceed the first value of the stability switch of the following linear
delayed system
d2 x μdx
+ + x = G p ( x ( T − τ+) − x ( T )). (40)
dT 2 dT
Its characteristic equation is given by

λ2 + μλ + 1 + G p − G p e−λ+τ = 0. (41)

Substituting λ = iv into (41) and separating the imaginary and real parts yields

− Ω p 2 + 1 + G p − G p cos(Ω p τ+) = 0, μΩ p + G p sin(Ω p τ+) = 0. (42)

304
Mathematics 2022, 10, 2676

By eliminating the trigonometric functions, Equation (42) becomes

Ω4p − (2 + 2G p − μ2 )Ω2p + 1 + 2G p = 0. (43)

For
μ2 μ2
G p > max{ − 1, + μ }, (44)
2 2
there are two different positive solutions of Equation (43) expressed as Ω p1 and Ω p2
(Ω p1 > Ω p2 > 0); the critical value of the delay for stability switch of S1 (0,0) is

1 1 − Ω2p1
τ0 = (2π − arccos(1 + )). (45)
Ω p1 Gp

For τ+ < τ0 , the delay position feedback can be regarded as a disturbed term.
Substituting the heteroclinic orbits (9) into the Melnikov function of the system (29) yields
√ √ "
2 2μ 8 2g 2 πω Gp
M± ( t 0 ) = − − + πω f csch( √ ) cos(ωt0 ) − l (τ+), (46)
3b 35b2 b 2 b 1
√ √ √
where l1 (τ+) = csch2 ( 22 τ+)(sinh( 2+
τ ) − 2+
τ ) > 0. The critical condition for heteroclinic
bifurcation can be expressed as

G p sech( πω
√ ) l1 ( τ
+)
f > f P (τ+) = f 0 + √ 2 , (47)
π 2bω

Accordingly, for G p > 0, the threshold of heteroclinic bifurcation f P (τ+) will increase
with time delay and be higher than f 0 . Fixing μ = 0.02 and Gp = 0.3, it can be calculated that
τ0 = 2.55.
Figure 9 shows the variation of f P (τ+) with the increase of time delay τ+ (+
τ << τ0 ). Then
the numerical values of f P (τ+) are obtained at which the boundary of the safe basin begin to
be unsmooth. Each numerical critical value of f P (τ+) is kept at two decimal places. We make
sure that if f is less than the numerical results f P (τ+), the boundary of safe basin will be
smooth. In Figure 9, the numerical results for the critical values of f are in agreement with
the analytical values, demonstrating that the threshold of the amplitude f for heteroclinic
bifurcation will increase monotonically with the delay for G p > 0 and τ+ < τ0 .

Figure 9. fP (τ) of the controlled system (29) versus time delay when Gp = 0.3 and ω = 0.80.

Figure 10 depicts the evolution of the safe basin of the controlled system (29) with
the delay τ+. In Figure 10a,e,i for τ+ = 0, safe basins as well as basins of attraction can
be considered as those of the uncontrolled system (6), illustrating that as the amplitude f
increases, safe basin will be eroded. Specifically, in Figure 10e for f = 0.30, only the basin of
attraction of the higher-amplitude attractor is left, which is seriously eroded by the white

305
Mathematics 2022, 10, 2676

region, showing the high possibility of excessive motion. For Gp = 0.3 and τ+ = 0.11, the
safe basin is obviously expanded. Besides, comparing Figure 10a–d, one can observe that
as τ+ increases, the safe basin changes from the union of red region and blue to the red
region itself, meaning that jump and initial-sensitive excessive motion are reduced by the
basin expansion of the lower-amplitude attractor. In Figure 10e–h, one can also draw this
conclusion. Note that Figure 10e,f, three color regions are mixed in the neighborhood of the
origin, which indicates the high probability of jump and initial-sensitive excessive motion.
As τ+ increases, they can be controlled. For f = 0.30 (see Figure 10i–l), even though the safe
basin is still fractal, its area becomes much larger with the increase of τ+. This shows that
the possibility of excessive motion is reduced. It follows from Figure 9 that the delayed
position feedback can suppress the erosion of the safe basin effectively when Gp is positive.

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

Figure 10. Sequences of safe basin of the system (29) with the increase of f and τ+ when Gp = 0.3.
(a) f = 0.11, τ+ = 0; (b) f = 0.11, τ+ = 0.05; (c) f = 0.11, τ+ = 0.15; (d) f = 0.11, τ+ = 0.35; (e) f = 0.15, τ+ = 0;
(f) f = 0.15, τ+ = 0.05; (g) f = 0.15, τ+ = 0.15; (h) f = 0.15, τ+ = 0.35; (i) f = 0.30, τ+ = 0; (j) f = 0.30, τ+ = 0.05;
(k) f = 0.30, τ+ = 0.15; (l) f = 0.30, τ+ = 0.35.

4.2. Delayed Velocity Feedback

4.2.1. Primary Resonant Response and Stability of Solutions
As in Section 4.1.1, by applying the Method of Multiple Scales, one can get the slow
flow equation of the delayed-velocity-feedback controlled system (30) as below:

da μ 3ga3 f sin ψ Gv (1 − cos τ+) dψ 3ba3 f cos ψ Gv sin τ+

= − a− + − a, a = a ( ω − 1) + + + a. (48)
dT 2 8 2 2 dT 8 2 2

306
Mathematics 2022, 10, 2676

The amplitude and phases of the system can be obtained via letting the right side of
Equation (48) be zero, given by

3g 3 3b
f sin ψ = μa + a + Gv (1 − cos τ+) a, f cos ψ = (2 − 2ω − a2 − Gv sin τ+) a. (49)
4 4
Thus, the amplitude of the approximated periodic solution x0 = a cos(ωT − ψ) can be
solved from the following equation:
2 2
3g 2 3b
(μ + a + Gv (1 − cos τ+)) a2 + (2 − 2ω − a2 − Gv sin τ+) a2 = f 2 . (50)
4 4
According to its characteristic equation, the stability switch will occur if there exists a
real number v+ satisfying
3ga2
v(μ + Gv (1 − cos τ+) + 2 ) = 0,
(μ+ Gv (1−cos τ+))2 (2−2ω − Gv sin τ+)2 27( g2 +b2 ) a4
(51)
− v2 + 4 + 4 + μ2 Gv (1 − cos τ+) + 3a2
4 ( μg + b (2ω − 2 + Gv sin τ+)) + 64 = 0.
3g 2
For Gv > 0, in Equation (51), one has μ + 2 a + Gv (1 − cos τ+) > 0, v+ = 0, and

(μ+ Gv (1−cos τ+))2 (2−2ω − Gv sin τ+)2 27( g2 +b2 ) a4

4 + 4 + μ2 Gv (1 − cos τ+) + 3a2
4 ( μg + b (2ω − 2 + Gv sin τ+)) + 64 = 0, (52)
implying that saddle-node bifurcation will occur in the delayed-velocity-feedback con-
trolled system. For Gv > 0, when τ+ is long enough to satisfy

4(μg + b(2ω − 2) + bGv sin τ+)2 − 3( g2 + b2 )((μ + Gv )2 + 2μGv + Gv 2 + 4(1 − ω )2 − 2Gv (2μ + Gv ) cos τ+ + 4(ω − 1) Gv sin τ+) < 0, (53)

there will be no positive roots for Equation (52), meaning no saddle-node bifurcation or
bi-stability. For example, given Gv = 0.3 and ω = 0.8, one can calculate from Equation (53)
that saddle-node bifurcation will disappear for τ+ > 0.68. Similarly to Section 4.1.1, although
delayed velocity feedback can be used to reduce the occurrence of saddle-node bifurcation,
it is hard to express the range of τ+ for SN bifurcation as a function of f. Fixing the value of f
and combining Equations (50), (52) and (53), one can obtain the condition for τ+ to reduce
saddle-node bifurcation. For instance, given f = 0.13, one can calculate that the condition of
τ+ to reduce saddle-node bifurcation is τ+ > 0.77.
The similar as in Section 4.1.1, for Gv = 0.3 and different values of time delay in
the velocity feedback, Figure 11 shows the variation of the response amplitude with the
excitation. It follows from the theoretical and numerical results depicted in Figure 11 that,
even though the response amplitude is not necessary to be suppressed with the increase of
time delay, the velocity feedback can reduce saddle-node bifurcation, and thus bi-stability
and jump. The controlling effect of velocity feedback on jump can also be verified by the
evolution of basins of attraction with the increase of time delay for f = 0.13, ω = 0.78 (see
Figure 12). To be different from the case of the delayed position feedback, as the delay
increases, the basin of attraction of the higher-amplitude attractor is expanded, whose
boundary becomes smoother and smoother. Finally, it becomes the only bounded attractor
(see Figure 12d). For ω = 0.80, if can be concluded from Figure 11b that the delayed velocity
feedback can suppress bi-stability. For the fixed value of ω, sequences of the basin of
attraction with the increase of f and τ+ will be discussed in detail in the next subsection.

307
Mathematics 2022, 10, 2676

(a) (b)

Figure 11. Variation of the response amplitude of the system (30) with the excitation for Gv = 0.3.
(a) f = 0.13; (b) ω = 0.80.

(a) (b) (c) (d)

Figure 12. Evolution of basins of attraction with the increase of τ for f = 0.13 and ω = 0.78. (a) τ+ = 0;
(b) τ+ = 0.2; (c) τ+ = 0.5; (d) τ+ = 0.8.

4.2.2. Heteroclinic Bifurcation under Delayed Velocity Feedback

As in Section 4.1.2, to apply the Melnikov method in the delayed-velocity-feedback
controlled system (30), we also need to treat the delayed feedback as a perturbed term of
this system. To this end, the delay τ+ should not exceed the first value of stability switch of
S1 (0,0) in the linear delayed system whose characteristic equation can expressed as

λ2 + μλ + 1 + Gv λ − Gv λe−λ+τ = 0 (54)

The first value of time delay for stability switch of S1 (0,0) will occur when λ = ±Ωv i.
Substituting it into Equation (54) and separating the imaginary and the real parts of
Equation (54) yields

− Ωv 2 + 1 = Gv Ωv sin(Ωv τ+), μΩv = Gv Ωv (cos(Ωv τ+) − 1). (55)

By eliminating the trigonometric functions, the equation above becomes

2
(Ω2v − 1) + (μ2 + 2μGv )Ωv 2 = 0. (56)

Since μ > 0, for Gv > 0, there will be no real roots of Equation (56), meaning that the
stability of the origin will not be changed. Then the delay velocity feedback of the delayed
system (30) can be regarded as a disturbed term. The Melnikov function of the controlled
system (30) can be expressed as

308
Mathematics 2022, 10, 2676

- +∞
M± ( t 0 ) = −√∞ (− μy ( T ) − gy3 ( T ) + f cos(ω ( T + t0 )) − Gv y( T ) + Gv y( T − τ+))y( T )dT
√ # √ (57)
= − 2 3b2μ − 8 2g
35b2
+ 2b πω f csch( 2πω 2Gv
2 ) cos( ωt0 ) − b l2 ,
√ √ √
where l2 (τ+) = csch2 ( 22+τ )( 2 − τ+coth( 22 τ+)) > 0. Accordingly, for
√ √
2bGv 2πω
f > f V (τ+) = f 0 + sinh( )l2 (τ+), (58)
bπω 2
And heteroclinic bifurcation may occur. In Equation (58), for Gv > 0, the threshold
of heteroclinic bifurcation f V (τ+) increases monopoly with the delay τ+. Given Gv = 0.3,
the variation of f V (τ+) with the increase of τ+ is presented in Figure 13. Numerical and
analytical results both illustrate that, for Gv > 0, the delayed velocity feedback can be used
to increase the threshold of the amplitude f for heteroclinic bifurcation, thus being effective
in reducing the probability of initial-sensitive excessive motion.

Figure 13. f V (τ) of the controlled system (30) versus time delay when Gv = 0.3.

Figure 14 shows the evolution of the safe basin of the controlled system (30) with the
increase of f and τ+. For τ+ = 0, Figure 14a,e,i are the same as Figure 10a,e,i, depicting the safe
basin of the uncontrolled system (6). Under a positive gain Gv = 0.3, when τ+ increases, the
area of safe basin of the delayed-velocity-feedback-controlled system will expand obviously.
To be different from the results of Section 4.1.2, for f = 0.11 and 0.15 where the two periodic
attractors coexisting in the uncontrolled system, as τ+ increases, the basin of attraction of the
higher-amplitude attractor is expanded and finally becomes the safe basin (see the blue
regions in Figure 14a–h). This illustrates that, although the delayed velocity feedback does
not necessary reduce the response amplitude, it reduces jump and excessive motion.

(a) (b) (c) (d)

Figure 14. Cont.

309
Mathematics 2022, 10, 2676

(e) (f) (g) (h)

(i) (j) (k) (l)

Figure 14. Sequences of safe basin of the system (30) with the increase of f and τ+ when Gv = 0.3.
(a) f = 0.11, τ+ = 0; (b) f = 0.11, τ+ = 0.20; (c) f = 0.11, τ+ = 0.50; (d) f = 0.11, τ+ = 0.80; (e) f = 0.15, τ+ = 0;
(f) f = 0.15, τ+ = 0.20; (g) f = 0.15, τ+ = 0.50; (h) f = 0.15, τ+ = 0.80; (i) f = 0.30, τ+ = 0; (j) f = 0.30, τ+ = 0.20;
(k) f = 0.30, τ+ = 0.50; (l) f = 0.30, τ+ = 0.80.

5. Discussion
In this paper, a typical relative rotation system is considered, and the phenomena
of jump and initial-sensitive excessive motion as well as their suppression via delayed
feedback is investigated. The Method of Multiple Scales and the Melnikov method are
applied to analyze the conditions for bi-stability and initial-sensitive excessive motion,
respectively. By introducing the basins of attraction and safe basin to describe the extent
of jump and initial-sensitive excessive motion, respectively, the point mapping approach
is employed to present the numerical results which matches the theoretical approach,
verifying the validity of the analysis. Some significant results are presented as follows.
(1) The variation of excitation may induce the coexistence of bistable periodic attractors,
which can be ascribed to saddle-node bifurcation.
(2) The increase of the excitation amplitude may cause initial-sensitive excessive motion,
which can be due to heteroclinic bifurcation.
(3) Under positive coefficients of the feedback gain, the delayed position feedback and
the delayed velocity feedback can reduce saddle-node bifurcation and heteroclinic
bifurcation so as to suppress jump and initial-sensitive excessive motion. Compara-
tively, the former can also reduce the amplitude of the response, while the latter may
not; the former works well if time delay does not exceed the first stability switch of
the trivial equilibrium, while the latter does not have that restriction.
This work presents a detailed analysis of jump and initial-sensitive excessive motion
of a typically relative rotation system, which may be beneficial for the performance im-
provement of rotors and main bearings. The relevant experimental investigations will be
included in our future work.

Author Contributions: Conceptualization, H.S.; formal analysis, H.S.; funding acquisition, H.S.;
Investigation, Z.C.; methodology, H.S.; project administration, H.S.; software, Z.C.; supervision, H.S.;
validation, H.S.; visualization, Z.C.; writing—original draft, Z.C. and H.S.; writing—review & editing,
H.S. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the National Natural Science Foundation of China, grant
number 11472176.

310
Mathematics 2022, 10, 2676

Institutional Review Board Statement: Not applicable.

Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Liang, Y.; Li, N. Optimal vibration control for nonlinear systems of tracked vehicle half-car suspensions. Int. J. Control. Autom.
Syst. 2017, 15, 1675–1683. [CrossRef]
2. Renson, L.; Noël, J.P. Complex dynamics of a nonlinear aerospace structure: Numerical continuation and normal modes. Nonlinear
Dyn. 2015, 79, 1293–1309. [CrossRef]
3. Shi, J.; Gou, X. Bifurcation and Erosion of Safe Basin for a Spur Gear System. Int. J. Bifurc. Chaos 2018, 28, 1830048. [CrossRef]
4. Li, J.; Wu, H. Bifurcation, chaos, and their control in a wheelset model. Math. Methods Appl. Sci. 2020, 43, 7152–7174. [CrossRef]
5. Luo, Z.; Wang, J. Research on vibration performance of the nonlinear combined support-flexible rotor system. Nonlinear Dyn.
2019, 98, 113–128. [CrossRef]
6. Li, Y.; Luo, Z. Dynamic modeling and stability analysis of a rotor-bearing system with bolted-disk joint. Mech. Syst. Signal Process.
2021, 158, 107778. [CrossRef]
7. Shi, P.; Liu, B. Chaotic motion of some relative rotation nonlinear dynamic system. Acta Phys. Sin. 2008, 57, 1321–1328.
8. Li, Z.; Gou, X. Erosion and bifurcation of safety-attraction basin for multi-state meshing gear transmission system under tooth
safety condition. J. Vbration Shock 2021, 40, 63–74.
9. Wang, K.; Guan, X. Precise periodic solutions and uniqueness of periodic solutions of some relative rotation nonlinear dynamic
system. Acta Phy. Sin. 2010, 59, 3648–3653. [CrossRef]
10. Xiao, L.; Xuan, C. The periodic solution problem of a relative rotation nonlinear dynamic system with time-varying stiffness. Acta
Phy. Sin. 2013, 62, 21–25.
11. Li, X.; Yan, J. The periodic solution problem of a relative rotation nonlinear system with nonlinear elastic force and generalized
damping force. Acta Phy. Sin. 2014, 63, 36–41.
12. Shi, P.; Han, D. Chaos and chaotic control in a relative rotation nonlinear dynamical system under parametric excitation. Chin.
Phys. B 2010, 19, 116–121.
13. Verichev, N. Chaotic torsional vibration of imbalanced shaft driven by a limited power supply. J. Sound Vib. 2012, 331, 384–393.
[CrossRef]
14. Liu, B.; Zhao, H. Bifurcation and chaos of some relative rotation system with triple-well Mathieu-Duffing oscillator. Acta Phy. Sin.
2014, 63, 174502.
15. Liu, S.; Tian, S. Chaos of a kind of nonlinear relative rotation system based on the effect of Coulomb friction. Acta Phy. Sin. 2015,
64, 247–254.
16. Zhu, L.; Li, Z. Evolutionary mechanism of safety performance for spur gear pair based on meshing safety domain. Nonlinear Dyn.
2021, 104, 215–239. [CrossRef]
17. Ju, J.; Wei, L. Dynamics and nonlinear feedback control for torsional vibration bifurcation in main transmission system of scraper
conveyor direct-driven by high-power PMSM. Nonlinear Dyn. 2018, 93, 307–321. [CrossRef]
18. Shang, H. Pull-in instability of a typical electrostatic MEMS resonator and its control by delayed feedback. Nonlinear Dyn. 2017,
90, 171–183. [CrossRef]
19. Wang, Q.; Wu, H. The effect of fractional damping and time-delayed feedback on the stochastic resonance of asymmetric SD
oscillator. Nonlinear Dyn. 2022, 107, 2099–2114. [CrossRef]
20. Zhao, Y.; Li, C. The delayed feedback control to suppress the vibration in a torsional vibrating system. Acta Phy. Sin. 2011, 60,
417–425.
21. Shang, H.; Han, Y. Suppression of chaos and basin erosion in a nonlinear relative rotation system by delayed position feedback.
Acta Phy. Sin. 2014, 63, 88–95.
22. Rezaei, M.; Khadem, E.S.; Friswell, I.M. Energy harvesting from the secondary resonances of a nonlinear piezoelectric beam
under hard harmonic excitation. Meccanica 2020, 55, 1463–1479. [CrossRef]
23. Siewe, S.M.; Hegazy, H.U. Homoclinic bifurcation and chaos control in MEMS resonators. Appl. Math. Model. 2011, 35, 5533–5552.
[CrossRef]
24. Danico, K.; Tanmoy, C.; Milan, C.; Sondipon, A.; Karličić, D.; Chatterjee, T.; Cajić, M.; Adhikari, S. Parametrically amplified
Mathieu-Duffing nonlinear energy harvesters. J. Sound Vib. 2020, 488, 115677.
25. Zhu, Y.; Shang, H. Multistability of the vibrating system of a micro resonator. Fractal Fract. 2022, 6, 141. [CrossRef]
26. Rega, G.; Lenci, S. Dynamical integrity and control of nonlinear mechanical oscillators. J. Vib. Control 2008, 14, 159–179. [CrossRef]
27. Liu, C.; Yan, Y.; Wang, W. Resonances and chaos of electrostatically actuated arch micro/nanoresonators with time delay velocity
feedback. Chaos Solitons Fractals 2019, 10, 109512. [CrossRef]

311
 mathematics

Article
The General Dispersion Relation for the Vibration Modes of
Helical Springs
Leopoldo Prieto *, Alejandro Quesada, Ana María Gómez Amador and Vicente Díaz

Department of Mechanical Engineering, Universidad Carlos III de Madrid, 28911 Leganes, Spain;
alejandro@ing.uc3m.es (A.Q.); amgomez@ing.uc3m.es (A.M.G.A.); vdiaz@ing.uc3m.es (V.D.)
* Correspondence: leprieto@ing.uc3m.es

Abstract: A system of mathematical equations was developed for the calculation of the natural
frequencies of helical springs, its predictions being compared with finite element simulation with
ANSYS® . Authors derive the general equations governing the helical spring vibration relative to the
Frenet trihedral representing the normal, binormal and tangent unit vectors to the spring medium
line. The dispersion relation ω = f (k ) has been obtained to model a wave traveling along the axis of
the wire.

Keywords: helical spring; vibration; Frenet trihedral; dispersion relation; natural frequency

MSC: 70B15

1. Introduction
Helical springs are one of the most frequently used elastic elements in mechanical
engineering. They are used in the most diverse structures as elastic energy accumulators
supplementing damping devices, e.g., automobile or railroad suspensions, and in advance
Citation: Prieto, L.; Quesada, A.; and return devices, e.g., camshafts and valves of internal combustion engines (Wahl [1],
Gómez Amador, A.M.; Díaz, V. The Shigley [2]; Kobelev [3]). One of the spring failure modes is caused by resonance vibration
General Dispersion Relation for the that occurs when the spring is excited with a periodic signal of frequency equal to its
Vibration Modes of Helical Springs. natural frequency.
Mathematics 2022, 10, 2698. https:// Love [4] develops equations to study the static response of helical springs subjected
doi.org/10.3390/math10152698
to large deformations. Stokes [5] investigates vibrations in the case of a spring subjected
Academic Editor: Dan B. Marghitu to shock loads. Gironnet and Louradour [6] determine more precise expressions for the
calculation of the natural frequencies of the helical spring. There are several numerical
Received: 24 June 2022
methods to determine the natural frequencies, the main ones being: the transfer matrix
Accepted: 28 July 2022
method (Yildirim [7]), the dynamic stiffness formulation (Lee [8]), and the pseudo-spectral
Published: 30 July 2022
method (Lee [9]). Becker [10] determines with the transfer matrix method the resonant
Publisher’s Note: MDPI stays neutral frequencies of a spring subjected to an axial compression load. Jiang [11] studies forced
with regard to jurisdictional claims in vibrations and wave propagation in springs using the Laplace transform.
published maps and institutional affil- Wahl [1] already proposed in the middle of the 20th century the equations that are
iations. widely used today for the design of springs. Among these equations, Equation (1), which
provides the stiffness constant of the spring subjected to an axial force, and Equation (2),
which provides the natural frequency of the spring placed between two parallel flat plates,
stand out.
Copyright: © 2022 by the authors.
Gd4
Licensee MDPI, Basel, Switzerland. kWahl = (1)
8D3 N
This article is an open access article
distributed under the terms and
where kWahl is the spring axial stiffness, G is the shear modulus of elasticity, d is the wire
conditions of the Creative Commons diameter, D is the helix diameter and N is the number of active coils.
Attribution (CC BY) license (https:// Where f is the natural harmonic frequencies of a spring in Hz, and W is the mass of the
creativecommons.org/licenses/by/ spring in kg. The fundamental frequency is determined for m = 1 and is usually the most
4.0/). important frequency in practice. Replacing in Equation (2) the value of the axial stiffness

Mathematics 2022, 10, 2698. https://doi.org/10.3390/math10152698 313 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 2698

we obtain Equation (3) also called Harignx’s equation [12], who experimentally verified
its validity. "
m kWahl
f = m = 1, 2, 3 . . . . (2)
2 W

d G
f = (3)
2πR2 N 32ρ
R being the helix radius and ρ the density of the material.
Equation (3) indicates that for a given material, the fundamental frequency of a helical
spring is proportional to the wire diameter and inversely proportional to the product of the
helix diameter and the number of active coils, i.e., the resonance frequency of the spring
depends on all the spring design parameters. If a higher frequency is desired for the same
spring diameter, it is sufficient to increase the wire diameter (this increases the spring
stiffness), and vice versa. If you want to vary the resonance frequency while keeping the
wire diameter constant, just change the spring diameter or the number of coils of the spring.
All the expressions shown above are raised with respect to an inertial reference system
whose main X axis coincides with the symmetry axis of the helical spring.
By design of the machine itself, the spring is always inserted between two masses in
relative oscillatory motion and its function is to keep them away from each other ensuring
at all times that they do not come into contact. The design of the helical spring depends
directly on the inertia between the masses at any given moment. In addition to the diameter
of the wire and the spring itself, the number of coils and the pitch are responsible for
achieving non-interaction between the mechanical elements in relative motion.
It is interesting to note that the behavior of the spring changes radically when its
operation is associated with a mass much greater than that of the spring itself. For example,
in the damping systems of a railway bogie such as the one in Figure 1. In that case, the
mass of the spring is much lower than the mass of the railway car it is supporting and
Equations (2) and (3) are no longer valid. Den Hartog [13] shows that in that configuration
the natural frequency follows Equation (4).
"
1 g
f = (4)
2π δest

where g is the gravitational constant and δest is the static deflection of the spring under the
weight of the suspended load.

Figure 1. Damping springs of a railway bogie (Díaz [14]).

Sauvage [15] and Campedelli [16] studied in detail the behavior of the springs of a
bogie and simulated in finite elements the anomaly presented by the springs during axial
compression called “effort de chasse”. This anomaly consists of a rotation and a lateral
displacement appearing on the bearing surface of the spring. The “effort de chasse”, also

314
Mathematics 2022, 10, 2698

called transverse flexibility of the spring, is caused by the forces and moments that appear
at the ends of the spring during its axial compression. Campedelli [16] determines that the
transverse flexibility is determined by both the geometry of the spring end and the way in
which the gap between the coils closes during axial compression. The “effort de chasse”
is an interesting anomaly in railway suspension springs since in their displacement the
springs could interfere with other mechanical elements of the bogie generating interferences
and breakages. To avoid this, the railway standard [17] requires a test for the springs of the
highest category according to the scheme in Figure 2a. From the test, both the direction
and the transverse bowing force (Φc) of the spring subjected to a defined axial load are
obtained. The direction of the transverse direction is marked on the spring by a permanent
system to take it into account in its mounting on the bogie. Figure 2b shows an example of
a bogie spring assembly compensating for the spring bowing forces.

(a) (b)

Figure 2. “Effort de chasse” or transverse flexibility: (a) schematic of the test device; (b) mounting on
a spring bogie compensating the “effort de chasse”.

Kobelev [18] also studies another type of anomaly in the behavior of springs such
as the transverse vibration of the spring once it enters into oscillations close to resonant
frequencies. Kobelev [18] found that the fundamental natural frequency of transverse
oscillations turns to be to zero when the lateral buckling of the spring occurs.
Yildirim [19] demonstrates that Equation (1), which is usually used to determine
the spring stiffness, is only valid for spring angles α ≤ 10◦ . Figure 3a shows the spring
geometry indicating the position of the angle α.

(a) (b)

Figure 3. Helical spring: (a) geometrical parameters; (b) Frenet trihedral.

Yildirim [19] indicates that Equations (1) and (2) only take into account the effect of
the torsion of the spring section when compressing it. However, as Yildirim [19] points out
when the axial compression force acts, both stresses due to torsion and bending moment
and forces normal and shear to the section appear. These last three effects are negligible
for spring angles α ≤ 10◦ where the torsional effect dominates. Yildirim [19] proposes

315
Mathematics 2022, 10, 2698

a global equation to determine the deformation of the spring subjected to an axial force.
Additionally, Kato [20] proposes recently an equation outside the elastic range to determine
large spring deformations.
In general, the previous authors focus the study of the spring behavior using a conven-
tional Cartesian system in which the static or dynamic displacements are referred to as the
first axis of symmetry of the helical spring, X. With this methodology the global behavior
of a spring as a whole can be known with precision, i.e., constituted by a series of coils of a
certain diameter and a certain thickness of the wire, contemplating in some of the occasions
the ending effect.
This type of methodology, therefore, allows us to analyze the elastic effect of the spring
as a whole, but it is not able to explain some mechanical behaviors intrinsic to its own
constitution, i.e., it does not study the phenomenon occurring in the coil itself. In this way,
anomalous spring behaviors such as those described above cannot be fully explained.
In order to achieve a reference system that meets these needs, the Frenet trihedral is
used in this article, since it is intrinsic to the wire of the spring itself, configuring a reference
system positioned according to the neutral line of the helix of the coil as shown in Figure 3b.
In the present study, the equations of motion of the coil of a helical spring are obtained
in order to determine its modes of vibration. This is an original and novel approach since
the usual practice is to obtain the modes of vibration of the complete spring referred to as a
reference system according to its joint motion. Additionally, a sensitivity analysis has been
performed for the main geometrical parameters of the coil: i.e., the helix diameter, the wire
diameter and pitch. The obtained vibration modes have been validated satisfactorily with
finite element calculations with ANSYS® R17 (release 17).

2. Model of Coupled Bending and Torsional Oscillation in a Spring

2.1. Frenet’s Trihedral
For the approach of the oscillation model of the coil, the coordinate system formed by
the Frenet trihedral, intrinsic to the neutral line of the spring, has been considered.
Figure 4 shows a differential element of the length of the curve defined by the midline
of the coil of a helical spring. The Frenet trihedral or intrinsic trihedral is positioned on
→ →→
it. It is an orthonormal trihedron ( t , n, b ) whose components are named tangent, normal
and binormal. The tangent vector is oriented to the neutral line of the loop. The normal is
directed towards the bending center of curvature of the loop. The binormal vector follows
the orientation defined by the other two. The vectors t and n form a tangent plane to the
loop, containing the element ds, and perpendicular to the binormal vector which is directed
towards the bending curvature of this element.

Figure 4. Intrinsic trihedral and principal directions of inertia of the section.

316
Mathematics 2022, 10, 2698

Considering as a scalar variable the arc length of the midline of the loop(s), we obtain
the well-known Frenet–Serret equations, Equation (5), which give us the derivative with
respect to s of the components of the trihedral as a function of the curvature of the midline
of the coil. ⎛ → ⎞ ⎛→⎞ ⎛ ⎞
⎛ ⎞ →
dt
t 0 Ω − Ω t
⎜ → ⎟ ⎜→⎟
ds 3 2 ⎜→⎟
⎜ d n ⎟ = ⎜ n  ⎟ = ⎝ − Ω3 0 Ω1 ⎠ ⎜ n⎟ (5)
⎝ ds→ ⎠ ⎝ → ⎠ ⎝→ ⎠
Ω2 − Ω1 0
db
ds
b  b

where Ω1 = τ0 , Ω2 = −k0 sin φ0 and Ω3 = k0 cos φ0 . Where k0 is the bending curvature,

τ0 the torsional curvature and φ0 the angle formed by the principal directions of inertia
with the Frenet trihedron.
Equation (5) can be expressed as a function of the vector product of the Darboux vector
→
Ω(s) by the components of the intrinsic trihedron obtaining Equation (6).
⎛ →⎞ ⎛→⎞
dt
t
⎜ ds→ ⎟ → ⎜→⎟
⎜dn ⎟ = Ω×⎜ ⎟ (6)
⎝ ds→ ⎠ ⎝n⎠ →
db b
ds

where the Darboux vector is defined according to Equation (7):

→ → → →
Ω ( s ) = Ω1 t + Ω2 n + Ω3 b (7)

The variation of any vector function, for example, the rotation and translation of the
section, with respect to the length s of the arc follows the expression of Equation (8a) that we
will simplify using the notation of Equation (8b) where we will call [ R] the tensor associated
→
with the operation Ω×. 
→ →
dV d V  → →
= +Ω×V (8a)
ds ds  S
( ) ( )
→ → →
V  = V  + [ R] V (8b)
⎛ ⎞
0 Ω3 − Ω2
[ R ] = ⎝ − Ω3 0 Ω1 ⎠ (8c)
Ω2 − Ω1 0
→

dV 
where ds  S is the derivative calculated as if the Frenet trihedron were fixed along the curve.
That is, the variation of the function along the curve has two terms, the first represents the
variation of the vector within the basis itself and the second represents the variation of the
basis itself along the curve.

2.2. Rotation of the Spring Cross-Section

The spring in its natural geometrical configuration undergoes a deformation caused
by external forces and moments. The plane cross-section remains plane and we neglect its
small deformation. As shown in Figure 5, the cross-section of the deformed part is obtained
→ →
by a translation η (s) of the center and a rotation ψ (s) of the section. Therefore, the point
P the intersection of the cross-section with the midline, will be transformed into the point
→ →
P1 (where PP1 = η (s)) and the perpendicular axes (j,k), coinciding with the normal and
→ →
binormal of the intrinsic trihedron before the deformation, will become the axes j1 and k1
→
which we also assume to be perpendicular to each other. The axis i1 is obtained as the
→ → →
vector product of the previous ones, i1 = j1 × k1 .

317
Mathematics 2022, 10, 2698

Figure 5. Schematic of the oscillations of the spring section.

Due to the shear produced by the shear forces, the i-axis is not transformed into the
i-axis1 by rotation and translation. Consequently, the tangent vector to a point on the
midline of the part will no longer coincide with the first axis of the trihedron as it did in the
→
initial equilibrium state but will form an angle β , called the shear angle, proportional to
the shear stress. If we do not consider the deformation due to the shear stress, the shear
angle is canceled, so that the tangent to the midline of the part will be perpendicular to the
section, or what is the same, will coincide with the first axis of the trihedron.
→
The derivative of ψ with respect to s is given to first order by Equation (9a,b).
&→' → →
ψ  = [ M N ] N + [ R] ψ (9a)

⎛ ⎞ ⎛ ⎞
1 1
0 0 I1∗ 0 0
⎟ ⎜ ⎟
k GI1
⎜ 1
[ MN ] = ⎝ 0 1
EI2 0 ⎠=⎜⎝0
1
I2∗ 0⎟⎠. (9b)
1 1
0 0 EI3 0 0 I3∗

where M N is the tensor associated with the torsional-flector torque increment, E is the mod-
ulus of elasticity, G is the shear modulus of elasticity, Ii is the geometric moment of inertia
→
and k1 is the torsion constant. N is the vector of internal torque increase. Its first component
is the torsional moment, and the other two components are the bending moments in the
normal and binormal directions. I1∗ is the torsional stiffness, I2∗ is the bending stiffness
along the normal axis, and I3∗ is the bending stiffness along the binormal axis.

2.3. Translation of the Spring Cross-Section

→
The derivative of the displacement vector η with respect to s is given to first order by
Equation (10a–c).
&→' → → →
η  = [ MF ] F + [S] ψ + [ R] η (10a)
⎛ 1 ⎞ ⎛ ⎞
EA 0 0 E∗ 0 0
⎜ 0 ⎟⎠ = ⎝ 0 G1∗ 0 ⎠
1
[ MF ] = ⎝ 0 k2 GA (10b)
0 0 1
k3 GA
0 0 G2∗
⎛ ⎞
0 0 0
[ S ] = 0 0 1⎠
⎝ (10c)
0 −1 0

318
Mathematics 2022, 10, 2698

where MF is the tensor associated with the forces and ki the shear constants, or the so-called
→
Timoshenko’s k-factors [21]. Being for symmetric sections k2 = k3 . F is the vector of internal
force increment. Its first component is the axial force, and the other two components are
the shear forces according to the normal and binormal. EA is the axial stiffness and k2 GA
is the shear stiffness.
Equations (9a,b) and (10a–c) constitute a system of six scalar equations for the twelve
→ → → →
components of ψ, η , F , N. The remaining six equations are obtained from the dynamic
study of the system for which the equations of the linear momentum and angular momen-
tum of the part element comprised between s and s + ds are posed.

2.4. Linear Momentum

→
Considering the linear momentum, the derivative of F with respect to s is given by
Equation (11). ( ) ..
→ → →
F  = λ η + [ R] F (11)

where λ = ρ· A is the linear density per unit length.

2.5. Angular Momentum

→
Considering the angular momentum, the derivative of N with respect to s is given by
Equation (12a,b), which takes into account the effects of rotational inertia introduced by
Rayleigh [22].
( ) ..
→ λ → → →
N  = [ I ] ψ + [S] F + [ R] N (12a)
A
⎛ ⎞
I1 0 0
[ I ] = 0 I2 0 ⎠
⎝ (12b)
0 0 I3
Table 1 summarizes the complete system of equations obtained and which describes
the behavior of the spring coil subjected to bending and torsion. It consists of four vector
equations in linear partial derivatives, non-homogeneous and of second order.

Table 1. System of Vectorial Equations to model spring behavior.

Vectorial Equations
&→' → →
ψ = [ M N ] N + [ R] ψ
&→' → → →
η = [ MF ] F + [S] ψ + [ R] η
( ) ..
→ → →
F = λ η + [ R] F
( ) ..
→ → → →
λ
N = A [ I ] ψ + [S] F + [ R] N

2.6. Dispersion Relation

If in the Table 1 system of equations we change the variable using Equation (13a–d),
we can simulate the movements in time of the spring and then perform a vibration analysis
→ → → →
of its oscillatory behavior in the vector variables ψ, η , F , N. We will assume a wave
traveling along the axis of the wire. Thus
→ → 
ψ (s, t) = Re ψ0 (s)·ei(ks−ωt) (13a)

→
→ 
η (s, t) = Re η0 (s)·ei(ks−ωt) (13b)

319
Mathematics 2022, 10, 2698

 
→ →
F (s, t) = Re F0 (s)·ei(ks−ωt) (13c)
 
→ →
N (s, t) = Re N0 (s)·ei(ks−ωt) (13d)

The new parameters introduced are the wavenumber, k, and the angular velocity of
the wave, ω (rad/s). Where the wavenumber k, is related to the length λ of the vibrating
wave, with Equation (14).
2π
k= (14)
λ
Substituting the system of equations in Table 1 we arrive at the following system in
Equation (15a–d).
→ →
[Λ] F − λω 2 η = 0 (15a)
→ → →
[Λ] η + [ MF ] F + [S] ψ = 0 (15b)
→ →
[Λ] ψ + [ M N ] N = 0 (15c)
→ λω 2 → →
[Λ] N − [ I ] ψ + [S] F = 0 (15d)
A
where the matrix [Λ] follows the expression of Equation (16).
⎛ ⎞
−ik k0 0
[Λ] = ⎝−k0 −ik τ0 ⎠ (16)
0 −τ0 −ik
→
Solving for η in Equation (15a) and replacing it in Equation (15b), and similarly
→
solving for N in Equation (15c) and replacing in Equation (15d) we obtain the following
vector Equation (17a,b):
→ →
[ A] F + [S] ψ = 0 (17a)
→ →
[ B] ψ + [S] F = 0 (17b)
where the matrices [ A], [ B] are calculated with Equations (18) and (19).

1
[ A] = [ Λ ]2 + [ M F ] (18)
λω 2
→ →
[ B] ψ + [S] F = 0 (19)
Equation (17a,b) constitutes a system of six equations with six unknowns
(F1 , F2 , F3 , ψ1 , ψ2 , ψ3 ) and with null independent terms. It is therefore a homogeneous
system and for it to have a solution other than the trivial, null solution, it must be verified
that its determinant is null, i.e., Equation (20).

 2 ∗ $ 2 % 
λω E − k + β 9 β 4 − 2ikΩ β 5 + 2ikΩ2 0 0 0 
 $ 3 % 
 2ikΩ + β λω 2 G1∗ − k2 + β 8 β 6 − 2ikΩ 0 0 λω 2 
 3 4 $ 1 % 
 β 5 − 2ikΩ2 β 6 + 2ikΩ1 λω 2 G2∗ − k2 + β 7 0 −λω 2 0 
 = 0 (20)
 0 0 0 k2 I1∗ + α4 − ω 2 λ1 − β 4 I3∗ + ikα1 − β 5 I2∗ − ikα2 

 0 0 1 − β 4 I3∗ − ikα1 k2 I2∗ + α5 − λ2 ω 2 − β 6 I1∗ + ikα3 

 0 1 0 β 5 I2∗ − ikα2 β 6 I1∗ + ikα3 −k2 I3∗ + ω 2 λ3 − α6 

Operating the determinant and simplifying we obtain the alphanumerical

Equation (21), which would be the dispersion relation, ω = f (k ), which relates the length
of a wave with its oscillation frequency in the spring coil.

m1 ω 12 + m2 ω 10 + m3 ω 8 + m4 ω 6 + m5 ω 4 + m6 ω 2 + m7 = 0 (21)

320
Mathematics 2022, 10, 2698

where the coefficients of the polynomial are calculated according to the expressions given
in Equation (A1) of Appendix A.
The real roots in ω 2 of Equation (21) will give us the number of branches of the
dispersion equation for each value of the wavenumber k.
If in the dispersion relation obtained for the spring we cancel the independent term,
m7 (which is equivalent to making ω → 0 ), we will have the solutions of the wave number
shown in Equation (22a,b).
k∗ = 0 (22a)
#
k∗∗ = Ω21 + Ω23 (22b)

That is to say, there exists a value of the wave number k different from zero (to which
corresponds, therefore, a finite wavelength) for which one of the real and positive roots of
the dispersion relation (degree 12) cancels, so that the branch corresponding to this solution,
for any value of k, is tangent to the x-axis in k = k∗∗ .
If we do k → 0 we will obtain the number of branches of the dispersion relation
(which will be equivalent to the number of real roots of the equation). For this case, the real
and positive branches are determined by Equation (23a–c):

ω=0 (23a)

ρ2 I1 I2 ω 4 − ω 2 [kG { AρI1 + ( β 1 + β 3 ) I2 } + EI2 ρ( I1 β 1 + I2 β 3 )]

(23b)
+kGEI2 [( β 1 + β 3 )( I1 β 1 + I2 β 3 ) + Aβ 3 ] = 0
ρ2 I2 ω 4 − ω 2 ρ[ EI2 ( β 1 + β 3 ) + kG [( β 1 I2 + A) + β 3 kGI1 ]]
(23c)
+( β 3 kGI1 + β 1 EI1 )( β 3 E + β 1 kG ) + β 3 kGEA = 0
The wave numbers k obtained from Equation (23a–c) are indicators of the spring
oscillation modes. Equation (23b) represents the bending oscillations of a differential
element of spring length in plane 1, 2 determined by the principal directions of inertia
of a cross-section (Figure 6). Equation (23c) represents the torsional oscillations of the
differential element around the principal direction 1 of Figure 6. These oscillation modes
will be explained also later in the finite elements section.

Figure 6. Spring Oscillation Modes.

The result of the oscillation is bending accompanied by torsion which tends to make
the ends of the coil open and close in time, oscillating according to the movement. This is
composed of two modes: opening and closing of the ends, the coil remaining at all times
inscribed in the original helix of the spring (of diameter D) and the other mode consists
of the ends of the coil entering and leaving the original helix of diameter D, forming an
ellipse. This last mode responds to the reason why the springs, under certain dynamic
conditions, pull the coils out of their original helix, and may impact other elements of the
machine, thus affecting its operation. Authors consider this can explain part of the “Effort
de Chasse” anomaly described in Section 1.

321
Mathematics 2022, 10, 2698

In this article, we are in the case of a freely oscillating coil. In the case of several coils
joined together, it would be necessary to consider the ligatures of their ends as shown in
Figure 7 for the end of a squared and ground spring. That is to say, with a helix angle of
zero degrees and a flat seat, which allows a better load transfer. The assembly thus defined
becomes an oscillatory mode in which the spring as a whole adopts the barrel configuration.
Pearson [23] studies the important contribution of the ends to the calculation of the spring
natural frequencies.

Figure 7. The end of squared and ground helical spring.

The dispersion relation allows us to obtain the normal modes of oscillation by applying
the appropriate boundary conditions. For example, in the case of a helical spring with one
end clamped and the other free, the boundary conditions would be as follows: the rotation
→ →
vector ψ and the translation vector η cancel out at the clamped end and the internal torque
→ →
N and the internal stress F cancel at the free end. Assuming the clamped end for s = 0 and
the free end for s = L, Equation (24) is satisfied which implies that the wave number k takes
the values of Equation (25).
→ →
ψ (0, t) = ψ ( L, t) = 0 (24)
n·π n·π
k= = √ n = 1, 3, . . . .. (25)
2L 2N h2 + π 2 D2
Entering the value of k calculated according to Equation (25) in the dispersion
Equation (21), the natural oscillation frequencies for the helical spring are obtained.
In order to validate the proposed theoretical model, a comparison with finite elements
and with the experimental tests of a helical spring is made below.

3. Application of Finite Elements

The previous calculation performed to obtain the natural frequencies of the system
has been compared with the results of the resonance frequencies obtained by numerical
simulation. In addition, a complete modal analysis is performed. In order to compare
the results, the oscillation modes of the one spring turn are calculated using standard
computational methods. The simulations have been carried out by modeling the spring
geometry by Finite Elements using ANSYS® software.
A three-dimensional geometry of the helicoid representing the neutral fiber of a helical
spring has been faithfully reproduced. The three-dimensional geometry has been generated
from SOLID EDGE software, and it has been imported by the FEM software in the *.iges
format. The SOLID EDGE software provides flexibility in the modelization and facilitates
the actual geometrical definition compared to the CAD tool integrated in ANSYS. The
spring is thus reproduced volumetrically.
We proceed to calculate in both cases, i.e., the theoretical model proposed in the
previous section with one solved in finite elements, in order to compare the results.

322
Mathematics 2022, 10, 2698

In finite elements, a meshing has been performed using extruded hexahedral ele-
ments of 2 mm characteristic size, which represents the circumference of the wire cross-
section (10 mm diameter) with 16 nodes in its contour. This results in 6000 nodes per turn,
54,000 nodes for the complete spring and 162,000 degrees of freedom.
The meshing strategy used consists of an extrusion method from the two-dimensional
meshing (with quadrilaterals) of the cross-section, which guarantees a regular meshing
that avoids geometric distortions of the elements and achieves a low bandwidth for the
stiffness matrix, resulting in a better performance of the iterative methods of resolution.
The SOLID185 element has been selected in its homogeneous form, which is frequently
used to model three-dimensional structures in ANSYS. The element consists of eight nodes
with three degrees of freedom per node: translations in the three directions of space.
SOLID185 uses the reduced selective integration method.
The material has been modeled linearly, with steel properties considering Young’s
modulus and Poisson’s ratio.
The problem has been plausibly solved using the Lanczos method for the determina-
tion of eigenfrequencies and eigenmodes.
In free-free solid conditions, the different natural frequencies of one spring turn
belonging to a helical spring, with different helix diameters and wire thickness have been
calculated. In all cases, the pitch of the helix has always been the same.
In all cases, the corresponding oscillation modes are studied. Table 2 shows the results.
Columns 4 to 9 show the frequency and the mode of vibration. Essentially, the first mode
corresponds to an opening between ends, such that the original helix is transformed into
an ellipsoidal helix (Figure 8a). In this case, the predominant deformation of the helix
is bending. The second mode corresponds to the relative motion between these ends
(Figure 8b) but preserves the circumference of the helix, which corresponds to a mode in
which the torsional effect of the helix is more predominant than its bending. In this second
case, the pitch of the coil would increase significantly.

Table 2. FEM calculation of the first natural frequencies (Hz) of helical spring.

Natural Frequencies (Hz) Calculated by ANSYS

D (mm) d (mm) Pitch (mm) Bending 1 Torsional 1 Bending 2 Torsional 2 Bending 3 Torsional 3
80 18 39.8 960 1218 2111 3184 4543 6322
80 22 39.8 1168 1468 2556 3733 5428 7414
80 28 39.8 1475 1823 3199 4435 6623 8254
108 18 39.8 539 664 1183 1815 2595 3721
108 22 25 667 792 1451 2155 3178 4339
108 22 39.8 657 806 1439 2162 3132 4349
108 22 50 648 817 1426 2169 3091 4355
108 28 39.8 832 1012 1815 2630 3899 5127
220 18 39.8 133 158 290 455 648 968
220 22 39.8 162 193 354 553 790 1169
220 28 39.8 207 245 450 695 1000 1454
N = 1 turn, E = 210 GPa, G = 84 GPa, ρ = 7800 kg/m3 , ν = 0.24407.

From these two basic modes, the deformation profiles are reproduced with increasing
inflection points. The third (Figure 8c) and fifth (Figure 8e) modes constitute double bending
and triple bending, respectively. Always without leaving the imaginary plane that would
contain the spiral if it were a ring in a plane and not helical, but distorting the circumference
that defines it. The fourth (Figure 8d) and sixth (Figure 8f) modes constitute double torsion
(with one inflection point) and triple torsion (with two inflection points), respectively.
If we continue to search for natural modes at increasing frequencies, increasingly com-
plex deformation profiles with small displacements and numerous inflection points appear.

323
Mathematics 2022, 10, 2698

(a) (b)

(c) (d)

(e) (f)

Figure 8. Modes of vibration of a spring: (a) Bending 1; (b) Torsional 1; (c) Bending 2; (d) Torsional 2;
(e) Bending 3; (f) Torsional 3.

With respect to the results shown in Table 2, it can be deduced that, in the case of
the first resonance mode (bending), keeping the helix diameter constant and varying the
wire thickness, the resonance frequency increases with increasing thickness. This is logical
because the stiffness of the spring increases. Figure 9 shows this trend. Moreover, as
shown in Figure 9, as the diameter of the helix increases, the resonance frequency decreases.
Despite maintaining the stiffness, the loop flexes with lower frequency values. A similar
situation occurs with the first torsional mode.

324
Mathematics 2022, 10, 2698

ϮϬϬϬ

ϭϱϬϬ
ĞŶĚŝŶŐϭсϴϬŵŵ
ĞŶĚŝŶŐϭсϭϬϴŵŵ
ϭϬϬϬ
&ƌĞƋƵĞŶĐLJ;,njͿ ĞŶĚŝŶŐϭсϮϮϬŵŵ
dŽƌƐŝŽŶĂůϭсϴϬŵŵ
ϱϬϬ
dŽƌƐŝŽŶĂůϭсϭϬϴŵŵ
dŽƌƐŝŽŶĂůϭсϮϮϬŵŵ
Ϭ
ϭϱ ϮϬ Ϯϱ ϯϬ
tŝƌĞĚŝĂŵĞƚĞƌĚ;ŵŵͿ

Figure 9. Sensitivity of the first resonance modes to the wire diameter, d.

Proceeding analogously for the first bending mode, the wire thickness is now kept
constant, and the average diameter of the helix is varied (Figure 10). The frequencies
decrease as the helix diameter increases. Moreover, as the wire diameter increases, the
resonant frequency increases. The situation is repeated for the first torsional mode of
vibration. It is a situation analogous to that of the previous paragraph.

ϮϬϬϬ
ϭϴϬϬ
ϭϲϬϬ
&ƌĞƋƵĞŶĐLJ;,njͿ

ϭϰϬϬ ĞŶĚŝŶŐϭĚсϭϴŵŵ
ϭϮϬϬ ĞŶĚŝŶŐϭĚсϮϮŵŵ
ϭϬϬϬ
ĞŶĚŝŶŐϭĚсϮϴŵŵ
ϴϬϬ
ϲϬϬ dŽƌƐŝŽŶĂůϭĚсϭϴŵŵ
ϰϬϬ dŽƌƐŝŽŶĂůϭĚсϮϮŵŵ
ϮϬϬ
dŽƌƐŝŽŶĂůϭĚсϮϴŵŵ
Ϭ
Ϭ ϱϬ ϭϬϬ ϭϱϬ ϮϬϬ ϮϱϬ
,ĞůŝdžŝĂŵĞƚĞƌ;ŵŵͿ

Figure 10. Sensitivity of the first resonance modes to the helix diameter, D.

Analyzing the variation of the frequency as a function of the variation of the average
diameter of the helix, it is observed (Figure 10) that the first bending and torsional mode
of the spring decrease as the mass increases due to a higher helix diameter. This has to
do with the wire diameter/helix diameter ratio and it can be seen that there is a certain
characteristic value of this ratio that linearizes the resonant behavior.
As shown in Figure 9 the relationship between frequencies and wire diameter (d) is
linear. This result is consistent with Equation (3).
It has also been considered whether the pitch of the spring can be an influential design
parameter when calculating resonances. Figure 11 shows that the pitch variation does not
affect the frequencies. Figure 11 shows the different modes obtained and commented.

325
Mathematics 2022, 10, 2698

ϱϬϬϬ
ϰϱϬϬ
ϰϬϬϬ

&ƌĞĐƵĞŶĐLJ;,njͿ
ϯϱϬϬ ĞŶĚŝŶŐϭ
ϯϬϬϬ dŽƌƐŝŽŶĂůϭ
ϮϱϬϬ
ϮϬϬϬ ĞŶĚŝŶŐϮ
ϭϱϬϬ dŽƌƐŝŽŶĂůϮ
ϭϬϬϬ
ĞŶĚŝŶŐϯ
ϱϬϬ
Ϭ dŽƌƐŝŽŶĂůϯ
ϮϬ ϯϬ ϰϬ ϱϬ ϲϬ
WŝƚĐŚ;ŵŵͿ

Figure 11. Sensitivity of the first resonance modes to the pitch of the helix.

Table 3 includes a comparison for the first harmonic of the results obtained with the
finite element simulation and the numerical model presented in this paper. The results of
the numerical model are obtained by solving ω in Equation (21) with the wavenumber k
calculated with Equation (22b) or Equation (25). In general, it is observed that the numerical
model overestimates the frequency values with an average difference of 3%.

Table 3. Comparison of the numerical model with FEM for the first harmonic (Hz).

Fundamental Frecuency (Hz)

D (mm) d (mm) Pitch (mm) Wavenumber k (m−1 ) Bending 1 FEM Bending 1 Numerical Model Equation (21)
80 18 39.8 24.69 960 989
80 22 39.8 24.69 1168 1190
80 28 39.8 24.69 1475 1495
108 18 39.8 18.39 539 555
108 22 25 18.47 667 680
108 22 39.8 18.39 657 670
108 22 50 18.32 648 665
108 28 39.8 18.39 832 840
220 18 39.8 9.08 133 145
220 22 39.8 9.08 162 170
220 28 39.8 9.08 207 211
N = 1 loop, E = 210 GPa, G = 84 GPa, ρ = 7800 kg/m3 , ν = 0.24407.

4. Conclusions
The study of a single coil has been proposed since, in general, the authors who
propose equations to model the vibration behavior of a spring consider the spring as a
whole, including its ends, but do not analyze the behavior of the coil itself. From the
analysis of a single coil, the singular behavior of the spring known as “effort de chasse”
described in the introduction could be justified.
For the study of the vibratory behavior of a spring, an analytical model has been
developed based on a reference system consisting of a Frenet trihedron whose tangent
vector is the midline of the coil. One of the results obtained from the model is the dispersion
relation that relates the wavelength to its oscillation frequency, ω = f (k ). Applying the
appropriate boundary conditions the dispersion relation allows us to calculate the normal
modes of oscillation.
#
Another of the results obtained is that there is a value of the wavenumber
k = Ω21 + Ω23 for which the roots of the dispersion equation cancel out. Being Ω1
and Ω3 the torsional curvature and the bending curvature, respectively.
Two clear modes of oscillation have been characterized, one in which the coil moves
according to the torsion of the cross-section increasing the pitch of the coil, and another

326
Mathematics 2022, 10, 2698

bending mode in which it moves by making an opening of its ends so that the original helix
is transformed into an ellipsoidal helix.
A spring has been modeled by finite elements with ANSYS® software in order to
contrast results with the numerical model, obtaining an average difference of 3% in the
fundamental oscillation frequency.

Author Contributions: Individual contributions are as follows: original concept: L.P., A.Q., A.M.G.A.
and V.D.; mathematical development: L.P. and A.M.G.A.; implementation, simulation and draft
preparation: L.P., A.M.G.A., A.Q. and V.D.; review, editing, validation, and formal analysis: L.P., A.Q.,
A.M.G.A. and V.D. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A
In this appendix, the expressions for the calculation of the coefficients of Equation (21)
are indicated. Equation (A1a–g) contains the expressions for the calculation of mi

m1 = φ1 ·φ5 (A1a)

m2 = φ1 ·φ6 + φ2 ·φ5 + φ9 (A1b)

m3 = φ1 ·φ7 + φ2 ·φ6 + φ3 ·φ5 + φ10 (A1c)
m4 = φ1 ·φ8 + φ2 ·φ7 + φ3 ·φ6 + φ4 ·φ5 + φ11 (A1d)
m5 = φ2 ·φ8 + φ3 ·φ7 + φ4 ·φ6 + φ12 (A1e)
m6 = φ3 ·φ8 + φ4 ·φ7 + φ13 (A1f)
m7 = φ4 ·φ8 (A1g)
With Equation (A2a–m) for the calculation of the parameters of the φi

φ1 = λ1 λ2 λ3 (A2a)

φ2 = −[λ2 λ3 a4 + λ1 λ2 a6 + λ1 λ3 a5 ] (A2b)
φ3 = {λ1 ( a5 a6 − c3 ) + λ2 ( a4 a6 − c5 ) + λ3 ( a4 a5 − c4 )} (A2c)
φ4 = c4 a6 + c5 a5 − a4 ( a5 a6 − c3 ) + b5 (A2d)
φ5 = λ 3
G1∗ G2∗ E∗ (A2e)
φ6 = −λ2 { G1∗ G2∗ a3 + E∗ ( G1∗ a1 + G2∗ a2 )} (A2f)
φ7 = λ{ a3 [ G1∗ a1 + G2∗ a2 ] + E∗ [ a2 a1 − c6 ] + G2∗ c2 − G1∗ c1 } (A2g)
φ8 = − a3 ( a1 a2 − c6 ) + c1 a2 − c2 a1 + b6 (A2h)
3 ∗
φ9 = −λ E λ1 (λ3 G1∗ + λ2 G2∗ ) (A2i)
$ %
φ10 = λ2 { λ 1 λ3 E∗ a ∗ ∗ ∗ ∗ ∗
2 + G1 a3 + λE G1 ( λ1 a6 + λ3 a4 ) + λ1 λ2 ( E a1 + G2 a3 )
(A2j)
+λE G2∗ (λ2 a4 + λ1 a5 ) + λλ1 E∗ }
∗

327
Mathematics 2022, 10, 2698

 
φ11 = λ{$λ2 E∗ G1∗ (c5 −% a4 a6 ) + G2∗ (c4 − a4 a5 )
−λ[ E∗ a2 + G1∗ a3 (λ1 a6 + λ3 a4 )
+( E∗ a1 + G2∗ a3 )(λ2 a4 + λ1 a5 )] (A2k)
−λ1 [λ2 ( a1 a3 − c1 ) + λ3 ( a2 a3 + c2 )]
+λ[λ1 ( E∗ b1 − a3 ) − E∗ a4 ]}
φ12 = λ{( a2 a3 + c2 )(λ1 a6 + λ3 a4 ) + (λ2 a4 + λ1 a5 )( a1 a3 − c1 )
−λ1 (b3 + a3 b1 ) $ % (A2l)
−λ[( E∗ a1 + G2∗ a3 )(c4 − a4 a5 ) + E∗ a2 + G1∗ a3 (c5 − a4 a6 )
−λa3 a4 + λE∗ ( a4 b1 + b2 )]}
φ13 = λ[( a2 a3 + c2 )(c5 − a4 a6 ) + ( a1 a3 − c1 )(c4 − a4 a5 ) + a3 (b1 a4 + b2 ) + a4 b3 + b4 ] (A2m)

With Equation (A3a–f) for the calculation of the parameters of the ai

a1 = k 2 + β 7 (A3a)

a2 = k 2 + β 8 (A3b)
a3 = k + β 9
2
(A3c)
a4 = k2 I1∗ + α4 (A3d)
a5 = k2 I2∗ + α5 (A3e)
a6 = k2 I3∗ + α6 (A3f)
With Equation (A4a–f) for the calculation of the parameters of the bi
!
b1 = 2 β26 I1∗ − 2k2 β 1 ( I2∗ + I3∗ ) (A4a)
! !
b2 = 2β 5 I2∗ β 4 β 6 I3∗ + 2k2 β 5 ( I1∗ + I2∗ ) − 2k2 α1 α2 β 6 − 2β24 I3∗ ( I1∗ + I3∗ ) (A4b)
! !
b3 = 2β 5 β 4 β 6 I1∗ + 2k2 β 5 ( I2∗ + I3∗ ) − 4k2 2β26 I1∗ − β24 ( I2∗ + I3∗ ) (A4c)
$ %$ %
b4 = 2 β 4 β 5 I2∗ I$3∗ − k2 α1%α2 β 4 β 5 − 4k2 β 6
−2k2 [ β$5 I2∗ I1∗ + ∗
% I2 ( β 5 β 3 + β 6 β 4 ) (A4d)
+ β 4 I3∗ I1∗ + I3∗ ( β 5 β 6 + β 2 β 4 )]
 
b5 = 2β 5 I2∗ β 6 β 4 I1∗ I3∗ − k2 α1 α3 − 2k2 α2 (α1 β 6 I1∗ + α3 β 4 I3∗ ) (A4e)
 
b6 = 2β 5 β 4 β 6 − 8k2 β24 + β25 + β26 (A4f)

With Equation (A5a–f) for the calculation of the parameters of the ci

c1 = β 1 β 3 + 4k2 β 2 (A5a)
 
c2 = − 4k2 β 3 + β24 (A5b)

c3 = β26 I1∗2 + k2 α23 (A5c)

c4 = β24 I3∗2 + k2 α21 (A5d)
c5 = β25 I2∗2 + k2 α22 (A5e)
c6 = β26 + 4k β 1
2
(A5f)
where Equation (A6a–i) indicates the calculation of the parameters β i

β 1 = Ω21 (A6a)

328
Mathematics 2022, 10, 2698

β 2 = Ω22 (A6b)
β3 = Ω23 (A6c)
β 4 = Ω1 Ω2 (A6d)
β 5 = Ω1 Ω3 (A6e)
β 6 = Ω2 Ω3 (A6f)
β7 = β1 + β2 (A6g)
β8 = β1 + β3 (A6h)
β9 = β2 + β3 (A6i)
where Equation (A7a–f) indicates the calculation of the parameters αi

α1 = Ω3 ( I1∗ + I2∗ ) (A7a)

α2 = Ω2 ( I1∗ + I3∗ ) (A7b)

α3 = Ω1 ( I2∗ + I3∗ ) (A7c)
α4 = β 3 I2∗ + β 2 I3∗ (A7d)
α5 = β 3 I1∗ + β 1 I3∗ (A7e)
α6 = β 1 I2∗ + β 2 I1∗ (A7f)
where Equation (A8a–c) indicates the calculation of the parameters λi

λI1
λ1 = (A8a)
A
λI2
λ2 = (A8b)
A
λI3
λ3 = (A8c)
A

References
1. Wahl, A.M. Mechanical Spring, 2nd ed.; McGraw-Hill: New York, NY, USA, 1963.
2. Shigley, J.E. Mechanical Engineering Design, 10th ed.; McGraw-Hill: New York, NY, USA, 2015.
3. Kobelev, V. Durability of Springs; Springer: Berlin/Heidelberg, Germany, 2018.
4. Love, A.E.H. A Treatise on the Mathematical Theory of Elasticity, 4th ed.; Dover Publications: Mineola, NY, USA, 1927.
5. Stokes, V.K. On the dynamic radial expansion of helical springs due to longitudinal impact. J. Sound Vib. 1974, 35, 77–99.
[CrossRef]
6. Gironnet, B.; Louradour, G. Comportement Dynamique des Resorts; Techniques de l’Ingénieur: Paris, France, 1983.
7. Yldirim, V. An efficient numerical method for predicting the natural frequencies of cylindrical helical springs. Int. J. Mech. Sci.
1999, 41, 919–939. [CrossRef]
8. Lee, J.; Thompson, D.J. Dynamic stiffness formulation, free vibration and wave motion of helical springs. J. Sound Vib. 2001, 239,
297–320. [CrossRef]
9. Lee, L. Free vibration analysis of cylindrical helical springs by the pseudo spectral method. J. Sound Vib. 2007, 302, 185–196.
[CrossRef]
10. Becker, L.E.; Chassie, G.G.; Cleghorn, W.L. On the natural frequencies of helical compression springs. Int. J. Mech. Sci. 2002, 44,
825–841. [CrossRef]
11. Jiang, W.; Wang, T.L.; Jones, W.K. The forced vibration of helical spring. Int. J. Mech. Sci. 1992, 34, 549–562. [CrossRef]
12. Haringx, J.A. On highly compressible helical springs and rubber rods, and their application for vibration-free mountings. Philips
Res. Rep. 1949, 4, 49–80.
13. Den Hartog, J.P. Mechanical Vibrations; Dover Civil and Mechanical Engineering; Courier Corporation: Chelmsford, MA,
USA, 1956.
14. Díaz, V. Automóviles y Ferrocarriles; Universidad Nacional de Educación a Distancia: Madrid, Spain, 2012; p. 287,
ISBN 978-84-362-6568-2.

329
Mathematics 2022, 10, 2698

15. Sauvage, G. Determining the Characteristics of Helican Springs: A simplification for Application in Railway Suspensions. Veh.
Syst. Dyn. Int. J. Veh. Mech. Mobil. 1984, 13, 43–59.
16. Campedelli, J. Modelisation Globale Statique des Systemes Mecaniques Hyperstatiques Pre-Charges Application a un Bogie
Moteur. Ph.D. Thesis, INSA Lyon, Lyon, France, 2002.
17. UNE-EN 13298; Railway Applications. Suspension Components. Helical Suspension Springs, Steel. The European Committee for
Standardisation: Brussels, Belgium, 2003.
18. Kovelev, V. Effect of static axial compression on the natural frequencies of helical springs. Multidiscip. Model. Mater. Struct. 2014,
10, 379. [CrossRef]
19. Yildirim, V. Axial Static Load Dependence Free Vibration Analysis of Helical Springs Based on the Theory of Spatially Curved
Barx. Lat. Am. J. Solids Struct. 2016, 13, 2852–2875. [CrossRef]
20. Kato, H.; Suzuki, H. Nonlinear deflection analysis of helical spring in elastic-perfect plastic material: Application to the plastic
extension of piano wire spring. Mech. Mater. 2021, 160, 103971. [CrossRef]
21. Timoshenko, S.P. On the correction for shear of the differential equation for transverse vibrations of prismatic bars. Philos. Mag.
1921, 41, 744–746. [CrossRef]
22. Rayleigh, J.W.S. The Theory of Sound; Dover Publications: Mineola, NY, USA, 1945.
23. Pearson, D. Modelling the ends of compression helical springs for vibration calculations. Proc. Inst. Mech. Eng. 1986, 200, 3–11.
[CrossRef]

330
 mathematics

Article
Fault Classification in a Reciprocating Compressor and
a Centrifugal Pump Using Non-Linear Entropy Features
Ruben Medina 1, *, Mariela Cerrada 2 , Shuai Yang 3 , Diego Cabrera 2 , Edgar Estupiñan 4 and René-Vinicio Sánchez
2, *

1 CIBYTEL-Engineering School, University of Los Andes, Mérida 5101, Venezuela

2 GIDTEC, Universidad Politécnica Salesiana, Cuenca 010105, Ecuador
3 National Research Base of Intelligent Manufacturing Service, Chongqing Technology and Business University,
19# Xuefu Avenue, Nan’an District, Chongqing 400067, China
4 Mechanical Engineering Department, University of Tarapaca, Arica 1130000, Chile
* Correspondence: ruben.djmedina@ieee.org (R.M.); rsanchezl@ups.edu.ec (R.-V.S.)

Abstract: This paper describes a comparison of three types of feature sets. The feature sets were
intended to classify 13 faults in a centrifugal pump (CP) and 17 valve faults in a reciprocating com-
pressor (RC). The first set comprised 14 non-linear entropy-based features, the second comprised
15 information-based entropy features, and the third comprised 12 statistical features. The clas-
sification was performed using random forest (RF) models and support vector machines (SVM).
The experimental work showed that the combination of information-based features with non-linear
entropy-based features provides a statistically significant accuracy higher than the accuracy provided
by the Statistical Features set. Results for classifying the 13 conditions in the CP using non-linear
entropy features showed accuracies of up to 99.50%. The same feature set provided a classification
accuracy of 97.50% for the classification of the 17 conditions in the RC.

Citation: Medina, R.; Cerrada, M.; Keywords: approximate entropy; non-linear systems; phase space reconstruction; fault classification;
Yang, S.; Cabrera, D.; Estupiñan, E.; random forest; support vector machines
Sánchez, R.-V. Fault Classification in
a Reciprocating Compressor and MSC: 28D20
a Centrifugal Pump Using
Non-Linear Entropy Features.
Mathematics 2022, 10, 3033. https://
doi.org/10.3390/math10173033 1. Introduction
Academic Editor: Andrey Jivkov In industrial applications, centrifugal pumps are equipment for transferring energy to
fluids to enable pipeline transportation. Similarly, reciprocating compressors are essential
Received: 30 June 2022
in petrochemical plants and refineries for gas and other fluid transportation. Reliable
Accepted: 17 August 2022
functioning of these types of equipment is essential in industry to avoid unexpected
Published: 23 August 2022
halting of processes and economic losses. Consequently, several condition monitoring (CM)
Publisher’s Note: MDPI stays neutral methods have been proposed for diagnosing faults in centrifugal pumps and reciprocating
with regard to jurisdictional claims in compressors [1–3].
published maps and institutional affil- The CM methods are based on sensing several variables from the target equipment,
iations. such as electrical current, vibration, sound, temperature, or acoustic emission. Some of
the most common variables are vibration signals recorded using accelerometer sensors.
Such a set of signals is processed for extracting features useful for fault detection and
classification. Traditional techniques for vibration signal analysis are divided into time,
Copyright: © 2022 by the authors.
frequency, and time-frequency methods. In general, signal processing techniques assume
Licensee MDPI, Basel, Switzerland.
that signals represent linear systems that comply with the stationary and periodicity
This article is an open access article
condition. Unfortunately, this assumption is not always met. In particular, in most rotatory
distributed under the terms and
machines, the vibration signals represent a non-linear system [4]. These constraints can be
conditions of the Creative Commons
Attribution (CC BY) license (https://
efficiently handled using chaos theory and non-linear dynamics techniques.
creativecommons.org/licenses/by/
In this work, we propose to use the approximate entropy (AppEn) [5,6] and variants,
4.0/).
such as the sample entropy (SampEn) [7,8] and fuzzy entropy (FuzzyEn) [9,10] for fault

Mathematics 2022, 10, 3033. https://doi.org/10.3390/math10173033 331 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 3033

classification. The emphasis of this research is feature extraction. The classification stage
can be performed using either classical or deep learning-based classifiers. However, as the
feature extraction is computationally expensive, selecting deep learning models would
impose an unnecessary cost. In consequence, we have selected two efficient classical models
corresponding to RF and the multi-class SVM [11,12]. The original contributions of this
research are the following:
• Extraction of a non-linear entropy-based feature set that provides high classification
accuracy using RF and SVM models. The accuracy attained by the SVM model trained
with the non-linear Entropy Features set was higher than the accuracy attained by a
CNN model trained with 2D spectrogram images for both the CP and the RC.
• Detailed comparison of three feature sets useful for classifying a large number of faults
in a CP and an RC. These are the non-linear Entropy Features, the Information Entropy,
and the Statistical Features sets. The first set is composed of the approximate entropy
and several variants, and the second set is composed of the combination of the wavelet
packet transform (WPT)-based features and the power spectrum entropy (PSE)-based
features. Finally, the third set is composed of classical time series statistical features.
• The non-linear Entropy Features set and the All Features set corresponding to the fusion
of the three feature sets when compared provided a classification accuracy of up to
99.59% for the CP and up to 97.90% for the RC. For the CP, 13 different conditions
were classified, and 17 valve conditions were classified for the RC.
The paper is organized as follows: in Section 3 the theoretical background is presented.
In Section 4, the test bed for acquiring the vibration signals from the CP and RP is presented.
In Sections 5 and 6, the methodology for extraction of the different sets of features is
described. The methodology for classifying faults using RF and SVM models based on
the extracted features is presented in Section 7. The results are described in Section 8, and
finally, the conclusion and future work are presented in Section 10.

2. Related Research
The related research is discussed in the following subsections. A compendium with
several publications is presented in Table A1 in terms of their main features. A brief revision
concerning these publications is presented in Sections 2.1–2.3.

2.1. Rotating Machinery

Although non-linear entropy methods have been used for classifying a small set of
faults in centrifugal pumps and reciprocating compressors, attaining high accuracy, their
application to other rotatory machines (RM) has also been reported. Research concerning
the correlation dimension (CD) is presented in [13]. The approach analyzes the vibration
signal measured from a roller element bearing. Results allow them to conclude that CD
features show significant statistical differences between a roller element close to failure and
a new one. The largest Lyapunov exponent (LLE) [14] has been used to show that faults
in non-linear rotatory systems can be detected early when the onset of the fault occurs.
This feature evolves as the fault progresses. The research reported in [4] combines the CD,
the approximate entropy, the LLE, and visualization of the phase space for detecting damage
in bearings and gearboxes. In [15], a method for classifying seven different conditions
in roller bearings and ten types of conditions on a gearbox is reported. The approach is
based on features extraction from the Poincaré plot and their classification using SVM.
The accuracy attained was 100% for the roller bearings and 99.3% for the gearbox dataset.
A combination of the spectral graph wavelet transform (SGWT) with detrended fluctuation
analysis (DFA) was used for denoising the vibration signals in RM [16]. The first stage is
converting the vibration signal into a graph signal. The SGWT is used for decomposing the
graph signal into scaling function coefficients and spectral graph wavelet coefficients. DFA
is used for selecting the level of decomposition of SGWT. A bi-dimensional representation
of vibration signals denoted as the complexity-entropy causality plane (CECP) useful
for fault classification in RM is reported in [17]. This 2D representation was obtained

332
Mathematics 2022, 10, 3033

with the normalized permutation entropy and the Jensen–Shannon complexity of time
series. The authors showed the utility of this representation for improving the classification
accuracy of faults in roller bearing and gear applications using SVM models. A review
of entropy algorithms and its variants in RM is reported in [18,19]. The reviews briefly
introduced several entropy methods and their application in fault detection in RM.

2.2. Reciprocating Compressor

An application of faults diagnosis in an RC using multi-scale entropy is reported in [20].
After performing the signals’ local mean decomposition (LMD), vibration signals’ multi-
scale entropy (MSE) features were calculated. Four conditions were classified using SVM,
obtaining a classification accuracy of 95.0%. In a previous study [2], a method for feature
extraction from vibration signals recorded from an RC is proposed to classify 13 different
conditions of combined faults in valves and roller bearings. The feature extraction method
was based on symbolic dynamics and complex correlation measures. The extracted features
were used for fault classification using several machine learning models, attaining accuracy
of up to 99%. Research concerning the fault diagnosis and severity of valve leakage in an RC
is reported in [21]. The diagnosis was performed using features extracted from the p − V
diagram using principal component analysis (PCA) and linear discriminant analysis (LDA).
The method showed that the p − V diagram’s features help classify two types of faults with
several levels of severity, attaining classification accuracies of up to 99.62%. The research
reported in [22] concerns the classification of faults based on images representing the
p − V diagram. A set of features is extracted from the images for classifying eight health
conditions in valves, piston rings, and valve springs. Accuracies of up to 97.9% were
obtained using artificial neural network (ANN) models. Research reported in [23] combines
the binary moth flame optimization (BMFO) algorithm with the K-nearest neighbor (KNN)
algorithm for classifying three types of valve faults in an RC, attaining accuracies higher
than 99%. Research concerning convolutional neural networks (CNNs) for classifying
faults in an RC is reported in [24]. The vibration signal was re-arranged into a 2D format
by bringing its time samples into lines that moved 45 degrees counterclockwise to reinforce
their features. This array was fed to the CNN that incorporated the attention mechanism
through squeeze-and-excitation (SE) modules. The maximum accuracy attained was 99.4%.
Classification of four different valve conditions using 1D and 2D CNNs is reported in [25].
The classification was performed using vibration signals in the case of the 1D CNN and
considering the fusion of seven channels (four channels with vibration signals and three
channels with pressure) in the case of 2D CNN. The classification was performed for
different SNR values of the signals, and the highest classification accuracy was 100% for
signals with more than 20 dB of SNR. The CNN used has a complex architecture.

2.3. Centrifugal Pump

The permutation entropy features are used in a fault detection method reported in [26].
The method was applied to hydraulic pump (HYP) fault detection. The vibration signal
was processed using resonance-based sparse signal decomposition (RSDD). A Multi-scale
hierarchical amplitude-aware permutation entropy (MHAAPE) method was applied for
feature extraction. Four fault conditions were classified using a probabilistic neural network,
SVM models, and extreme learning machines (ELM). The approach attained an accuracy
of up to 100%. A method for faults diagnosis in a CP is reported in [27]. The method is
based on applying complementary ensemble empirical mode decomposition (CEEMD) to
vibration signals. The decomposed signals or intrinsic mode functions are processed for
extracting the SampEn. The set of features is used for classification; RF attained classification
accuracies of 94.58% for classifying five conditions. A feature extraction approach from
vibration signals is reported [28] to classify faults in a CP. The discriminant feature extraction
method includes three stages. The first stage is the healthy baseline signal selection.
The second stage is devoted to cross-correlating the healthy baseline signals with the
vibration signals of different fault classes. A set of features is extracted from the resultant

333
Mathematics 2022, 10, 3033

correlation sequence. In the third stage, time, frequency, and time–frequency domain raw
hybrid features are extracted from baseline signal and vibration signals from faulty classes.
A new set of discriminant features is composed of the correlation coefficient between raw
hybrid feature pools. The combination of all extracted features is fed to an SVM model for
classification. The method is used to classify four conditions in a CP, attaining accuracy
of up to 98.4%. A method for fault classification in a CP based on vibration signals is
reported in [29]. A feature extraction pre-processing is used to obtain a time-frequency
representation using the continuous wavelet transform (CWT). The CWT is converted
to gray images fed to a CNN model with an adaptive learning rate. Accuracies of 100%
are obtained for classifying four different fault conditions. A fault classification method
of a multi-stage CP reported in [30] starts by selecting the fault-specific frequency band
and continues by extracting statistical features in time, frequency, and wavelet domains
from this band. The next step is to select a low dimensional features vector using the
informative ratio PCA. The classification of four conditions in a multi-stage CP using KNN
models. The accuracy attained was 100%. Fault detection and classification for water
pump bearings based on features extracted from the instantaneous power spectrum is
reported in [31]. The instantaneous power spectrum (IPS) was obtained with the voltage
and current measured, and several features were extracted from the IPS for classifying
three different conditions at different load levels. The classification was performed using
an extreme gradient boosting (XBG) model [32]. The application of CNN for classifying
faults in a CP based on acoustic images is reported in [33]. In this application, the sound
signals were acquired from a CP test rig where five different health conditions were
implemented. The sound signals were converted into acoustic images using the analytical
wavelet transform (AWT). The acoustic images were fed to the CNN for classification.
The attained accuracy was 100%.

2.4. Anomaly Detection

Anomaly detection corresponds to identifying events within a stream of data features
that deviate significantly from most available data. Anomaly detection has applications
in many fields. This topic is relevant because the non-linear entropy-based features have
found applications in the context of anomaly detection. Examples of applications have
been reported in [34–39]. Several authors have proposed anomaly detection applications
centered around deep learning. Approaches such as CNN, autoencoders, and generative
adversarial networks [40–42] can be used for anomaly detection implementation. Research
concerning anomaly detection in RM is active. An example is reported in [41] for anomaly
detection in gearboxes where a prototype selection algorithm is used for improving the
training of the feature extractor implemented with a CNN. The research reported in [40]
uses an autoencoder for learning to model the healthy state of the rotating machine. Such
a learning process is performed using vibration signals representing the healthy state.
The anomaly is detected using a threshold-based approach to the reconstruction error of
data acquired from an unknown state. The algorithm was validated with the IMS bearings
dataset. Other recent applications of anomaly detection in RM and also in other fields are
reported in [41,43–46].

3. Theoretical Background
3.1. Phase Space Reconstruction
The phase space reconstruction method is based on the embedding theorem of Taken,
reported in [47]. The theorem postulates that we can recover the equivalent dynamics of
a non-linear system by using time delays from a recorded time-domain signal. The ap-
proximate dynamics correspond to a 1D projection of the system trajectory. The theorem is
mainly applied to univariate time series [47]; however, the multivariate version has also
been reported in [48]. In this research, we are mainly concerned with univariate embedding.
The average mutual information (AMI) enables the selection of the time delay or lag
denoted τ [49]. The AMI is plotted versus τ. According to [50], the selection of τ could be

334
Mathematics 2022, 10, 3033

the location of the first local minimum of the AMI. The system’s embedding dimension,
denoted as m, could be selected by applying the false nearest neighbors (FNN) method, as
reported in [51].

3.2. Approximate Entropy

Given a signal denoted x including N samples, the AppEn [5,6] measures the reg-
ularity of the time-series. The time-series regularity and dynamics are estimated con-
sidering a higher order dimensional space generated by a series of vectors represented
by time-delayed components of x. The AppEn is calculated for a time-series of length
N, representing a system with embedding dimension m, lag τ, and a distance threshold
between vectors denoted r. In general, the calculation of the AppEn should consider a
window of the signal with length N large enough to reduce the bias in the estimation [6].
However, a limitation is that extending N can significantly increase the computational
cost. AppEn has the advantage of being robust to noise and outliers (see [52] for a detailed
review concerning entropy calculation methods).

3.3. Sample Entropy

The SampEn is obtained through modification of the AppEn algorithm [7]. The mod-
ification consists of the elimination of self-matches in the evaluation of the distance be-
tween lagged vectors. The SampEn avoids the bias introduced in ApEn by considering
self-matches. The resultant features are more consistent and unbiased, even when their
calculation uses short-length signal windows. In addition, the calculation of SampEn is
more straightforward than the calculation of AppEn.

3.4. Fuzzy Entropy

The FuzzyEn measure [10] is a variant of the AppEn that modifies the function used
for estimating the similarity between lagged vectors. The algorithm’s modification is based
on changing the Heaviside function for a fuzzy exponential membership function. The cal-
culation of FuzzyEn also incorporates the modifications introduced for the calculation of
the SampEn.

3.5. Shannon Entropy and Other Measures of Signal Complexity

The entropy measure enables quantification of the complexity of a non-linear dy-
namical system. The Shannon entropy (ShannEn) quantifies the amount of information
in a system. In addition, the conditional entropy quantifies the rate of information gen-
eration [53,54]. Other entropy measures, such as corrected conditional entropy, can be
estimated based on these measures. A detailed review of ShannEn estimation is presented
in [55].

3.6. Permutation Entropy

The permutation entropy (PermEn) is based in partitions [56]. When the time series x
takes many values, their range can be replaced with a symbol sequence {si }. This is done
by defining the partition of the signal range x = P1 ∪, . . . , ∪ Pξ . Based on this partition, the
symbol si = j is defined if xi is in Pj . The PermEn is calculated for different embedding
dimensions m; however, for practical purposes m = 3, 4, . . . , 7. Denote by π the permutation
of order m with probability given as p(π ) = #{Cardinality o f permutation π }/N − m + 1;
then, the PermEn is calculated using Equation (1):

PermEn = − ∑ p(π )logp(π ) (1)

This research used the method reported in [57] for computing the PermEn. The method
is based on pre-calculating values for successive ordinal patterns of order m. This ap-
proach allows the calculation of PermEn in large time series using overlapped successive
time windows.

335
Mathematics 2022, 10, 3033

3.7. The Correlation Dimension

The system’s phase space enables the calculation of complexity [58] that is quantified
using several non-linear features such as SampEn. In addition, another feature is the
CD which quantifies the number of independent lagged vectors required for defining
a system [59]. Details concerning the estimation of this feature are presented in [60].

3.8. Detrended Fluctuation Analysis

The fractal scaling properties quantified in short intervals of time-series [61] can be
performed using DFA. This approach enables quantitative analysis of non-stationary time-
series [62]. Their calculation requires performing the integration of the signal, subdividing
such a signal into several segments, finding the trend of the signal using least-square
approximations, and calculating the average fluctuation of the signal around the trend.
Details of the estimation of this feature are presented in [63].

3.9. Largest Lyapunov Exponent

The Lyapunov exponent λ quantifies the relative motion of trajectories in the phase-
space of a system [64]. Its value allows for determining if a signal is chaotic or periodic.
A system can be represented by several Lyapunov exponents determined by the embed-
ded dimension. However, the largest one, denoted LLE, is typically used. There are
several algorithms for calculating the Lyapunov exponent [65]. The method proposed by
Rosenstein et al. [66] was implemented in this research.

3.10. Information Entropy and Chaotic Time Series

Information entropy was proposed by Claude Shannon [67] in the context of theoret-
ical communication modeling. A collection of messages sent through a communication
channel can be considered as independent random events with probabilities pi ≥ 0, where
∑in=1 pi = 1. Information entropy is defined as:
n
H = − ∑ pi log pi (2)
i =1

The entropy concept is also used in classical thermodynamics in the context of heat
transfer [68]. Gibbs and Boltzmann attained a definition of entropy with an expression
similar to Equation (2) [69]. When dealing with discrete-time signals represented as a vector,
the entropy concept has also been applied to select the best basis for orthogonal wavelet
packets [70]. The authors showed that the assumption that the Karhunen–Loeve basis is
the best, even for a single vector which enables the calculation of several entropy features.
Consider the time-series x where the vector can be arranged in decreasing order such that
xi = 0 for large i. Then, pi = | xi |2 x−2 . Without loss of generality, the time-series can
be normalized so x−2 = 1. The following features can be estimated: the log energy
entropy, the Shannon information entropy, the norm entropy, the threshold entropy, and the
“sure” entropy.

3.10.1. Tsallis and Rényi Entropy

In the context of systems with complex behavior (including, for instance, chaotic
motion or fractal evolution), Tsallis [71] has proposed an alternative definition of entropy as:
 
N
K
1 − ∑ pi
q
HTq = (3)
q−1 i =1

336
Mathematics 2022, 10, 3033

where K is a positive constant and q denotes any real number. Tsallis showed that when
q → 1, HTq tends to the classical Boltzmann–Gibbs–Shannon entropy represented by
Equation (2). Similarly, q is used in the Rényi entropy [72] to amplify probabilities as:
 
N
K
log ∑ pi
q
HRq = (4)
q−1 i =1

Both entropies are parameterized representations of the Boltzmann–Gibbs–Shannon

entropy, where one parameter is used for weighting the probabilities to evaluate different
fractal properties of the system.

3.10.2. Power Spectral Entropy

The complexity can also be quantified using the power spectral density of time-
series [73,74]. This feature has been used for fault detection in gearboxes and roller bearings.
The normalized power spectral density of the time series x is denoted P(ωi ), and it can be
expressed as:
P ( ωi )
pi = . (5)
∑ i P ( ωi )
This equation enables the calculation of the power spectral entropy (PSE) as:
n
S = − ∑ pi log pi (6)
i =1

The PSE is calculated for each time frame and is a vector denoted as S. When the
signal has high complexity, the spectrum of the time series tends to be uniform, and the
PSE is high. In contrast, when the spectrum has a narrow peak, the PSE feature is lower,
and the signal is less complex.

3.11. Machine Learning Techniques

3.11.1. Random Forest
The RF algorithm is a machine learning algorithm that combines a set of weak learners
using ensemble methods to attain a high-quality classification or regression model [75].
The most common weak learners in the RF model are tree-type classifiers used for growing
forests of randomized trees using bootstraps resampling techniques [76].

3.11.2. Support Vector Machines

The machine learning model proposed by Vapnik [77] is the SVM. In its more basic
form, the SVM is used for solving problems of binary classification where the data in
each of the classes can be separated using a hyperplane that can be constructed based on
the support vectors, as detailed in [78]. In practical applications, the SVM incorporates
higher-dimensional transformation represented as kernels for transforming the feature
space to increase class separability. Similarly, multi-class support vector machines are also
available, as reported in [11].

4. Experimental Test-Bed
In this research, two different vibration signal datasets were considered. The first was
acquired from a CP and the second from an RC.

4.1. Centrifugal Pump Test-Bed

In this research, a multi-stage vertical CP is considered. The CP model 3SV10GE4F20
was a 2 HP pump operated at 3500 rpm. The pump had ten stages for a nominal flow
of 15 GPM. The vibration signals were recorded using model 603C01 accelerometers.
The module NI 9234 connected to a laptop computer was used to acquire the vibration signal
using a sampling frequency of 50 kHz. The physical description and locations of sensors

337
Mathematics 2022, 10, 3033

are shown in Figure 1. The signal acquisition was performed under controlled operating
conditions concerning the room temperature, relative humidity, and environmental noise.

(a) (b)
Figure 1. CP and sensors. (a) Multi-stage vertical CP. (b) Locations of sensors. The vibration signals
were recorded using A1, A2, A3, and A4 accelerometers.

Six experimental conditions were considered related to the discharge pressure. The con-
ditions are denoted as Ci with i ranging from 1 to 6. The pressure started at 5.5 bar,
increasing by 1 bar for each condition. C1 corresponds to a discharge pressure of 5.5 bar.

4.1.1. Faults of the CP

A set of 13 impeller faults were configured in the CP. One condition corresponded to
a CP where all stages were healthy. Three conditions involved pitting at the entrance of
the impeller blades (PEB), three conditions involved pitting at the output of the impeller
blades (POB), three conditions involved impeller channel blockade (ICB), and finally, three
conditions involved an imbalanced impeller (IB). The pitting and impeller blockade can
affect several stages of a pump with different levels of severity. For instance, in fault P3,
the severity level for stage 6 was 1, and the severity increased with the stage number,
and it was 5 for stage 10. The pitting severity levels were attained by changing the number
and diameter of holes in the blades. The ICB severity levels were attained by varying the
percent of blocked channels in the impeller. The impeller imbalance severity was obtained
by increasing the surface of the cut portion of the impeller front cover. Table 1 presents the
list of configured faults.

Table 1. Faults conditions in the CP.

Fault Label Impeller Fault Stages Status Severity

P1 Healthy All stages healthy healthy
P2 PEB Faulty stages 9-10 1-2
P3 PEB Faulty stages 6-10 1-5
P4 PEB Faulty stages 5-10 3-8
P5 POB Faulty stages 9-10 1-2
P6 POB Faulty stages 6-10 1-5

338
Mathematics 2022, 10, 3033

Table 1. Cont.

Fault Label Impeller Fault Stages Status Severity

P7 POB Faulty stages 5-10 3-8
P8 ICB Faulty stages 10 1
P9 ICB Faulty stages 7-10 1-4
P10 ICB Faulty stages 5-10 1-6
P11 IB Faulty stages 10 1
P12 IB Faulty stages 7-10 1-4
P13 IB Faulty stages 5-10 1-6

4.1.2. CP Vibration Signal Dataset

The vibration signal acquisition was performed by maintaining the environmental
condition within a controlled range. The motor rotation was set to 3600 rpm. A pressure
condition Ci was set, and the vibration signal was acquired during 10 s. Ten repetitions
were performed for each of the faults conditions P1 to P13. Consequently, 60 vibration
signals were recorded for each condition. A total of 780 vibration signals are included in
the dataset, including all fault conditions.

4.2. Reciprocating Compressor Test-Bed

The RC test bed is centered around a compressor model EGB-250 with two stages.
The compressor is driven by an induction motor with a power of 5.5 HP that operates at
3470 rpm.
A detailed description of the test bed is described in [2]. The test bed incorporates
several accelerometer sensors for acquiring the vibration signals. The accelerometers are
denoted as A1, A2, A3, and A4 and were located as shown in Figure 2. Module NI9234 was
used to acquire the vibration signals from model 603C01 accelerometers. The sampling
frequency was 50 kHz. The acquired vibration signals were transferred to a laptop computer.

Figure 2. Sensors’ locations in the RC. The accelerometers are denoted as A1, A2, A3, and A4.

4.2.1. Faults of the Reciprocal Compressor

The test bed enables the configuration of a set of 17 fault conditions concerning the
valves. Four faults were configured for compressor stages and the discharge and inlet
valves. The faults configured were (a) valve seat wear, (b) corrosion of the valve plate,
(c) fracture of the valve plate, and broken spring. Table 2 presents the completed list of
configured faults.

339
Mathematics 2022, 10, 3033

Table 2. Valve faults in the RC.

Fault Label Stage and Valve Type Fault Type

P1 All stages Healthy
P2 Second stage, discharge Valve seat wear
P3 Second stage, discharge Corrosion of valve plate
P4 Second stage, discharge Fracture of valve plate
P5 Second stage, discharge Broken Spring
P6 Second stage, inlet valve Valve seat wear
P7 Second stage, inlet valve Corrosion of valve plate
P8 Second stage, inlet valve Fracture of valve plate
P9 Second stage, inlet valve Broken Spring
P10 First stage, discharge Valve seat wear
P11 First stage, discharge Corrosion of valve plate
P12 First stage, discharge Fracture of valve plate
P13 First stage, discharge Broken Spring
P14 First stage, inlet valve b Valve seat wear
P15 First stage, inlet valve b Corrosion of valve plate
P16 First stage, inlet valve b Fracture of valve plate
P17 First stage, inlet valve b Broken Spring

4.2.2. RC Vibration Signal Dataset

Each vibration signal has a duration of 10 s. The motor rotated at 3462 rpm, represent-
ing a crankshaft rotation of 768 rpm. The operating conditions of the RC were kept constant
during the vibration signal acquisition. In particular, the tank pressure was kept at 3 bar,
the inlet pressure and discharge pressure were kept within the nominal range, and the
surface temperature of the valve cover was maintained. Similarly, the environmental con-
ditions corresponding to the room temperature and relative humidity were maintained
within a predefined range. A total of 15 vibration signals were acquired for each fault
condition, and a total of 255 vibration signals were acquired for each sensor.

5. Feature Extraction from the CP

5.1. Non-Linear Entropy Based Features
Each vibration signal in the CP dataset was processed for extracting a set of 29 features.
The set of features was composed of a subset of 14 features corresponding to non-linear
entropy features (Table A2) defined in Equations (A1)–(A14), a subset of 15 features cor-
responding to information-based features (Table A3) defined in Equations (A15)–(A29).
The first step consisted of performing the phase space reconstruction [48] for estimating
the spatial dimensions m and the lag τ of the system. The entropy-based features were
estimated using these parameters. Each feature set estimated in a widow was considered a
feature sample. Five non-contiguous segments of size 12,000 samples from each vibration
signal were considered for calculating features. As the number of vibration signals files
was 780, the total number of feature samples was 780 × 5 = 3900.

5.2. Statistical Features

Additionally, a set of 12 statistical features (Table A4), defined in Equations (A30)–(A41),
were extracted to compare to the entropy features set and be combined with other feature
sets to evaluate their accuracy. The procedure for extracting the statistical features was
similar to extracting the non-linear features. Three-thousand nine-hundred (3900) samples
for the CP dataset were extracted and considered for classification.

340
Mathematics 2022, 10, 3033

6. Feature Extraction from the RC

6.1. Entropy Based Features
The procedure for extracting the features from the RC was similar to extracting the
features from the CP. However, the number of segments with 12,000 samples was increased
to ten. This increment was justified because this dataset had fewer files (255). A set of
29 entropy-based features was extracted from each signal segment. In total, 2550 feature
samples were obtained for this dataset.

6.2. Statistical Features

A set of 2550 feature samples was obtained from the RC vibration signal dataset.
The procedure was similar to the case of the CP. Each feature sample comprised 12
statistical features extracted from each of the windows considered, and each signal was
subdivided into ten fragments.

7. Classification of Faults Using RF and SVM

Each feature dataset was used as input to both classical machine learning algorithms.
Each dataset was randomly split into a training set (85%) and a test set (15%). The training
set was used for 10-fold cross-validation, and the accuracy of the partitioned trained model
was calculated with the test set. This entire process (including the splitting) was repeated
ten times, and the accuracy is the average obtained during the repetitions.
The accuracy and the area under the Receiver Operator Curve (ROC) are two relevant
metrics for classifier performance. Several metrics [79] were used for evaluating the
classification results. The confusion matrix elements were used for defining the metrics [80].

8. Results
8.1. Results for the CP Dataset
The set of fault conditions for the CP included several severity levels in the impeller
(see Table 1). Examples of the vibration signals for the CP dataset are shown in Figure 3.
Four conditions are shown, and in this subset, only subtle differences in signal amplitude
can be visually detected. Any fault signature could be affected by noise.
In Figure 4, the AppEn for four vibration signals of the CP dataset are shown. Each
of the vibration signals represents a fault condition. The utility of non-linear features for
detecting fault conditions in vibration signals is shown in the following paragraphs. The
feature estimation in each signal starts by performing the phase space reconstruction [47]
aimed at estimating the system dimension m and the lag τ. These parameters are used
for calculating the AppEn for several values of r. The AppEn for the CP signals increases
with r until attaining a maximum and decreases with the increase in r. In this example,
the differences are evident between each impeller’s conditions.
The LLE is calculated using four vibration signals extracted from the CP dataset.
The calculation procedure was explained in Section 3.9. Results concerning the LLE are
shown in Figure 5. The LLE is negative for the healthy class, class P6, and class P13.
However, it is positive for class P10. The sign of the Lyapunov exponent could be related to
the presence of chaos when the LLE is positive. When the LLE is negative, the time series
could represent a system with periodic dynamics, as suggested in [65].
A non-integer dimension usually characterizes non-linear systems that could have
a chaotic nature. Generally, when dealing with chaotic systems, the Rényi dimension is used
as a feature of strange attractors that enables the estimation of several feature dimensions.
The D0 is the fractal dimension, D1 is the information dimension, and D2 is the CD [81].
Like AppEn estimation, the first step corresponds to the phase space reconstruction that
provides the embedding dimension m. Parameter m is the upper bound of the CD [82].
The CD feature was estimated for several vibration signals of the CP dataset. The results
are shown in Figure 6. This set of signals has estimated embedding dimensions of 5 for P1
and P6, and 4 for P10 and P13. The CD is lower than 5 and higher than 4 for the signals

341
Mathematics 2022, 10, 3033

from classes P1 and P6. Similarly, for classes P10, and P13, the CD values are higher than 3
and lower than 4.

(a) (b)

(c) (d)

Figure 3. Vibration signals from the CP dataset. (a) Signal from class P1. (b) Signal from class P6.
(c) Signal from class P10, (d) Signal from class P13.

(a) (b)

(c) (d)

Figure 4. AppEn for several signals from the pump dataset. (a) ApEn for a signal from class P1,
(b) ApEn for a signal from class P6, (c) ApEn for a signal from class P10, (d) ApEn for a signal from
class P13.

342
Mathematics 2022, 10, 3033

(a) (b)

(c) (d)

Figure 5. LLE for several signals from the CP dataset. (a) LLE for class P1, (b) LLE for class P6,
(c) LLE for class P10, (d) LLE for class P13.

(a) (b)

(c) (d)

Figure 6. CD for several signals from the CP dataset. (a) CD for a signal from class P1, (b) CD for a
signal from class P6, (c) CD for a signal from class P10, (d) CD for a signal from class P13.

The percentages of accuracy for classifying the fault considering the CP dataset using
RF and SVM are presented in Table 3. The accuracy is presented for each of the models and
feature sets considered. The vibration signal channels A2, A3, and A4 were considered
for feature extraction. Vibration signal channel A1 was not considered for two reasons.
Firstly, the computational time of the feature extraction is high, and to keep the total feature
extraction time feasible with the available hardware, we only considered three channels.

343
Mathematics 2022, 10, 3033

Secondly, we selected the channels close to the mechanical connection with the motor,
where a roller bearing is located, because the pump stages in this neighborhood have
a higher probability of failing.

Table 3. Percentages of accuracy for fault classification using the vibration signals of the CP dataset.

Features Model A2 A3 A4
Statistical RF 67.76 68.34 68.50
SVM 68.50 70.51 71.03
Entropy-based RF 98.09 97.83 94.70
SVM 99.50 98.96 97.30
Info-Statistical RF 84.60 84.80 93.81
SVM 86.26 86.96 96.32
All Features RF 97.76 99.38 99.37
SVM 99.59 99.69 99.81

The lowest classification accuracy was attained using RF trained with the Statistical
Features set. In this case, the accuracy attained from channel A2 was 67.76%, for channel A3
was 68.34%, and for channel A4 was 68.50%. The accuracy attained using SVM with this
set of features was 68.5% for channel A2, 70.51% for channel A3, and 71.03% for channel
A4. The highest accuracy was 99.59% for channel A2, 99.69% for channel A3, and 99.81%
for channel A4. This accuracy was attained with the All Features set and SVM. With this
set of features, using the RF model, the accuracy was 97.76% for channel A2, 99.38% for
channel A3, and 99.37% for channel A4.
A comparison of accuracy considering both machine learning models and feature sets
is presented in Figure 7. In the box plot, a red line represents the median of the data that
divides the box or inter-quartile range in two parts. This box represents 50% of the data.
The horizontal line over the box represents the upper quartile that indicates that 75% of
the data have values below this quartile. The horizontal line below the box represents
the lower quartile that indicates that 25% of the data have values below this quartile. The
red cross below or over the box represents the outliers. The accuracy is presented for
channel A2 in box plots calculated based on the ten repetitions performed for obtaining ten
different cross-validated models. The highest accuracy (99.50% and 99.59% in Table 3) was
attained based on the Entropy Features and the complete set of features, All Features, with
the SVM model. The lower accuracies (67.76% and 68.50%) for the RF and SVM models
were attained by considering the Statistical Features set. The combination of information
entropy and statistical features (InfoStat Features) provided an intermediate accuracy value
(84.60% and 86.96%) for both considered models.
Table 4 presents several performance metrics [79,80,83] results expressed in percent-
ages. The included metrics are the sensitivity, the specificity, the error, the false positive
rate, and the area under the curve (AUC). The metrics were calculated during the cross-
validation results of the SVM model trained with the All Features set extracted from vibration
signals recorded in the channel A2 from the CP. The highest sensitivity of 100.00% was
attained by classes P1, P6, P12, and P13. In contrast, the lowest sensitivity of 98.67% was
attained by class P9. The highest specificity of 100% was attained by classes P1, P3, P5,
and P12. The lowest specificity of 99.87% was attained by classes P4 and P8. Concerning
the AUC [84], the highest value of 100% was attained by classes P1 and P12, and the lowest
AUC value was 99.19%.

344
Mathematics 2022, 10, 3033

Figure 7. Comparison of accuracy for the machine learning models tested, considering several
combinations of features from channel A2 acquired from the CP.

Table 4. Performance metrics (in percent) obtained with the SVM model and the All Features set.
The model was trained using vibration signals in channel A2 from the CP dataset.

Class Sensitivity Specificity Error FPR AUC

P1 100.00 100.00 0.00 0.00 100.00
P2 99.12 99.96 0.88 0.00 99.74
P3 99.78 100.00 0.22 0.00 99.99
P4 99.11 99.87 0.89 0.13 99.19
P5 98.68 100.00 1.32 0.00 99.94
P6 100.00 99.98 0.00 0.02 99.89
P7 98.90 99.96 1.10 0.04 99.73
P8 99.55 99.87 0.45 0.13 99.20
P9 98.67 99.93 1.33 0.07 99.50
P10 99.56 99.98 0.44 0.02 99.87
P11 99.33 99.91 0.67 0.09 99.42
P12 100.00 100.00 0.00 0.00 100.00
P13 100.00 99.93 0.00 0.07 99.56

8.2. Results for the RC Dataset

A set of vibration signals for the RC is presented in Figure 8. The vibration signals
presented are for the healthy class P1, and for class P6 representing the valve seat wear (in
the inlet valve) of the second stage. The bottom row shows the vibration signal for class
P11 corresponding to corrosion of the valve plate located in the discharge valve of the first
stage. Finally, the vibration signal for the class P16 represents the fracture of the valve
plate for inlet valve b of the first stage. The appearance of this vibration signal set differs
from that of the CP dataset’s signals. In this dataset, the signals are more peaked, and each
prominent peak indicating an event is surrounded by signals with lower amplitudes.
The AppEn estimated for each of the signals shown in Figure 8 is shown in Figure 9.
The curve shape of AppEn is different from that of the CP. The AppEn in the RC is higher
for low amplitudes and decreases as the threshold of amplitudes represented by the radius
r increases.

345
Mathematics 2022, 10, 3033

The calculation of the LLE for the set of vibration signals selected from the RC dataset
is shown in Figure 10. Although the LLE value is different for each signal presented, all
the estimated values are positive, suggesting that the system is chaotic [65]. The CD is
presented in Figure 11 for the set of signals of the RC. The CD for these signals has different
values according to the fault type. The estimated values of CD are close to the embedded
dimensions [82] (four, calculated using the phase space reconstruction method).

(a) (b)

(c) (d)

Figure 8. Vibration signals from the RC dataset. (a) Signal from the class P1, (b) signal from class P6,
(c) signal from class P11, (d) signal from class P16.

(a) (b)

(c) (d)

Figure 9. AppEn for several signals from the RC dataset. (a) ApEn for a signal from the class P1,
(b) ApEn for a signal from the class P6, (c) ApEn for a signal from the class P11, (d) ApEn for a signal
from the class P16.

346
Mathematics 2022, 10, 3033

(a) (b)

(c) (d)

Figure 10. LLE for several signals from the RC dataset. (a) LLE for class P1, (b) LLE for class P6,
(c) LLE for a signal from the class P11, (d) LLE for class P16.

(a) (b)

(c) (d)

Figure 11. CD for several signals from the RC dataset. (a) CD for a signal from the class P1, (b) CD for a
signal from the class P6, (c) CD for a signal from the class P11, (d) CD for a signal from the class P16.

Concerning the RC, we selected three vibration channels from the four available.
The selected channels were A1, A2, and A3. We excluded A4 to save computational time
because A4 was far from the valve’s location. The percent accuracy attained using RF and
SVM is presented in Table 5. Four feature sets are considered: the first is the Statistical
Features set, and the second is the Entropy Features. The third is the InfoStat Features set.

347
Mathematics 2022, 10, 3033

Finally, we have the All Features set. The lowest accuracy was attained with the Statistical
Features set. In this case, the accuracy attained by the RF model was 83.96% for vibration
signal channel A1, 74.07% for A2; and 72.33% for A3. The accuracy attained by the SVM
model was 86.91% for vibration signal channel A1, 76.91% for A2, and 75.55% for A3.
The highest accuracy was attained using the All Features set. The accuracy attained by the
RF model was 95.35% for vibration signal channel A1, 94.40% for A2, and 88.82% for A3.
The SVM model attained a classification accuracy of 97.90% for signal channel A1, 95.47%
for A2 and 93.63% for A3. The rest of the feature combinations attained intermediate
accuracy close to the optimal accuracy.

Table 5. Percentage accuracy for fault classification using the vibration signals of the RC dataset.

Features Model A1 A2 A3
Statistical RF 83.96 74.07 72.33
SVM 86.91 76.55 75.55
Entropy based RF 95.35 94.12 86.21
SVM 97.57 94.55 91.87
Info-Statistical RF 93.96 90.66 85.78
SVM 96.83 94.27 90.38
All Features RF 95.35 94.40 88.82
SVM 97.90 95.47 93.63

A comparison of results obtained during the ten repetitions of the cross-validation

is presented in Figure 12 for vibration signal A1. The comparison is presented using box
plots. The accuracies attained by the Statistical Features set with RF and SVM were the
lowest (83.96% and 86.91% in Table 5). The feature sets Entropy Features, InfoStat Features,
and All Features attained higher accuracy with both machine learning models. In particular,
the highest accuracies were attained by the All Features set with RF and SVM (95.35% and
97.90%). The accuracies attained by the Entropy Features were slightly lower, corresponding to
95.35% and 97.57% with the RD and SVM models, respectively. The accuracies attained by the
InfoStat Features set were 93.96% and 96.83% with the RF and SVM models, respectively.

Figure 12. Comparison of accuracy for the machine learning models tested, considering several
combinations of features extracted from channel A1 of the RC dataset. In the box plot, a red line
represents the median of the data. The horizontal line over the box represents the upper quartile. The
horizontal line below the box represents the lower quartile. The red cross, below or over the box,
represents the outliers.

348
Mathematics 2022, 10, 3033

A set of performance metrics [79,80,83] expressed in percentages is presented in Table 6.

Such metrics were obtained with the SVM model trained with the All Features set extracted
from vibration signals recorded in channel A1 from the RC dataset. The highest sensitivity
of 100.00% was attained by classes P1, P2, P6, P7, P9, P10, and P11; and 93.99% attained by
class P3 was the lowest. The highest specificity of 100% was attained by classes P1, P2, P10,
and P17. The highest AUC value of 100.00% was attained by classes P1, P2, and P10; and
the lowest value of 97.42% was attained by class P5.

Table 6. Performance metrics (in percent) obtained with the SVM model and the All Features set.
The model was trained using vibration signals in channel A1 from the RC dataset.

Class Sensitivity Specificity Error FPR AUC

P1 100.00 100.00 0.00 0.00 100.00
P2 100.00 100.00 0.00 0.00 100.00
P3 96.07 99.73 3.93 0.27 97.70
P4 99.56 99.92 0.44 0.08 99.33
P5 93.99 99.70 6.01 0.30 97.42
P6 100.00 99.97 0.00 0.03 99.78
P7 100.00 99.78 0.00 0.22 98.26
P8 97.01 99.92 2.99 0.08 99.25
P9 100.00 99.95 0.00 0.05 99.57
P10 100.00 100.00 0.00 0.00 100.00
P11 100.00 99.86 0.00 0.14 98.91
P12 94.56 99.89 5.44 0.11 98.95
P13 99.13 99.92 0.87 0.08 99.32
P14 99.56 99.95 0.44 0.05 99.55
P15 99.56 99.92 0.44 0.08 99.33
P16 97.42 99.92 2.58 0.08 99.27
P17 98.29 100.00 1.71 0.00 99.95

9. Discussion
The proposed methodology was implemented on a laptop computer. The laptop
was equipped with an intel(R) Core(TM) i7-6700HQ CPU, 12 GB of RAM memory, and
a graphics card NVIDIA GeForce GTX 950M. The Matlab software was used to implement
the feature extraction and classification. Concerning the computational time, the Entropy
Features set required an average time of 311.85 s when extracted from the five windows
(each with a length of 12,000-time samples) of the vibration signals from the CP dataset.
In contrast, the Information Entropy features require 0.97 s, and the Statistical Features set
requires only 0.156 s. In the RC, when the sampling rate is similar to that of the CP and the
window length is similar, the computational time is similar to the time required in the CP.
In the case of the CP, multiple pairwise comparisons of machine learning models
are presented in Figure A1. A one-way analysis of variance by ranks was performed
using a Kruskal–Wallis approach [85]. The test showed no significant statistical differences
between results attained by either machine learning models when trained with the non-
linear Entropy Features set or All Features set. Results attained by the SVM machine learning
models trained with the non-linear Entropy Features set or the All Features set are statistically
significantly superior to than those of RF or SVM models trained with the Statistical Features
or the combination InfoStat Features. The comparison also shows that the non-linear Entropy
Features set provides higher accuracy without combining with the Statistical Features set.

349
Mathematics 2022, 10, 3033

The plot representing the Kruskal–Wallis with multiple one-way comparisons is

presented in Figure A2 for the RC. The combination of features corresponding to Entropy
Features and All Features sets attained higher accuracies using both machine learning models,
and there were no significant statistical differences between them. In contrast, there were
significant statistical differences in the results attained with the Statistical Features set
using both machine learning models. Additionally, the results obtained by the features
sets Entropy Features, and All Features with the SVM model were statistically significantly
superior to the results obtained with SVM using the InfoStat Features set.
A comparison with a deep-learning method was performed. For this purpose, the spec-
trogram of the vibration signal was calculated and fed to a CNN for fault classification.
The spectrogram was calculated considering a Blackman widow of size 81.92 ms that was
moved along the time axis with an overlap of 81.44 ms. The feature array had a size of
96 × 128, for each signal, and it was standardized using the Z-score normalization. The fea-
ture array for each dataset was fed to a CNN for 10-fold cross-validation. Results of the
comparison are presented in Table 7. Concerning the case of the CP, the highest accuracy
attained by the CNN was 98.86% for vibration channel A2, whereas using the Entropy
Features set with SVM provided a classification accuracy of 99.50%. The accuracy attained
by the RF model was 98.09%, which was slightly lower than the accuracy provided by the
CNN. Results for the RC show that the highest accuracy attained by the CNN models was
96.74% when trained with signals from the A1 vibration channel. In contrast, the accuracy
provided by the SVM trained with the Entropy Features extracted from the same channel
was 97.57%. In summary, the accuracy provided by the SVM trained with the Entropy
Features was higher than the accuracy attained by the CNN trained with the spectrogram of
the vibration signal.

Table 7. Comparison of performances of the tested models obtained during the 10-fold cross-
validation between entropy features with classical models and the spectrogram with a CNN model.

Machine Signal Accuracy Accuracy Accuracy

Entropy Features Entropy Features Spectrogram
+SVM +RF +CNN
A2 99.50 98.09 98.86
Centrifugal
A3 98.96 97.83 97.36
Pump
A4 97.30 94.70 98.61
A1 97.57 95.35 96.74
Reciprocating
A2 94.55 94.12 95.55
Compressor
A3 91.87 86.21 93.98

Results can be compared to previous research that used non-linear entropy features
with classical models. In [27], the average accuracy for classifying five different conditions
in a CP was 94.58%. The work reported in [28] used an SVM for classifying four different
conditions in a CP using vibration signals. The method attained an average accuracy of
98.4%. In [21] the p − V, was used with LDA to classify six different conditions, attaining
an accuracy of 99.62% in an RC. Several papers reported CNN-based approaches for fault
classification in CP and RC [24,25,29,33]. These methods can highly accurately classify a
small set of faults in a CP or an RC. In contrast, our method was able to classify 13 types of
faults in a CP and 17 different conditions in an RC with accuracies close to those of these
deep learning-based methods.

10. Conclusions
The non-linear Entropy Features set enables fault classification in a CP or an RC with
high accuracy. In this research, 13 impeller faults in a CP incorporating different degrees of

350
Mathematics 2022, 10, 3033

severity were classified using RF and SVM models. Additionally, a set of 17 valve faults in
an RC were also classified.
The research showed that the non-linear Entropy Features set provides high accuracy
(99.50% for the CP and 97.57% for the RC) considering the SVM model. The accuracy
attained using the Entropy Features set and the RF model was 98.09% for the CP and 95.35%
for the RC. Results showed no significant statistical differences between Entropy Features
and All Features. However, both feature sets provided accuracies statistically significantly
higher than the accuracy provided by the Statistical Features set. The Statistical Features set
provided the lowest accuracy in this application. In this case, the accuracy attained with
the SVM was 71.03% for the CP and 76.55% for the RC. Concerning the comparison of
machine learning models, when both models RF and SVM were trained with the Entropy
Features set, there were no statistically significant differences in the accuracy of results.
The results suggest that the Entropy Features set or the All Features set could be used for
fault classification in centrifugal pumps and reciprocating compressors, where multiple
fault types in impellers and valves should be diagnosed at early stages.
The novelty of the approach is to propose using the entropy features for classifying
faults in centrifugal pumps and reciprocating compressors with high accuracy. In addition,
we showed that the entropy-based feature sets provide high classification accuracy even
for a large number of faults in both the CP and the RC. Concerning the comparison with
respect to a deep learning model, the accuracy results obtained by the SVM trained with
the Entropy Features set (99.50% for the CP and 97.57% for the RC) were superior to those
obtained by a CNN trained with the spectrogram (98.86% for the CP and 96.74% for the RC).
The limitation of the proposed approach is the high computational cost necessary for
extracting the entire set of entropy-based features.
Further research is oriented at reducing the computational time required for extracting
the non-linear Entropy Features set and exploring their implementation using multi-scale
approaches. We also plan to select some of the entropy-based features that would be useful
for anomaly detection in RM, combined with deep learning approaches, such as generative
adversarial networks and autoencoders.

Author Contributions: Conceptualization, R.-V.S., M.C. and R.M.; methodology, D.C. and M.C.;
software, R.M.; validation, R.M., E.E. and R.-V.S.; formal analysis, S.Y.; investigation, E.E.; resources,
S.Y.; data curation, D.C.; writing—original draft preparation, R.M.; writing—review and editing,
M.C. and R.-V.S.; visualization, D.C.; supervision, R.-V.S.; project administration, R.-V.S.; funding
acquisition, R.-V.S. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by the MoST Science and Technology Partnership Program
(KY201802006) and National Research Base of Intelligent Manufacturing Service, Chongqing Tech-
nology and Business University, and by the Universidad Politécnica Salesiana through the GIDTEC
research group.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data used in this research are available upon reasonable request to
the corresponding authors. We also plan to prepare a paper describing the datasets and submit it to
the MDPI journal Data to release them for public use.
Conflicts of Interest: The authors declare no conflict of interest.

351
Mathematics 2022, 10, 3033

Abbreviations
The following abbreviations are used in this manuscript:

CP Centrifugal Pump
RC Reciprocating Compressor
CM Condition Monitoring
CD Correlation Dimension
LLE Largest Lyapunov Exponent
LMD Local Mean Decomposition
SVM Support Vector Machine
CEEMD Complementary Ensemble Empirical Mode Decomposition
SampEn Sample Entropy
RF Random Forest
ELM Extreme Learning Machine
AMI Average Mutual Information
FNN False Nearest Neighbor
AppEn Approximate Entropy
FuzzyEn Fuzzy Entropy
ShannEn Shannon Entropy
PermEn Permutation Entropy
DFA Detrended Fluctuation Analysis
LLE Larger Lyapunov Exponent
S Power Spectral Entropy
PEB Pitting at the Entrance of Impeller Blades
POB Pitting at the Output of the Impeller Blades
ICB Impeller Channel Blockage
IB Imbalance Impeller
ROC Receiver Operator Curve
AUC Area Under the Curve
HP Horse Power
RMS Root Mean Square
RPM Revolutions per minute
GPM Gallons per minute
RM Rotating Machinery
CNN Convolutional Neural Network
SGWT Spectral Graph Wavelet Transform
CECP Complexity-Entropy Causality Plane
MSE Multi-Scale Entropy
PCA Principal Component Analysis
LDA Linear Discriminant Analysis
BMFO Binary Moth Flame Optimization
SE Squeeze and Excitation
HYP Hydraulic Pump
RSDD Resonance-based sparse signal decomposition
MHAAPE Multi-scale Hierarchical Amplitude Aware Permutation Entropy
ELM Extreme Learning Machine
CEEMD Complementary Ensemble Empirical Mode Decomposition
CWT Continous Wavelet Transform
XBG Extreme Gradient Boosting
AWT Analitical Wavelet Transform

352
Mathematics 2022, 10, 3033

Appendix A. Additional Tables

Appendix A.1. Related Research Compendium

Table A1. Related research and characteristics.

Contributors Signal Methods Application Advantages Limitations

Li et al. [21] Pressure PCA RC Accuracy Six types of fault
vibration LDA 99.62%
Patil et al. [23] Vibration BMFO RC Accuracy Can get
Signal KNN >99% trapped in
local minimum
Lv et al. [22] Pressure ANN RC Accuracy Several
97.9% Structuring
Elements
Zhao et al. [24] Vibration CNN RC Accuracy Complex CNN
99.4% architecture
Xiao et al. [25] Vibration CNN RC Acc = 100% Complex CNN
Pressure Architecture,
Phase 4 Conditions
Ahmad et al. [28] Vibration SVM CP Accuracy 4 conditions
98.4%
Hasan et al. [29] Vibration CNN CP Acc = 100% 4 conditions
Ahmad et al. [30] Vibration KNN CP Acc = 100% 4 conditions
Irfan et al. [31] Voltage XGB Water pump Acc = 100% Threshold
Current bearings Selection
Kumar et al. [33] Sound CNN CP Sound Sensor Acoustic noise
Zhao et al. [20] Vibration SVM RC Weak fault LMD
MSE detection Mode mixing
Wang et al. [27] Vibration RF + CEEMD CP Accurate High
SampEn Comp. Cost
Zhou et al. [26] Vibration RSDD + RF HYP Accurate Large set
MHAAPE of parameters
Xin et al. [16] Vibration DFA + SGWT RM Retain fine Lacks more
signatures testing
Radhakrishnan et al. [17] Vibration CECP + SVM RM Robust and Setting of
easy parameters

Appendix A.2. Entropy-Based Features

Table A2. Entropy features extracted.

Feature Equation
ApEn [5,6] ApEn(m, k, τ, N ) = φm (r ) − φm+1 (r ), k = 0.06 (A1)

ApEn [5,6] ApEn(m, k, τ, N ) = φm (r ) − φm+1 (r ), k = 0.21 (A2)

FuzzyEn [10] FuzzyEn(m, n, k) = lnφm (n, r ) − lnφm+1 (n, r ), k = 0.06 (A3)

FuzzyEn [10] FuzzyEn(m, n, k) = lnφm (n, r ) − lnφm+1 (n, r ), k = 0.21 (A4)

353
Mathematics 2022, 10, 3033

Table A2. Cont.

Feature Equation
 
A m (r )
SampEn [7] SampEn(m, r ) = −ln m , k = 0.06 (A5)
B (r )
 m 
A (r )
SampEn [7] SampEn(m, r ) = −ln m , k = 0.21 (A6)
B (r )
ShannonEn [54] ShanEn(m) = − ∑ pm logpm (A7)
m

CondEn [54] CondEn(m/m − 1) = ShanEn(m) − ShanEn(m − 1) (A8)

CCondEn [54] CCondEn(m) = CondEn(m/m − 1) + perc(m)ShanEn(1) (A9)

( m +1) ! −1
1
PermEn [57] PermEn(m) = lnM −
M ∑ q j lnq j (A10)
j =0

∂lnCm (r )
CorDim [59] CD (m) = lim (A11)
r →0 ∂lnr
1 δZ (t)
LLE [66] LLE = lim ln (A12)
t→∞ t δZ0 

DFA1(α1) [66] α1 = τ (1), τ (1) = h (1) − 1 (A13)


DFA2(α2) [66] α2 = τ (2), τ (2) = 2h(2) − 1 (A14)

Appendix A.3. Information-Based Features

Table A3. Information entropy features.

Feature Equation
WPT-Shannon [70,86] WPT − Shannon(s) = − ∑ s2i log(s2i ) (A15)
i
WPT-Norm [70,86] WPT − Norm(s) = − ∑ |si | p , p≥1 (A16)
i

WPT-LogEn [70,86] WPT − LogEn(s) = − ∑ log(s2i ) (A17)

i
WPT-Thres [70,86] WPT − Thres(s) = #{i, / |si | > p}, | si | > p (A18)

WPT-Sure [70,86] WPT − Sure(s) = n + ∑ min(s2i , p2 ) − #{i, / |si | ≤ p} (A19)

i
 
N
K
log ∑ pi
q
RényiEn [72] HRq = (A20)
q−1 i =1
 
N
K
1 − ∑ pi
q
TsallisEn [71] HTq = (A21)
q−1 i =1

PSE-mean [73,74] PSE − mean = mean(S) (A22)

PSE-std [73,74] PSE − std = std(S) (A23)

PSE-rms [73,74] PSE − rms = rms(S) (A24)

354
Mathematics 2022, 10, 3033

Table A3. Cont.

Feature Equation
rms(S)
PSE-shape [73,74] PSE − shape = (A25)
1
N ∑i | si |
PSE-MaxToRms [73,74] PSE − MaxToRms = max (S)/rms(S) (A26)

PSE-median [73,74] PSE − median = median(S) (A27)

PSE-Skew [73,74] PSE − Skew = skewness(S) (A28)

PSE-Kur [73,74] PSE − Kur = kurtosis(S) (A29)

Appendix A.4. Statistical Features

Table A4. Statistical features.

Feature Equation
N
1
Mean μ=
N ∑ xi (A30)
i =1
0
1
11 N
Root Mean Square (RMS) RMS = 2 ∑ ( xi )2 (A31)
N i =1
0
1
11 N
Standard deviation σ = 2 ∑ ( x i − μ )2 (A32)
N i =1

N ∑iN=1 ( xi − μ)4
Kurtosis Kurtosis = ! (A33)
2 2
∑iN=1 ( xi − μ)
Maximum value Max = max (xn ) (A34)

max (xn )
Crest factor CrestFact = (A35)
rms(xn )
N
1
Rectified mean value RMV =
N ∑ | xi | (A36)
i =1
rms(xn )
Shape factor ShapeFact = (A37)
1
N∑iN=1 | xi |
max (xn )
Impulse factor ImpulseFact = (A38)
1
N∑iN=1 | xi |
N
1
Variance Var =
N ∑ ( x i − μ )2 (A39)
i =1
Minimum value Min = min(xn ) (A40)

N ∑iN=1 ( xi − μ)3
Skewness Skewness = (A41)
σ3

355
Mathematics 2022, 10, 3033

Appendix B. Additional Figures

Figure A1. Kruskal–Wallis comparison between the machine learning models tested. The plot shows
several combinations of features calculated from the channel A2 recorded from the CP. There are
statistical significant differences between the groups labeled with red and the group labeled with
blue, between the groups SVM-Entropy and SVM-AllFeat and the groups labeled with blue and gray.
In general there are statistically significant differences, when the projection of the horizontal lines of a
group over the horizontal axis does not have any interception with the projection of another group.

Figure A2. Kruskal–Wallis comparison between the machine learning models tested, considering
several combinations of features extracted from the channel A1 of the RC dataset. There are statistical
significant differences between the groups labeled with red and the group labeled with blue. In
general there are statistically significant differences, when the projection of the horizontal lines of a
group over the horizontal axis does not have any interception with the projection of another group.

References
1. Ali, S.M.; Hui, K.; Hee, L.; Leong, M.S. Automated valve fault detection based on acoustic emission parameters and support
vector machine. Alex. Eng. J. 2018, 57, 491–498. [CrossRef]
2. Cerrada, M.; Macancela, J.C.; Cabrera, D.; Estupiñan, E.; Sánchez, R.V.; Medina, R. Reciprocating compressor multi-fault
classification using symbolic dynamics and complex correlation measure. Appl. Sci. 2020, 10, 2512. [CrossRef]
3. Sharma, V.; Parey, A. Performance evaluation of decomposition methods to diagnose leakage in a reciprocating compressor
under limited speed variation. Mech. Syst. Signal Process. 2019, 125, 275–287. [CrossRef]

356
Mathematics 2022, 10, 3033

4. Soleimani, A.; Khadem, S. Early fault detection of rotating machinery through chaotic vibration feature extraction of experimental
data sets. Chaos Solitons Fractals 2015, 78, 61–75. [CrossRef]
5. Pincus, S.M. Approximate entropy as a measure of system complexity. Proc. Natl. Acad. Sci. USA 1991, 88, 2297–2301. [CrossRef]
[PubMed]
6. Pincus, S.M.; Huang, W.M. Approximate entropy: Statistical properties and applications. Commun. Stat. Theory Methods 1992,
21, 3061–3077. [CrossRef]
7. Richman, J.S.; Moorman, J.R. Physiological time-series analysis using approximate entropy and sample entropy. Am. J. Physiol.
Heart Circ. Physiol. 2000, 278, H2039–H2049. [CrossRef]
8. Weippert, M.; Behrens, M.; Rieger, A.; Behrens, K. Sample entropy and traditional measures of heart rate dynamics reveal
different modes of cardiovascular control during low intensity exercise. Entropy 2014, 16, 5698–5711. [CrossRef]
9. Zhao, L.; Wei, S.; Zhang, C.; Zhang, Y.; Jiang, X.; Liu, F.; Liu, C. Determination of sample entropy and fuzzy measure entropy
parameters for distinguishing congestive heart failure from normal sinus rhythm subjects. Entropy 2015, 17, 6270–6288. [CrossRef]
10. Chen, W.; Wang, Z.; Xie, H.; Yu, W. Characterization of surface EMG signal based on fuzzy entropy. IEEE Trans. Neural Syst.
Rehabil. Eng. 2007, 15, 266–272. [CrossRef]
11. Wang, Z.; Xue, X. Multi-class support vector machine. In Support Vector Machines Applications; Springer: Berlin/Heidelberg,
Germany, 2014; pp. 23–48. [CrossRef]
12. Escalera, S.; Pujol, O.; Radeva, P. On the decoding process in ternary error-correcting output codes. IEEE Trans. Pattern Anal.
Mach. Intell. 2010, 32, 120–134. [CrossRef] [PubMed]
13. Janjarasjitt, S.; Ocak, H.; Loparo, K. Bearing condition diagnosis and prognosis using applied nonlinear dynamical analysis of
machine vibration signal. J. Sound Vib. 2008, 317, 112–126. [CrossRef]
14. Sun, Y. Fault Detection in Dynamic Systems Using the Largest Lyapunov Exponent. Ph.D. Thesis, Texas A & M University,
College Station, TX, USA, 2012.
15. Medina, R.; Macancela, J.C.; Lucero, P.; Cabrera, D.; Sánchez, R.V.; Cerrada, M. Gear and bearing fault classification under
different load and speed by using Poincaré plot features and SVM. J. Intell. Manuf. 2020, 33, 1031–1055. [CrossRef]
16. Dong, X.; Li, G.; Jia, Y.; Li, B.; He, K. Non-iterative denoising algorithm for mechanical vibration signal using spectral graph
wavelet transform and detrended fluctuation analysis. Mech. Syst. Signal Process. 2021, 149, 107202. [CrossRef]
17. Radhakrishnan, S.; Lee, Y.T.T.; Rachuri, S.; Kamarthi, S. Complexity and entropy representation for machine component
diagnostics. PLoS ONE 2019, 14, e0217919. [CrossRef]
18. Li, Y.; Wang, X.; Liu, Z.; Liang, X.; Si, S. The entropy algorithm and its variants in the fault diagnosis of rotating machinery:
A review. IEEE Access 2018, 6, 66723–66741. [CrossRef]
19. Huo, Z.; Martínez-García, M.; Zhang, Y.; Yan, R.; Shu, L. Entropy measures in machine fault diagnosis: Insights and applications.
IEEE Trans. Instrum. Meas. 2020, 69, 2607–2620. [CrossRef]
20. Zhao, H.-Y.; Wang, J.-D.; Xing, J.-J.; Gao, Y.-Q. A feature extraction method based on LMD and MSE and its application for fault
diagnosis of reciprocating compressor. J. Vibroeng. 2015, 17, 3515–3526.
21. Li, X.; Ren, P.; Zhang, Z.; Jia, X.; Peng, X. A p&minus;V Diagram Based Fault Identification for Compressor Valve by Means of
Linear Discrimination Analysis. Machines 2022, 10, 53. [CrossRef]
22. Lv, Q.; Cai, L.; Yu, X.; Ma, H.; Li, Y.; Shu, Y. An Automatic Fault Diagnosis Method for the Reciprocating Compressor Based on
HMT and ANN. Appl. Sci. 2022, 12, 5182. [CrossRef]
23. Patil, A.; Soni, G.; Prakash, A. A BMFO-KNN based intelligent fault detection approach for reciprocating compressor. Int. J. Syst.
Assur. Eng. Manag. 2021, 1–13. [CrossRef]
24. Zhao, D.; Liu, S.; Zhang, H.; Sun, X.; Wang, L.; Wei, Y. Intelligent fault diagnosis of reciprocating compressor based on attention
mechanism assisted convolutional neural network via vibration signal rearrangement. Arab. J. Sci. Eng. 2021, 46, 7827–7840.
[CrossRef]
25. Xiao, S.; Nie, A.; Zhang, Z.; Liu, S.; Song, M.; Zhang, H. Fault Diagnosis of a Reciprocating Compressor Air Valve Based on Deep
Learning. Appl. Sci. 2020, 10, 6596. [CrossRef]
26. Zhou, F.; Liu, W.; Yang, X.; Shen, J.; Gong, P. A new method of health condition detection for hydraulic pump using enhanced
whale optimization-resonance-based sparse signal decomposition and modified hierarchical amplitude-aware permutation
entropy. Trans. Inst. Meas. Control 2021, 43, 3360–3376. [CrossRef]
27. Wang, Y.; Lu, C.; Liu, H.; Wang, Y. Fault diagnosis for centrifugal pumps based on complementary ensemble empirical mode
decomposition, sample entropy and random forest. In Proceedings of the 2016 12th World Congress on Intelligent Control and
Automation (WCICA), Guilin, China, 12–15 June 2016; IEEE: Piscataway, NJ, USA, 2016; pp. 1317–1320. [CrossRef]
28. Ahmad, Z.; Rai, A.; Maliuk, A.S.; Kim, J.M. Discriminant feature extraction for centrifugal pump fault diagnosis. IEEE Access
2020, 8, 165512–165528. [CrossRef]
29. Hasan, M.J.; Rai, A.; Ahmad, Z.; Kim, J.M. A fault diagnosis framework for centrifugal pumps by scalogram-based imaging and
deep learning. IEEE Access 2021, 9, 58052–58066. [CrossRef]
30. Ahmad, Z.; Nguyen, T.K.; Ahmad, S.; Nguyen, C.D.; Kim, J.M. Multistage centrifugal pump fault diagnosis using informative
ratio principal component analysis. Sensors 2021, 22, 179. [CrossRef]

357
Mathematics 2022, 10, 3033

31. Irfan, M.; Alwadie, A.S.; Glowacz, A.; Awais, M.; Rahman, S.; Khan, M.K.A.; Jalalah, M.; Alshorman, O.; Caesarendra, W. A novel
feature extraction and fault detection technique for the intelligent fault identification of water pump bearings. Sensors 2021,
21, 4225. [CrossRef]
32. Chen, T.; Guestrin, C. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM Sigkdd International Conference
on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794. [CrossRef]
33. Kumar, A.; Gandhi, C.; Zhou, Y.; Kumar, R.; Xiang, J. Improved deep convolution neural network (CNN) for the identification of defects
in the centrifugal pump using acoustic images. Appl. Acoust. 2020, 167, 107399. [CrossRef]
34. Howedi, A.; Lotfi, A.; Pourabdollah, A. An entropy-based approach for anomaly detection in activities of daily living in the
presence of a visitor. Entropy 2020, 22, 845. [CrossRef]
35. Callegari, C.; Giordano, S.; Pagano, M. Entropy-based network anomaly detection. In Proceedings of the 2017 International
Conference on Computing, Networking and Communications (ICNC), Santa Clara, CA, USA, 26–29 January 2017; IEEE:
Piscataway, NJ, USA, 2017; pp. 334–340. [CrossRef]
36. Palmieri, F. Network anomaly detection based on logistic regression of nonlinear chaotic invariants. J. Netw. Comput. Appl. 2019,
148, 102460. [CrossRef]
37. Wu, S.; Moore, B.E.; Shah, M. Chaotic invariants of lagrangian particle trajectories for anomaly detection in crowded scenes. In
Proceedings of the 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Francisco, CA,
USA, 13–18 June 2010; IEEE: Piscataway, NJ, USA, 2010; pp. 2054–2060. [CrossRef]
38. Garland, J.; Jones, T.R.; Neuder, M.; Morris, V.; White, J.W.; Bradley, E. Anomaly detection in paleoclimate records using
permutation entropy. Entropy 2018, 20, 931. [CrossRef] [PubMed]
39. Wang, S.; Lu, M.; Kong, S.; Ai, J. A Dynamic Anomaly Detection Approach Based on Permutation Entropy for Predicting
Aging-Related Failures. Entropy 2020, 22, 1225. [CrossRef] [PubMed]
40. Ahmad, S.; Styp-Rekowski, K.; Nedelkoski, S.; Kao, O. Autoencoder-based condition monitoring and anomaly detection method
for rotating machines. In Proceedings of the 2020 IEEE International Conference on Big Data (Big Data), Atlanta, GA, USA,
10–13 December 2020; IEEE: Piscataway, NJ, USA, 2020; pp. 4093–4102. [CrossRef]
41. de Paula Monteiro, R.; Lozada, M.C.; Mendieta, D.R.C.; Loja, R.V.S.; Bastos Filho, C.J.A. A hybrid prototype selection-based deep
learning approach for anomaly detection in industrial machines. Expert Syst. Appl. 2022, 204, 117528. [CrossRef]
42. Li, D.; Chen, D.; Jin, B.; Shi, L.; Goh, J.; Ng, S.K. MAD-GAN: Multivariate anomaly detection for time series data with generative
adversarial networks. In Proceedings of the International Conference on Artificial Neural Networks, Munich, Germany,
17–19 September 2019; Springer: Berlin/Heidelberg, Germany, 2019; pp. 703–716. [CrossRef]
43. Kumar, S.R.; Iniyal, U.; Harshitha, V.; Abinaya, M.; Janani, J.; Jayaprasanth, D. Anomaly Detection in Centrifugal Pumps Using
Model Based Approach. In Proceedings of the 2022 8th International Conference on Advanced Computing and Communication
Systems (ICACCS), Coimbatore, India, 25–26 March 2022; IEEE: Piscataway, NJ, USA, 2022; Volume 1, pp. 427–433. [CrossRef]
44. Dutta, N.; Kaliannan, P.; Subramaniam, U. Application of machine learning algorithm for anomaly detection for industrial pumps.
In Machine Learning Algorithms for Industrial Applications; Springer: Berlin/Heidelberg, Germany, 2021; pp. 237–263.[CrossRef]
45. Dutta, A.; Karimi, I.A.; Farooq, S. PROAD (Process Advisor): A Health Monitoring Framework for Centrifugal Pumps. Comput.
Chem. Eng. 2022, 163, 107825. [CrossRef]
46. Charoenchitt, C.; Tangamchit, P. Anomaly Detection of a Reciprocating Compressor using Autoencoders. In Proceedings of the
2021 Second International Symposium on Instrumentation, Control, Artificial Intelligence, and Robotics (ICA-SYMP), Bangkok,
Thailand, 20–22 January 2021; IEEE: Piscataway, NJ, USA, 2021; pp. 1–4. [CrossRef]
47. Rand, D.; Young, L.S. Dynamical Systems and Turbulence. In Lecture Notes in Mathematics; Springer: Cham, Switzerland, 1981;
Volume 898. [CrossRef]
48. Vlachos, I.; Kugiumtzis, D. State space reconstruction for multivariate time series prediction. arXiv 2008, arXiv:0809.2220.
https://doi.org/10.48550/arXiv.0809.2220.
49. Abarbanel, H.D.; Brown, R.; Sidorowich, J.J.; Tsimring, L.S. The analysis of observed chaotic data in physical systems. Rev. Mod.
Phys. 1993, 65, 1331. [CrossRef]
50. Fraser, A.M.; Swinney, H.L. Independent coordinates for strange attractors from mutual information. Phys. Rev. A 1986, 33, 1134.
[CrossRef]
51. Kennel, M.B.; Brown, R.; Abarbanel, H.D. Determining embedding dimension for phase-space reconstruction using a geometrical
construction. Phys. Rev. A 1992, 45, 3403. [CrossRef]
52. Ribeiro, M.; Henriques, T.; Castro, L.; Souto, A.; Antunes, L.; Costa-Santos, C.; Teixeira, A. The entropy universe. Entropy 2021,
23, 222. [CrossRef] [PubMed]
53. Azami, H.; Escudero, J. Amplitude-and fluctuation-based dispersion entropy. Entropy 2018, 20, 210.
doi: 10.3390/e20030210. [CrossRef] [PubMed]
54. Porta, A.; Baselli, G.; Liberati, D.; Montano, N.; Cogliati, C.; Gnecchi-Ruscone, T.; Malliani, A.; Cerutti, S. Measuring regularity by
means of a corrected conditional entropy in sympathetic outflow. Biol. Cybern. 1998, 78, 71–78. [CrossRef] [PubMed]
55. Feutrill, A.; Roughan, M. A Review of Shannon and Differential Entropy Rate Estimation. Entropy 2021, 23, 1046. [CrossRef]
56. Bandt, C.; Pompe, B. Permutation entropy: A natural complexity measure for time series. Phys. Rev. Lett. 2002, 88, 174102.
[CrossRef]

358
Mathematics 2022, 10, 3033

57. Unakafova, V.A.; Keller, K. Efficiently measuring complexity on the basis of real-world data. Entropy 2013, 15, 4392–4415.
[CrossRef]
58. Steven, V. Heart rate variability linear and nonlinear analysis with applications in human physiology. Diss. Abstr. Int. 2010, 71.
59. Lu, C.; Sun, Q.; Tao, L.; Liu, H.; Lu, C. Bearing health assessment based on chaotic characteristics. Shock Vib. 2013, 20, 519–530.
[CrossRef]
60. Rolo-Naranjo, A.; Montesino-Otero, M.E. A method for the correlation dimension estimation for on-line condition monitoring of
large rotating machinery. Mech. Syst. Signal Process. 2005, 19, 939–954. [CrossRef]
61. Ihlen, E.A. Introduction to multifractal detrended fluctuation analysis in Matlab. Front. Physiol. 2012, 3, 141. [CrossRef]
62. JiaQing, W.; Han, X.; Yong, L.; Tao, W.; Zengbing, X. Detrended Fluctuation Analysis and Hough Transform Based Self-Adaptation
Double-Scale Feature Extraction of Gear Vibration Signals. Shock Vib. 2016, 2016, 3409897. [CrossRef]
63. Golińska, A.K. Detrended fluctuation analysis (DFA) in biomedical signal processing: Selected examples. Stud. Log. Gramm.
Rhetor. 2012, 29, 107–115.
64. Henry, B.; Lovell, N.; Camacho, F. Nonlinear dynamics time series analysis. In Nonlinear Biomedical Signal Processing: Dynamic
Analysis and Modeling; Wiley: Hoboken, NJ, USA, 2012; Volume 2, pp. 1–39. [CrossRef]
65. Wolf, A.; Swift, J.B.; Swinney, H.L.; Vastano, J.A. Determining Lyapunov exponents from a time series. Phys. D Nonlinear Phenom.
1985, 16, 285–317. [CrossRef]
66. Rosenstein, M.T.; Collins, J.J.; De Luca, C.J. A practical method for calculating largest Lyapunov exponents from small data sets.
Phys. D Nonlinear Phenom. 1993, 65, 117–134. [CrossRef]
67. Shannon, C.E. A mathematical theory of communication. ACM SIGMOBILE Mob. Comput. Commun. Rev. 2001, 5, 3–55. [CrossRef]
68. Gao, J.; Liu, F.; Zhang, J.; Hu, J.; Cao, Y. Information entropy as a basic building block of complexity theory. Entropy 2013,
15, 3396–3418. [CrossRef]
69. Goldstein, S.; Lebowitz, J.L.; Tumulka, R.; Zanghì, N. Gibbs and Boltzmann entropy in classical and quantum mechanics. arXiv
2019, arXiv:1903.11870.
70. Coifman, R.R.; Wickerhauser, M.V. Entropy-based algorithms for best basis selection. IEEE Trans. Inf. Theory 1992, 38, 713–718.
[CrossRef]
71. Tsallis, C. Possible generalization of Boltzmann-Gibbs statistics. J. Stat. Phys. 1988, 52, 479–487. [CrossRef]
72. Rényi, A. On measures of entropy and information. In Contributions to the Theory of Statistics, Proceedings of the Fourth Berkeley
Symposium on Mathematical Statistics and Probability, Davis, CA, USA, 20–30 June 1960; The Regents of the University of California:
Davis, CA, USA, 1961; Volume 1.
73. Sharma, V.; Parey, A. A review of gear fault diagnosis using various condition indicators. Procedia Eng. 2016, 144, 253–263.
[CrossRef]
74. Pan, Y.; Chen, J.; Li, X. Spectral entropy: A complementary index for rolling element bearing performance degradation assessment.
Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci. 2009, 223, 1223–1231. [CrossRef]
75. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
76. Efron, B.; Tibshirani, R.J. An Introduction to the Bootstrap; CRC Press: Boca Raton, FL, USA, 1994.
77. Vapnik, V.N. An overview of statistical learning theory. IEEE Trans. Neural Netw. 1999, 10, 988–999. [CrossRef]
78. Scholkopf, B. Support vector machines: A practical consequence of learning theory. IEEE Intell. Syst. 1998, 13, 4. [CrossRef]
79. Hossin, M.; Sulaiman, M. A review on evaluation metrics for data classification evaluations. Int. J. Data Min. Knowl. Manag.
Process 2015, 5, 1. [CrossRef]
80. Sokolova, M.; Lapalme, G. A systematic analysis of performance measures for classification tasks. Inf. Process. Manag. 2009,
45, 427–437. [CrossRef]
81. Wang, W.; Chen, J.; Wu, Z. The application of a correlation dimension in large rotating machinery fault diagnosis. Proc. Inst.
Mech. Eng. Part C J. Mech. Eng. Sci. 2000, 214, 921–930. [CrossRef]
82. Boon, M.Y.; Henry, B.I.; Suttle, C.M.; Dain, S.J. The correlation dimension: A useful objective measure of the transient visual
evoked potential? J. Vis. 2008, 8, 6. [CrossRef] [PubMed]
83. Brown, J. Classifiers and their metrics quantified. Mol. Inform. 2018, 37, 1700127. [CrossRef] [PubMed]
84. Bradley, A.P. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognit. 1997,
30, 1145–1159. [CrossRef]
85. Pohlert, T. The pairwise multiple comparison of mean ranks package (PMCMR). R Package 2014, 27, 9.
86. Leite, G.d.N.P.; Araújo, A.M.; Rosas, P.A.C.; Stosic, T.; Stosic, B. Entropy measures for early detection of bearing faults. Phys. A
Stat. Mech. Its Appl. 2019, 514, 458–472. [CrossRef]

359
 mathematics

Article
Lateral Dynamic Simulation of a Bus under Variable Conditions
of Camber and Curvature Radius
Ester Olmeda 1, *, Enrique Roberto Carrillo Li 2 , Jorge Rodríguez Hernández 2 and Vicente Díaz 1

1 Mechanical Engineering Department, Research Institute of Vehicle Safety (ISVA),

University Carlos III de Madrid, Avda. de la Universidad, 30, 28911 Leganes, Madrid, Spain
2 Mechanical Engineering Section, Engineering Department, Pontificia Universidad Catolica del Perú,
Lima 15088, Peru
* Correspondence: eolmeda@ing.uc3m.es

Abstract: The objective of this paper is to describe a model for the simulation of the lateral dynamics
of a vehicle, specifically buses, under variable trajectory conditions, such as camber and radius
of curvature; in addition, a variable speed is added as a simulation parameter. The objective of
this study is the prevention of vehicle rollover and sideslip. An 8 degrees of freedom model was
developed, considering a front and a rear section of the bus with its respective suspension system,
and both sections have been connected by a torsion spring that emulates the torsional stiffness of the
vehicle chassis. A Panhard bar is also added at the rear as an additional element to the suspension
and the behavior of the bus when it is added is analyzed. This model also allows the evaluation
of the force on each suspension component, which allows for future controllability of the active
suspension components. The results show the dynamic behavior of the vehicle, and some indicators
are introduced to show the possible sideslip or rollover. As a conclusion, the influence of the road
parameters on the dynamic behavior of the bus and the effect of the Panhard bar on the dynamic
behavior of the bus can be pointed out.

Citation: Olmeda, E.; Carrillo Li, E.R.; Keywords: lateral dynamics; track parameters; vehicle mathematical modeling; safety vehicle index
Rodríguez Hernández, J.; Díaz, V.
Lateral Dynamic Simulation of a Bus MSC: 70E50
under Variable Conditions of Camber
and Curvature Radius. Mathematics
2022, 10, 3081. https://doi.org/
10.3390/math10173081 1. Introduction
Academic Editor: Armin Fügenschuh Vehicle dynamics deals with the mechanical modeling of vehicles as well as the
mathematical description and analysis of vehicles.
Received: 30 June 2022
Vehicle research models have undergone major changes in recent decades, from the
Accepted: 18 August 2022
traditional grouped parameter model to the finite element model, the dynamic substructure
Published: 26 August 2022
model and the dynamic multibody model; and from linear models to nonlinear models
Publisher’s Note: MDPI stays neutral with nonlinear stiffness or damping. The responses of these models can be obtained
with regard to jurisdictional claims in theoretically or numerically depending on the complexity of the resulting mathematical
published maps and institutional affil- models. This complexity is closely linked to the number of degrees of freedom considered
iations. for the respective modeling [1].
According to [2–4], among others, there are two ways to perform modeling of mechan-
ical systems: an empirical approach and an axiomatic or theoretical approach. In this work,
the axiomatic modeling method has been used for a vehicular system. The mathematical
Copyright: © 2022 by the authors.
or virtual model is obtained through the application of the Newton-D’Alembert principle,
Licensee MDPI, Basel, Switzerland.
obtaining a system of ordinary differential equations.
This article is an open access article
In general, the analytical modeling of any vehicle involves analyzing the interrela-
distributed under the terms and
conditions of the Creative Commons
tionships between a large number of components, as well as their response to different
Attribution (CC BY) license (https://
external disturbances and the effects of these on the rest of the components. Because of this,
creativecommons.org/licenses/by/ different models must be proposed to predict the behavior of the vehicle in the presence of
4.0/). a set of external stimuli in order to determine a specific type of response. Such modeling

Mathematics 2022, 10, 3081. https://doi.org/10.3390/math10173081 361 https://www.mdpi.com/journal/mathematics

Mathematics 2022, 10, 3081

implies leaving aside some components and disturbances with little or no relevance to the
phenomenon we are trying to understand.
In the case of the present work, the objective is to analyze the lateral dynamics of a
vehicle, and therefore to know the vertical and lateral response of it, as well as the vertical
and lateral movements of its components due to vertical disturbances and transverse loads;
therefore, longitudinal disturbances are left aside. In addition, the kinematics of the vehicle
in the longitudinal axis and its rotation around the z-axis will not be determined.
In recent years, different vehicle models have been proposed for the analysis of lateral
dynamics and accident consequences in case of rollover or sideslip. In [5], a restricted lateral
dynamics model of articulated vehicles and an algorithm for the estimation of sideslip
angle and cornering stiffness were proposed. The articulated vehicle was modeled by using
the bicycle model, the linear tire model, and the modified Dug-off model. In [6] is presented
an application of combined dynamic and finite element (FE) simulations to evaluate the
rollover crash of a bus. A nonlinear mathematical model with six degrees of freedom is
shown in [7] from which the effect of different factors on the stability of the vehicle on a
curved trajectory (braking angle, vehicle speed and acceleration a.s.o.) is studied. In [8], a
model that simulates the lateral rollover test including the effect of a variable position of
the center of gravity and calculates the maximum speed at which the bus can travel in a
curve without rolling over is proposed. The influence of several parameters (i.e., height
of the center of gravity, weight distribution between axles, chassis torsional stiffness, etc.)
on bus roll stability during a lane change maneuver, has been analyzed by means of six
degrees of freedom beam finite element model in [9].
The aim of lateral dynamics is to describe the behavior of a vehicle when it is subjected
to the action of accelerations and/or forces in its lateral direction in order to prevent a
possible rollover or sideslip of the vehicle. These solicitations, whose magnitudes are a
function of speed, commonly appear when the vehicle travels on a curved track. Different
studies show that a high number of traffic accidents take place on horizontal curves, where
the risk of accidents is significantly higher than on other road sections [10–12]. According
to several recent studies, the radius of curvature represents the most important factor in the
geometric design of horizontal curves, and the accident rate is directly related to it [13,14].
The most significant parameters of a curved path for this study are radius of curvature
and camber since these geometrical characteristics of a curved path influence the lateral
behavior. Due to these solicitations, vehicle driving can be difficult or even dangerous
when driving on a curved path. The vehicle is subjected to a lateral force caused by
centripetal acceleration acting on the center of gravity of each of the vehicle’s components.
Consequently, in order to achieve a balance, the adherence forces and the normal forces
between the road and the tire create a balance that prevents the vehicle from rollover or
sideslip. However, these adhesion forces have a limit given by Coulomb’s dry friction law
and another limit can also be considered when the normal force of any wheel reaches a
value of zero, the moment in which the vehicle rolls over. These limits are directly related
to the geometrical factors of the curve and to the environmental conditions that may cause
these limits to vary.
Therefore, the study of track characteristics in lateral vehicle dynamics is also of great
relevance. the safety issue of horizontal road curves was evaluated by [15] according to the
American Association of Highway Transportation Officials (AASHTO) standard. Different
parameters, such as vehicle weight, vehicle dimensions, longitudinal slopes, and vehicle
velocity are evaluated in the geometric design of horizontal curves using a multi-body
dynamic simulation process. A combination of simple circular and clothoid transition
curves with several longitudinal upslopes and downslopes were designed.
Ref. [16] studied the radius of curvature of the real vehicle trajectory under driver’s
instantaneous emergency steering maneuvers. It was calculated based on the bicycle model.
They analyzed the possibility of rollover taking into account the curve entry velocity and camber.
In this study, a trajectory composed of three paths (a straight path, a path of variable
radius of curvature and a path of constant curvature) is modeled. For simulation purposes,

362
Mathematics 2022, 10, 3081

only geometric variations are considered in order to obtain the values of the adherence and
normal forces present, while a vehicle follows a simulation trajectory composed of these
three paths with a different radius of curvature. This design fits to realistic transport routes
since it has been calculated and parameterized with curves of agreement (trajectories that
connect a straight and circular trajectory in such a way that it starts with an infinite radius
of curvature and ends in a constant radius of curvature) used in road design.
A bus model is provided using rigid multibody systems with which a virtual or
mathematical model (system of ordinary differential equations) has been generated. The
passenger bus is then modeled in two sections, front and rear. The front part contemplates
several subsystems, such as the unsprung masses, the unsprung masses are connected by
a stability or torsion bar, the flexible wheels, etc. These parts are represented by a set of
inertia, stiffness, and damping elements. The rear part has been modeled in an analogous
way although, unlike the front, in this there is only one unsprung mass and a Panhard
type bar is used to improve its vertical and lateral dynamic behavior. For this model, the
suspension system and tire stiffness as well as the damping of both bodies have been
modeled as linear components.
The Panhard bar is a suspension element that prevents lateral displacements of an axle
and also restricts the relative vertical movement between the suspended and unsuspended
mass. Independent rear suspension systems do not require a Panhard bar. The bar is
attached to two points that allow only the up and down movement of its ends, so that it
can only move in the vertical plane. The forces that restrict the movement between the
connected elements take part in the dynamics of these, since the forces and the consequent
moments that originate, both in the non-suspended mass and in the suspended, intervene
in their dynamics. Some previous studies have incorporated a Panhard bar in their vehicle
models. In [17], a roll plane model of a road vehicle is provided incorporating the kinematic
constraint provided by a Panhard bar. The results show that the location and orientation of
the Panhard bar significantly influences the kinetostatic roll properties of the suspension
when the vehicle is subjected to vertical and lateral forces. In [18], a common front axle
suspension is designed for four different tractor models, in which a Panhard bar is used to
improve safety and driving stability. Related with the Panhard bar [19], it compares the
structural durability behavior of commercial vehicles Panhard bars of a ferritic cast nodular
iron with the behavior of rods of an austempered.
Both sections of the bus are connected by a torsional spring (defined by its torsional
stiffness) to emulate the dynamic influence of the vehicle chassis. Due to the different
longitudinal position of the two sections and, consequently, the different roll angle, the
chassis exerts a torsional moment on both sections, which influences the lateral stability.
This bus model travels along the simulated trajectory and allows the simulation of the
lateral dynamics of a bus when it travels on a cambered curve.

2. Geometric Path Description

In Figure 1, a schematic model of the path considered for this work is shown.

z
D

Figure 1. Path’s geometric parameters.

363
Mathematics 2022, 10, 3081

These dimensions are subsequently applied to modeling of the road, with the radius
of curvature ρ and camber angle α as geometric parameters defining the dynamic behavior
of the vehicle on a curved path. The track’s width Bt and roadway width Bw define the
maximum value of the camber height and the vehicle width, respectively; the track width
also defines the number of lanes on the road. In order to simplify the simulation, it has
been assumed that the vehicle is driving in the middle roadway lane.

2.1. Kinematic Effects

The radius of curvature of a track section measures the magnitude of the curvature in
that section. According to the formulation deduced by dynamics, the normal acceleration
ay , caused by said curvature, is defined by the following equation:

ay = v2 /ρ (1)

This acceleration, mainly dependent on the speed (v) and inversely proportional to
the radius of curvature, is translated into an inertial force, opposite in sense to the normal
acceleration and oriented on the horizontal of the plane in which it is located. This inertial
force is commonly known as centrifugal force, Fi :

v2
Fi = mT (2)
ρ

where mT is the total mass of the vehicle.

The camber is defined as a relative elevation between the ends of the track, such
that a component of the vehicle’s weight compensates the inertial force caused by normal
acceleration; its length is measured between the ends of the track according to the vertical
direction (Figure 1).
Figure 2 shows a resultant force R as the vectorial sum of the weight and inertial force
(centrifugal force). As an effect of the inclination, there is a component of this resultant
force, which is perpendicularly pointed to the path’s plane R N , which keeps an adherence
force between the wheel and the path, and a perpendicular oriented component R T of the
resultant force R, which is responsible for the skidding and rollover of the vehicle; when
the resultant force R is completely oriented in the perpendicular direction of the path’s
plane, it is said that the vehicle is traveling at design velocity v D .

57

)L

K*
5
P7 J
51

\
&%

)I  1
)I 1

Figure 2. Force Schema for a vehicle traveling over a curved path.

For a skidding situation, it is said that the parallel component of R force is higher
than the adherence force determined by the Coulomb’s dry friction force; on the other

364
Mathematics 2022, 10, 3081

hand, when a rollover is produced, the vehicle will turn into the external side of the curve,
it means that it will turn over one or more wheel from the external side of the vehicle.
Consequently, we can affirm that when the normal force of one or more wheels in the inner
side are zero the rollover is produced, or is at least imminent. To analyze the vehicles
rollover, a fixed point is required thus the concept of roll center is introduced as a reference
point, which despite the strains that suspension suffers, it remains static on the vehicle
frontal plane. From this point, all moments are calculated so it is important to highlight
that the higher the vehicle’s mass center, the higher the possibility of rollover.

2.2. Kinds of Path

Three path types with different geometric characteristics (curvature radius and camber)
have been defined as shown in Table 1.
Table 1. Geometrical characteristic variations for type of path.

Type of Path Camber Angle Curvature Radius

Straight path None Infinite
Transition path Variable Variable
Circular path Constant Constant

The transition curves enhance the horizontal curve safety as they help to gradually
apply camber and centrifugal force. Since the transition path is the only one which varies
its parameters, it is the only one that defines parameter variation equations. Such paths
connect a straight and a circular path so that begins with an infinite curvature radius and
ends at a constant curvature radius; furthermore, over those paths, the camber angle starts
to increase or decrease. To describe those paths a variety of mathematical curves are applied
to determine their points in the space. Among those curves, the most known are Bernoulli’s
lemniscate and the clothoid, which is the most widely applied [15]. The use of a clothoid
transition curve between tangent and a circular curve has been suggested to increase the
safety margin of safety against sideslip, especially when the camber exceeds 12% [20]. So,
the clothoid has been used in this study.
A clothoid is defined by the following equation:
ρ ( s ) · s = A2 (3)

where s is the traveled distance over the curve and ρ(s) is the curvature radius as a function
of the traveled distance, it means that there is a linear correspondence through a constant.
This could be interpreted as while the vehicle is moving forward over the path, the magnitude
of the curvature radius increases or decreases. A clothoid curve is depicted in Figure 3.

Figure 3. Clothoid curve.

365
Mathematics 2022, 10, 3081

The position vector is described by the following equation which is expressed in terms
of Fresnel functions:
    
→ √ s s
r (s) = [ x (s); y(s)] = A π FresnelC √ ; FresnelS √ (4)
A π A π

One of the most important functions is avoid the appearance of sudden uncompen-
sated acceleration, thus it is required to evaluate the magnitude of the over acceleration,
which is the rate of variation of the acceleration. According to [21], for railway vehicles,
it is recommended that this value must not overcome 0.2 m/s3 ; nevertheless, for a road
vehicle, this value should be controlled by the maneuvering of the car’s rudder. For this
simulation, this value must be settled as an input parameter, hence the maneuvering of the
vehicles is not being simulated as it is a random parameter.

2.3. Simulation Path Parameterization

Once all geometric characteristics of the path are known, it is necessary to build,
in a parametric way, the simulation path over which the vehicle will travel during the
simulation. For that reason, the following equations are developed to describe the camber
and curvature radiuses over the clothoid:

A2
ρ(t) = (5)
v(t)t
zb
z(t) = v(t)t (6)
sb
In the previous expressions, zb is the constant camber value in the circular path and
sb is the total length of the clothoid. The velocity is going to be considered as a constant
value for simulation path generation purpose. Due to the vehicle’s behavior is studied
just a transition and circular path are required, hence in a straight path there is no risk of
rollover because lateral acceleration is not present.
From Equation (6) an equivalence to determine sb,min clothoid curve minimum length
can be calculated. According with [21], the maximum noncompensated lateral acceleration
should be 1 m/s2 for comfort and safety reasons. From the given conditions, the following
expression can be developed:
 √ 2 
9.81 zb v2 Bw −zb 2
Bw − ρmin Bw vD
< sb,min (7)
0.2
From this expression, the clothoid’s constant can be determined:
√
A= sb,min ρb (8)

Once the transition path is defined, a circular path can be added giving as a result the
whole simulation path.
For this work, a simulation path was developed considering the following parameters:
• Camber over the circular path (Final value): 0.25 m
• Curvature radius over the circular path (Final value): 400 m
• Track’s width: 0.9 m
• Number of tracks: 3
• Maximum over acceleration: 0.15 m/s3
• Path’s total length: 1500 m
• Average Speed: 120 km/h
Figure 4 shows a plane view of the track and the center of the circular path:

366
Mathematics 2022, 10, 3081

Figure 4. Simulation path—plane view.

Taking as a reference the average speed, function for camber angle and curvature
radius can be developed in terms of time (Figure 5).

D E

Figure 5. (a) Curvature radius vs. time and (b) camber angle vs. time for simulation.

3. Vehicle Modeling
As it has been described, to simplify the vehicle analysis, the bus has been divided
into two sections, front and rear, both are joined by a torsional spring, which emulates the
stiffness belonging to the bus structure.

3.1. Frontal Section

For the front section, the unsprung mass includes elements, such as the mass of the
tires and other components of the steering system, all of them are modelled as punctual
masses, which have only one degree of freedom on the perpendicular direction of path’s
plane. The Figure 6 shows the theoretical representation of the front section of the bus:

367
Mathematics 2022, 10, 3081

I
U

K* 

] P  , N V
K %
F.R.C F V 
F V P K (
\ N V N 7% ]
N Z
] P F Z
F Z
N Z

Bt
D

Figure 6. Bus front section—mathematical model.

The model proposed is based on [22,23]. This model considers not only the stiffness of
the wheels but also the damping of that element so that it is possible to get a more realistic
bus mechanical behavior. Additionally, a sway bar is considered as passive element, this
component has the functionality of link the vertical displacement of the unsprung masses. It is
performed taking advantage of the lever arms generated by the displacement giving as a result
a torsional moment along the bar so that restrict the movement of both unsprung masses, it is
desirable to avoid excessive vibrations, as well as a large displacement of the roll center.

3.2. Rear Section

In the rear section, the unsprung mass is modeled as a unique body, which is linked
with the bus chassis through the suspension system so that the unsprung mass has two
degrees of freedom, the vertical displacement, and a rotation around the longitudinal axis.
Most of the components are quite like the front section except for a rigid element known as
a Panhard bar which, as mentioned above, is used to avoid lateral displacement, and restrict
relative vertical displacement and rotation between the unsprung and sprung masses. In
terms of modeling, the Panhard bar is considered to be embedded in the unsprung mass
and articulated to the sprung mass.
A model of the rear section is depicted in Figure 7.

I
U

K* 
] P  ,  FV N V
F V K% 
N V I K( 
F Z
P  ,  N Z
F Z
]
N Z
\

Bt
D

Figure 7. Bus rear section—mathematical model.

368
Mathematics 2022, 10, 3081

Regarding the camber angle and curvature radius, the values of both parameters are
different from the front section, hence they are in different positions over the path; for that
reason, a time interval delay is introduced to simulate that effect.

3.3. Complete Model

From the aforementioned description, a torsional spring, with a torsional stiffness k E that
links both sections to emulate the stiffness of the bus structure chassis. Altogether, an 8 degree
of freedom model is composed by the two sections. The whole model is shown in Figure 8.

I

]
P  , 

NV  FV 
FV  N(
NV 
] I
P  , 
N Z FZ NZ
FZ 
I
]

P  ,

N V FV FV  NV 
]
P N7% P
N Z FZ FZ  ]
NZ

Figure 8. Complete bus model.

4. Kinetic Analysis of the Vehicle

4.1. Frontal Section
Hence, the forces on some vehicle components will be analyzed and a few parameters
will be modified to study the influence it has on behavior, the motion equations are going
to be determined by applying the D’Alembert principle. Firstly, equations belonging to the
suspended mass are obtained and then the equations corresponding to the unsprung masses.
From the geometric parameters of the road, which has been previously described, it is
possible to notice that the horizontal acceleration is the known centripetal acceleration pro-
duced by the vehicle’s travel over a curved road, therefore the only unknown acceleration
is vertical acceleration. Both accelerations can be decomposed into two orthogonal compo-
nents, one parallel and the other normal to the road plane (Figure 9). Transposing those
accelerations to the equivalent system and analyzing the so-called dynamic equilibrium,
motion equations are going to be obtained.

369
Mathematics 2022, 10, 3081

P \*
,I

\

P ]* D  I
]
Figure 9. Acceleration decomposition.

Suspended mass rotation, its angular speed and its vertical displacement are responsi-
ble for the forces on the suspension’s system components and each of them is linked to the
displacement of the correspondent unsprung mass connected to them. From those forces
and considering the center of gravity location, the moments around its roll center can be
calculated. In Figure 10, the front of the body diagram for dynamic equilibrium is shown.

E
E
P ]

7(VW K %
P \ K*  

)G  )V 
]
)V )G 
&  % )
Figure 10. Front cabin’s free body diagram for dynamic equilibrium.

From the free body diagram, the following differential equations are settled. For the
vertical displacement z1 :
.. . . .
!
m1 z1 = − k s1 (z1 + b1 φ1 − z2 ) + c1 (z1 + b1 φ1 − z2 )
. . .
!
− k s2 (z1 − b2 φ1 − z3 ) + c2 (z1 − b2 φ1 − z3 ) (9)
2
+m 1v
ρ (t)
sin[α1 (t) + φ1 (t)]
1

For rotation φ1 :
! .. . . .
!
I1 + m1 (hG1 − h B1 )2 φ1 = b2 k s2 (z1 − b2 φ1 − z3 ) + c2 (z1 − b2 φ1 − z3 )
. . .
!
−b1 k s1 (z1 + b1 φ1 − z2 ) + c1 (z1 + b1 φ1 − z2 )
 2
 ! (10)
− m1 ρ v(t) cos(α1 (t) + φ1 (t)) (hG1 − h B1 )
1
+[(m1 g sin α1 (α1 (t) + φ1 (t)))(hG1 − h B1 )] + Test

This equation considers a resistant torque, which is caused by the torsional stiffness
of the vehicle’s structure; said moment is denoted by Test , which will be determined later.
For the unsprung masses, the analysis is quite similar as that carried out for the suspended
mass, with the difference of the intervention of the torsion bar, which links the movements

370
Mathematics 2022, 10, 3081

of both suspended masses. Figure 11 shows the free body diagram and equivalent system
of both unsprung masses.

)V  )G  )%7
)V )G  )%7
{
{ P
P
)ZG  P ]* 
)ZG  P ]*  ] )Z 
]
)Z
Figure 11. Unsprung frontal masses free body diagram.

In Figure 11, it is possible to note that there is a force produced by the sway bar FTB ,
which is opposite at each unsprung mass. This element takes advantage of the lever arms
generated by each unsprung mass, which due to the displacement difference between the
ends of the bar generates a torsion moment.
The displacement can be calculated as follow:

F1 lb3
z2 = (11)
3Eb Ib

where F1 is the force applied by the left unsprung mass, lb is the arm bar’s length, Eb Young
Modulus of the bar material and Ib is the bar’s inertia moment. The force applied at each
end of the bar can be determined by the following equation:

3z2 Eb Ib
F1 (z2 ) = (12)
lb3

From Equation (12), a correlation between unsprung mass displacements and force in
the bar can be obtained. Figure 12 shows the free body diagram of the sway bar.

O7

* - R )

O%
)

Figure 12. Free body diagram—sway bar.

From the force calculated in (12), the torsion moment over the bar is calculated:

3z2 Eb Ib
Mt1 = F1 lb = (13)
lb2

In a similar way, it is possible calculate the moment generated by the unsprung right
mass as follow:
3z3 Eb Ib
Mt2 = F2 lb = (14)
lb2

371
Mathematics 2022, 10, 3081

Due to the moment exerted by both arms being opposite, the resultant torsion moment
is calculated from the difference between them:
3 Eb Ib
MTB = Mt1 − Mt2 = ( z2 − z3 ) (15)
lb2

The resultant torsion moment generates a rotation angle on the bar:

MTB lT
ΦB = (16)
G J0

This angle can be related to the displacement arm difference as follows:

3 Eb Ib lT
z Δ = zΔ = Φ lb = ( z2 − z3 ) (17)
lb G J0

From Equation (17) and due to the linear correlation between the displacement and
the unsprung mass displacements, an elastic constant is determined:

9 Eb2 Ib2 lT
k tb = (18)
lb4 G J0

Considering the influence of the sway bar and applying the acceleration decomposition
and transposing them to the first member of the equivalence (Dynamic equilibrium), the
following motion equations are developed:
Unsprung left mass:
.. 2 . . .
m2 z2 − m2 ρ v(t) sin[α1 (t) + φ1 (t)] = k s1 (z1 + b1 φ1 − z2 ) + c1 (z1 + b1 φ1 − z2 )
1 . (19)
−k TB (z2 − z3 ) − k w1 z2 − cw1 z2

Unsprung right mass:

.. . . .
m3 z3 = k s2 (z1 − b1 φ1 − z3 ) + c2 (z1 − b1 φ1 − z3 )
.
+k TB (z2 − z3 ) − k w2 z3 − cw2 z3 (20)
v2
+m3 ρ (t) sin[α1 (t) + φ1 (t)]
1

4.2. Rear Section

In a similar way as the front section, the free body diagram of the suspended mass for
dynamic equilibrium is depicted in Figure 13.

E
E

7(VW K% 
P \*  K*  
)3Y
] )G  )V 

)3K )G  P  J
)V 
& % 3
Figure 13. Rear cabin’s free body diagram for dynamic equilibrium.

372
Mathematics 2022, 10, 3081

In comparison with the front section, here there is a noticeable effect of the Panhard
bar, which is represented by the forces FPh and FPv , which are the horizontal and vertical
component of the force along the bar. Additionally, for the acceleration decomposition, the
influence of the rotation angle of both masses, unsprung and suspended, must be taken
into account, including the torsional torque produced by the structure. From this and the
free body diagram the following motion equations are obtained:
For the displacement z4
 .. 2

m4 z4 − ρ v(t) sin[α2 (t) + φ4 (t) + φ5 (t)] = −k s3 (z4 + b3 φ4 − z5 − b5 φ5 )
2
−k s4 (z4 − b4 φ4 − z5 + b6 φ5 ) (21)
. . . .
−cs3 (z4 + b3 φ4 − z5 − b5 φ5 )
. . . .
−cs4 (z4 − b4 φ4 − z5 + b6 φ5 ) + FPv

For the rotation φ4 :

! .. . . . . !
I4 + m4 (hG2 − h B2 )2 φ4 = b4 k s4 (z4 − b4 φ4 − z5 + b6 φ5 ) + cs4 (z4 − b4 φ4 − z5 + b6 φ5 )
. . . . !
−b3 k s3 (z4 + b3 φ4 − z5 − b5 φ5 ) + cs3 (z4 + b3 φ4 − z5 − b5 φ5 ) − FPv
 2
 !
− m4 ρ v(t) cos[α2 (t) + φ4 (t) + φ5 (t)] (hG2 − h B2 ) − Test
2
+[(m4 g sin[α2 (t) + φ4 (t) + φ5 (t)])(hG2 − h B2 )]
−[ FPh (hG2 − hS2 − h B2 )]
(22)
For the unsprung mass, the analysis is similar obtaining the free body diagram for the
dynamic equilibrium showed in Figure 14.
E
E )G  )V 
)3Y )3K

)G 
)V 
P \* 
)Z  )ZG  K( 
P J

] )Z )ZG 

&%3PQV

Figure 14. Rear unsprung mass free body diagram for dynamic equilibrium.

From the free body diagram, the following motion equations are obtained:
For the displacement z5
.. . .
m5 z5 = −k w3 (z5 + b5 φ5 ) − cw3 (z5 + b5 φ5 ) − k w4 (z5 − b6 φ5 )
. .
−cw4 (z5 − b6 φ5 ) + k s3 (z4 + b3 φ4 − z5 − b5 φ5 )
. . . . (23)
+k s4 (z4 − b4 φ4 − z5 + b6 φ5 ) + cs3 (z4 + b3 φ4 − z5 − b5 φ5 )
. . . . m5 v2
+cs4 (z4 − b4 φ4 − z5 + b6 φ5 ) − FPv + ρ2 ( t )
sin(α2 (t) + φ5 (t))

373
Mathematics 2022, 10, 3081

For the rotation φ5 :

⎡ . . ⎤
! .. k w4 (z5 − b6 φ5 ) + cw4 (z5 − b6 φ5 )
⎢ ⎥
I5 + m5 (h E2 )2 φ5 = b6 ⎣ −k s4 (z4 − b4 φ4 − z5 + b6 φ5 ) ⎦
. . . .
⎡ − c s4 ( z 4 − b4 φ 4 − z 5 + b 6 φ⎤5 ) + FPB
k s3 (z4 + b3 φ4 − z5 − b5 φ5 )
⎢ . . . . ⎥ (24)
−b5 ⎣ +cs3 (z4 + b3 φ4 − z5 − b5 φ5 ) ⎦
. .
−k w3 (z5 + b5 φ5 ) − cw3 (z5 + b5 φ5 )
(m5 g sin[α2 (t) + φ5 (t)])−
+h E2  2

m5 ρ v(t) cos[α2 (t) + φ5 (t)] + FPh
2

The forces produced by the Panhard bar require a kinematic analysis of both ends of it.
Such a component could be modeled as a bar element on which its inner force acts along
itself, therefore the force’s direction is already known.
Considering only the vertical displacement at the ends of the bar, Figure 15 shows the
displacement of each point connected to the Panhard bar:
E
E E
E ]  E I
I
KV
] I

]
]  E I

Figure 15. Rear masses displacements at the ends connected to the Panhard Bar.

From that displacement, the bar vertical displacement can be analyzed as it is shown
in Figure 16.

I
3'
]   E
I
K ]   E

%W

Figure 16. Panhard Bar vertical displacement.

From this analysis, the vertical dimension of the bar can be calculated as follow:

h = h0 + z5 − b6 φ5 − z4 − b3 φ4 (25)

With the vertical dimension determined, the bar’s length can be calculated directly:
#
 PD = h2 + Bt2 (26)

374
Mathematics 2022, 10, 3081

Additionally, its orientation:

 
Bt
ϕ B = arctan (27)
h0 + z5 − b6 φ5 − z4 − b3 φ4

Later, the bar stiffness can be calculated instantly as follow:

EA
kP = (28)
 PD

where E is the material’s Young Modulus and A its transversal area.

As a final step the influence of the torsional stiffness of the structure is seen on the
motion equations, so the torsional moment originated this element is equivalent to the
following expression:
Test = k E (φ4 − φ1 ) (29)
This component links both bus sections, so all differential equations developed are coupled.

5. Simulation and Results

Due to the complexity of the model, a state space model must be implemented, so that
the motion equations settled previously can be arranged as a differential equation system,
which can be divided into two matrices, one as a state matrix A which is entirely lineal, and
an input matrix B, which due to the Panhard bar and its dependency to the variable states
is highly non-lineal.
From the state space equation:

dx = Ax + Bu
(30)
y = Cx + Du

The state vector is defined as follow:

. . . . . . . . !
x = z1 z1 z2 z2 z3 z3 φ1 φ1 z4 z4 z5 z5 φ4 φ4 φ5 φ5 (31)

Due to its dimensions and simplicity, the state matrix which is a 16 × 16 square
matrix will not be shown on this document, but the input matrix will be presented for its
non-linear formulation:
⎛ ⎞
0
⎜ v2
sin(α(t) + φ1 ) ⎟
⎜ ρ(t) ⎟
⎜ ⎟
⎜ 0 ⎟
⎜ ⎟
⎜ v2
sin α ( α ( t ) + φ ) ⎟
⎜ ρ(t) 1 ⎟
⎜ ⎟
⎜ 0 ⎟
⎜ ⎟
⎜ v2
sin α ( α ( t ) + φ ) ⎟
⎜ ρ(t) 1 ⎟
⎜ ⎟
⎜ 0 ⎟
⎜   ⎟
⎜ 2
− m1 ρv(t) cos(α(t)+φ1 ) (hG1 −h B1 )+(m1 g sin(α(t)+φ1 ))(hG1 −h B1 ) ⎟
⎜ ⎟
⎜ I1B ⎟
⎜ ⎟
⎜ 0 ⎟ (32)
⎜ 2 ⎟
⎜ v
sin ( α ( t − 1 ) + φ + φ ) + FPv ⎟
⎜ ρ ( t −1) 4 5 m 4 ⎟
⎜ ⎟
⎜ 0 ⎟
⎜ 2 FPv ⎟
⎜ v
sin ( α ( t − 1 ) + φ ) − ⎟
⎜ ρ ( t −1) 5 m 5 ⎟
⎜ 0 ⎟
⎜⎛ ⎞ ⎟
⎜ 2 ⎟
⎜ ⎝ −m4 ρ(t−1) cos(α(t − 1) + φ4 + φ5 ) ⎠
v
⎟
⎜ (hG2 −h B2 )− FPh (hS2 )+ FPv b3 ⎟
⎜ +m4 g sin(α(t − 1) + φ4 + φ5 ) ⎟
⎜ ⎟
⎜ I4B ⎟
⎜ ⎟
⎜ 0 ⎟
⎝ 
v2
 ! ⎠
(m5 g sin(α(t−1)+φ5 ))− m5 ρ(t−1) cos(α(t−1)+φ5 ) + FPh h E2 + FPv b6
I5B

375
Mathematics 2022, 10, 3081

From the equations given in this matrix, it is highly dependent on the state matrix so
that matrix B is a function of x. Although there is a control vector u and matrix C, output,
and matrix D, feedforward, this matrix will be replaced by an identity matrix and a row
vector with ones, hence the main purpose of the settled system is simulation. In further
works, this matrix can be replaced to obtain a control system. All equations for the path
and other elements, such as a Panhard bar, are modeled in a Simulink® algorithm.
Once the differential equations are expressed into the state space equation, it is neces-
sary to numerically fix all the mechanic parameters which have been taken from [24] are
expressed in the following list:
m1 = 7000 kg—Front suspended mass
m2 = 500 kg—Front left unspring mass
m3 = 500 kg—Front right unspring mass
I1 = 125, 000 kg · m2 —Front suspended mass inertia respect to its mass center
c1 = 600 N · s/m—Damping coefficient of the left front suspension element
c2 = 600 N · s/m—Damping coefficient of the right front suspension element
b1 = 0.8 m—Distance from suspended mass center to the left front suspension
b2 = 0.8 m—Distance from suspended mass center to the right front suspension
k s1 = 200, 000 N/m—Stiffness coefficient of the left front suspension element
k s2 = 200, 000 N/m—Stiffness coefficient of the right front suspension element
h B1 = 0.75 m—Roll center front suspended mass height respect to the path’s plane.
hG1 = 1.55 m—Mass center height of the suspended front mass
k TB = 1, 000, 000 N/m—Stiffness coefficient of the sway bar
k w1 = 921, 607 N/m—Stiffness coefficient of the front left tire
k w2 = 921, 607 N/m—Stiffness coefficient of the front right tire
cw1 = 800 N · s/m—Damping coefficient of the front left tire
cw2 = 800 N · s/m—Damping coefficient of the front right tire
k E = 600, 000, 000 N · m/rad—Torsional stiffness coefficient of the vehicle structure
m4 = 7000 kg—Read suspended mass
m5 = 1000 kg—Rear left unspring mass
I4 = 125, 000 kg · m2 —Rear suspended mass inertia respect to its mass center
I5 = 1000 kg · m2 —Rear unspring mass inertia respect to its mass center
c3 = 600 N · s/m—Damping coefficient of the left rear suspension element
c4 = 600 N · s/m—Damping coefficient of the right rear suspension element
b3 = 0.8 m—Distance from suspended mass center to the left rear suspension
b4 = 0.8 m—Distance from suspended mass center to the right rear suspension
b5 = 0.8 m—Distance from unspring mass center to the left rear tire
b6 = 0.8 m—Distance from unspring mass center to the right rear tire
k s3 = 200, 000 N/m—Stiffness coefficient of the left rear suspension element
k s4 = 200, 000 N/m—Stiffness coefficient of the right rear suspension element
h B2 = 0.75 m—Roll center rear suspended mass height respect to the path’s plane.
hG2 = 1.55 m—Mass center height of the suspended rear mass
h E2 = 0.5 m—Mass center height of the unspring rear mass
k w3 = 921, 607 N/m—Stiffness coefficient of the rear left tire
k w4 = 921, 607 N/m—Stiffness coefficient of the rear right tire
cw3 = 800 N · s/m—Damping coefficient of the rear left tire
cw4 = 800 N · s/m—Damping coefficient of the rear right tire
Also, for the Panhard bar, the following parameters are taken:
E = 2.1 × 105 N/mm2 —Panhard bar material’s Young modulus
d P = 30 mm—Panhard bar diameter
From the constants given above, the following results are obtained from the simulation
of the vehicle’s travel over a transition and circular path, as previously given (Figure 17).

376
Mathematics 2022, 10, 3081

Figure 17. Simulation results—Displacement and rotation angles.

Furthermore, the instantaneous Panhard’s bar force and angle are also obtained from
the simulation (Figure 18).

377
Mathematics 2022, 10, 3081

Figure 18. Panhard bar—force and inclination angle.

Those results show the bus dynamic behavior when it is traveling over a transition and
circular curve; in fact, it is noticeable that while the bus is over the clothoid its position and
rotation state variables are increasing constant and when it arrives at the circular path they
become oscillating, consequently, the Panhard bar length and angle have the same behavior.

5.1. Effects of the Panhard Bar

The addition of the Panhard bar has a stabilization function on the mechanical behavior
of the bus, the main effect is appreciated on the rear mass rotation angles, which have
an effect on the displacements and angles in the front section. It can be said that the
Panhard bar shrinks the oscillations of each state variable as can be appreciated, making a
comparison between Figures 17 and 19.

Figure 19. Simulation results without Panhard bar—displacement and rotation angles.

378
Mathematics 2022, 10, 3081

5.2. Panhard Bar Influence over the Structural Torsional Moment

As can be observed in the rotation angles comparison, there is an increment in the
values in the system without a Panhard bar compared with the system with a Panhard bar.
In fact, this tendency makes a change in the sense of the torsional moment produced along
the structure of the bus, as can be seen in Figure 20.

Figure 20. Structural Torsion with and without Panhard bar.

The positive values of the structural moment present in the simulation with the
Panhard bar show a better stability due to this moment acts in opposition to the vehicles
rollover making the vehicle more stable; on the other hand, the absence of a Panhard bar
produces a negative structural moment which contributes to an increased risk of rollover.

5.3. Vehicle Lateral Stability Analysis

Although these results describe how the vehicle will behave, it does not show if the
vehicle rolls over or slips; for that reason, it is necessary to handle the obtained results and
the states derivatives in a new analysis that evaluates the normal and friction force on each
tire. For this purpose, a new free body diagram is performed for the dynamic equilibrium
in the front section (Figure 21).

P ]
)OL
:O
:Y K*  ] 

, %I P ] 
P ]  :O  )O L K(  ] 

:O  )O L
:Y 
:Y  )I 
1
K(  ]  1
)I

Figure 21. Free Body Diagram for front section dynamic equilibrium.

On the diagram shown, the inertial forces, which are obtained from the acceleration’s
decomposition used for the motion equations, and weight of each body is already decomposed.

379
Mathematics 2022, 10, 3081

The inertia forces on the vertical direction are defined using the states derivatives,
however, the lateral inertia forces and the weight decomposition are calculated as follow:

m1 v2
Fl1i = cos[α1 (t) + φ1 (t)] (33)
ρ1 ( t )

m2 v2
Fl2i = cos[α1 (t)] (34)
ρ1 ( t )
m3 v2
Fl3i = cos[α1 (t)] (35)
ρ1 ( t )
Analyzing the equilibrium on the lateral direction:

Ff 1 + Ff 2 = Fl1i + Fl2i + Fl3i − Wl1 − Wl2 − Wl3 (36)

Furthermore, the inertial forces and torques required to be used into the state equations
to calculate the normal forces at each tire. Analyzing the equilibrium taking moments from
the left frontal wheel and factorizing N2 :
..
[Wl2 − Fl2i ](h E1 − $z2 ) + [Wl3 − F
%l3i ](h E1 − z3$) −..I1B φ1 %
.. /( Bt ) (37)
+[Wl1 − Fl1i ](hG1 − z1 ) + m1 z1 − Wv1 ( Bvia /2) + m3 z3 − Wv3 ( avia )

Analyzing the equilibrium on the vertical direction, N1 is obtained:

.. .. ..
N1 = Wv1 + Wv2 + Wv3 − N2 − m1 z1 − m2 z2 − m3 z3 (38)

The magnitudes of the normal forces indicate if the vehicle rolls over or not when
any normal force becomes zero. On the other hand, in the lateral direction there are two
unknown variables, which are the friction force and just one equation, so that determines
the value of each friction force becomes an unsolvable problem, therefore a new procedure
to evaluate if the vehicle slips or not will be performed, using the value of both forces in a
conditional way:
Ff 1 + Ff 2 ≤ Ff 1,max + Ff 2,max (39)
In addition, it can be expressed as:

0 ≤ N1 μs + N2 μs − Ff 1 − Ff 2 (40)

The maximum values of friction force will be determined using the so-called Coulomb’s
dry friction approach. For that evaluation, it is assumed that both tires are not skidding, so
the friction coefficient is the static friction coefficient. The conditional form establishes that
when the value becomes negative it means that the vehicle is slipping. This analysis is also
applied for the rear section, with a similar free body diagram (Figure 22).

380
Mathematics 2022, 10, 3081

P \  )O L

:O 
:Y 
]
K*  
, %  I

P \  )O L
]
K(  
:O 
:Y 

)I 

, %  I 1

)I  1

Figure 22. Free Body Diagram for rear section dynamic equilibrium.

The equations for the lateral inertia forces:

m4 v2
Fl4i = cos[α2 (t) + φ4 (t) + φ5 (t)] (41)
ρ2 ( t )

m5 v2
Fl5i = cos[α2 (t) + φ5 (t)] (42)
ρ2 ( t )
Analyzing the equilibrium on the lateral direction:

Ff 3 + Ff 4 = Fl4i + Fl5i − Wl4 − Wl5 (43)

Similar to the front section, analyzing the equilibrium taking moments from the left
wheel and factorizing N4 :
.. ..
[Wl5 − Fl5i$](h E2 − z5 ) −
% IB4 φ4 − IB5
$ φ5 ..+ [Wl4 −%Fl4i ](hG2 − z4 )
N4 = − .. /Bt (44)
+ m4 z4 − Wv4 ( Bvia /2) + m5 z5 − Wv5 ( Bt /2)

Analyzing the equilibrium on the vertical direction, N1 is obtained:

.. ..
N3 = WV4 + WV5 − N4 − m4 z4 − m5 z5 (45)

Finally, the skidding condition is introduced for the rear section:

Ff 3 + Ff 4 ≤ Ff 3,max + Ff 4,max (46)

In other terms:
0 ≤ N3 μs + N4 μs − Ff 3 − Ff 4 (47)
For the simulation, a parameter more is necessary:
μk = 0.4—Static friction coefficient between the tire and path.
The results for the normal forces at each tire, while the bus is traveling over the
simulation path, are shown on the following graphics (Figure 23).

381
Mathematics 2022, 10, 3081

Figure 23. Normal forces on each tire considering Panhard bar.

About the friction force, the skidding condition shows the remaining friction force so
that when it becomes negative the vehicle slips. Figure 24 shows the remaining friction
force at each section, while the bus is traveling over the simulation:

Figure 24. Remaining friction force at each section.

A comparison of the normal force can be made using a model without a Panhard bar
to analyze the effect of this element in the lateral stability. The normal force values of the
bus without Panhard bar are presented in Figure 25.

382
Mathematics 2022, 10, 3081

Figure 25. Normal forces on each tire without Panhard bar.

From Figure 25, it is possible to observe that the normal force in the left side reach or
even pass negative values, which implies that the vehicle loses adherence with the path. It
demonstrates that the Panhard bar improves the adherence between the path and the tires.

6. Conclusions
A model of a bus is presented, elaborated by means of rigid multibody systems with
which a virtual or mathematical model (system of ordinary differential equations) has been
generated. This model allows the simulation of the lateral dynamics of a bus when it travels
on a cambered curve.
Based on the proposed model, a simulation algorithm has been designed to analyze
the dynamic behavior of a bus in transit on a cambered curve. This model is adjusted to
realistic transport routes since it has been calculated and parameterized with curves of
agreement (trajectories that connect a straight and circular trajectory in such a way that it
starts with an infinite radius of curvature and ends in a constant radius of curvature) used
in road design.
The modeling has been carried out in such a way that the total system representing
the bus has been divided into two fundamental parts: the front part of the vehicle, which in
turn contemplates several subsystems, such as the unsprung masses, the unsprung masses
connected by a stability or sway bar, the flexible wheels, etc. In turn, the rear part has
been modeled in an analogous way although, unlike the front part, in this one there is
only one unsprung mass and a Panhard type bar is used to improve its vertical and lateral
dynamic behavior.
Therefore, the model used to model the bus dynamic behavior integrates two modi-
fiable sections that allow for testing some new models for the calculation of the stiffness
of the suspension system or of the tires, etc. For this first approximation, the suspension
system and tire stiffness as well as the damping of both bodies have been modeled as linear
components. This is responsible for the quasi—linear correlation between the geometric
parameters and the tendency shown in the results: the state variables increase as the geo-
metric parameters of the trajectory increase and oscillate, trying to reach a steady state in
time while the trajectory parameters are constant.
According to the comparison between the results with and without a Panhard bar, it
can be stated that the Panhard bar has a stabilization function, which reduces the oscillations

383
Mathematics 2022, 10, 3081

of each state variable. Additionally, the parameters which modify the Panhard bar stiffness
can be treated as a control variable on an active suspension system.
From the effects on the rotation angles that the Panhard bar has, it is also possible to
highlight that the presence of this element reduces the rotation angles of the mass linked to
it, this improves the lateral stability of the vehicle, increasing the value of the normal forces
between the tires and the road, as can be concluded by comparing Figures 23 and 25.
The torsional structural stiffness is a parameter that must be considered when the
stability of the vehicle is analyzed, as presented in Figure 20, where the direction of the
structural moments affects the vehicle lateral stability.
Due to the weak effect of the non-linearities, some state variables have some distur-
bances which can be treated as noise.

Author Contributions: Conceptualization, E.O., J.R.H. and V.D.; methodology, E.O., E.R.C.L., J.R.H.
and V.D.; software, E.R.C.L.; formal analysis, E.O., E.R.C.L., J.R.H. and V.D.; investigation, E.O.,
E.R.C.L., J.R.H. and V.D.; resources, E.O., E.R.C.L., J.R.H. and V.D.; data curation, E.O., E.R.C.L., J.R.H.
and V.D.; writing—original draft preparation, E.O. and E.R.C.L.; writing—review and editing, E.O.,
E.R.C.L., J.R.H. and V.D.; visualization, E.O., E.R.C.L., J.R.H. and V.D.; supervision, E.O. and V.D.;
funding acquisition, V.D. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Yang, S.; Chen, L.; Li, S. Dynamics of Vehicle-Road Coupled System; Springer: Berlin/Heidelberg, Germany, 2015.
2. Popp, K.; Schiehlen, W. Ground Vehicle Dynamics; Springer: Berlin/Heidelberg, Germany, 2010.
3. Rill, G.; Schaeffer, T. Grundlagen und Methodik der Mehrkörpersimulation—Vertieft in Matlab-Beispielen, Übungen und Anwendungen;
Springer: Berlin/Heidelberg, Germany, 2014.
4. Schramm, D.; Hiller, M.; Bardini, R. Vehicle Dynamics—Modelling and Simulation; Springer: Berlin/Heidelberg, Germany, 2014.
5. Jeong, D.; Ko, G.; Choi, S.B. Estimation of sideslip angle and cornering stiffness of an articulated vehicle using a constrained
lateral dynamics model. Mechatronics 2022, 85, 102810. [CrossRef]
6. Seyedi, M.; Jung, S.; Wekezer, J. A comprehensive assessment of bus rollover crashes: Integration of multibody dynamic and
finite element simulation methods. Int. J. Crashworthiness 2020, 27, 273–288. [CrossRef]
7. Niculescu-Faida, O.-C.; Niculescu-Faida, A. Vehicle dynamics modeling during moving along a curved path. Mathematical
model usage on studying the robust stability. UPB Sci. Bull. 2008, 4, 49–60.
8. Gauchia, A.; Olmeda, E.; Aparicio, F.; Díaz, V. Bus mathematical model of acceleration threshold limit estimation in lateral
rollover test, Veh. Syst. Dyn. 2011, 49, 1695–1707. [CrossRef]
9. Gauchia, A.; Díaz, V.; Boada, M.J.L.; Olatunbosun, O.; Boada, B.L. Bus Structure Behaviour under Driving Manoeuvring and
Evaluation of the Effect of an Active Roll System. Int. J. Veh. Struct. Syst. 2010, 2, 14–19. [CrossRef]
10. Eftekharzadeh, S.F.; Khodabakhshi, A. Safety evaluation of highway geometric design criteria in horizontal curves at downgrades.
Int. J. Civ. Eng. 2014, 12, 326–332.
11. Yin, Y.; Wen, H.; Sun, L.; Hou, W. The Influence of Road Geometry on Vehicle Rollover and Skidding. Int. J. Environ. Res. Public
Health 2020, 17, 1648. [CrossRef] [PubMed]
12. Wang, B.; Hallmark, S.; Savolainen, P.; Dong, J. Crashes and near-crashes on horizontal curves along rural two-lane highways:
Analysis of naturalistic driving data. J. Saf. Res. 2017, 63, 163–169. [CrossRef] [PubMed]
13. Geedipally, S.R.; Pratt, M.P.; Lord, D. Effects of geometry and pavement friction on horizontal curve crash frequency. J. Transp.
Saf. Secur. 2019, 11, 167–188. [CrossRef]
14. Bogenreif, C.; Souleyrette, R.R.; Hans, Z. Identifying and measuring horizontal curves and related effects on highway safety. J.
Transp. Saf. Secur. 2012, 4, 179–192. [CrossRef]
15. Abdollahzadeh Nasiri, A.S.; Rahmani, O.; Abdi Kordani, A.; Karballaeezadeh, N.; Mosavi, A. Evaluation of Safety in Horizontal
Curves of Roads Using a Multi-Body Dynamic Simulation Process. Int. J. Environ. Res. Public Health 2020, 17, 5975. [CrossRef]
[PubMed]
16. Xu, J.; Xin, T.; Gao, C.; Sun, Z. Study on the Maximum Safe Instantaneous Input of the Steering Wheel against Rollover for Trucks
on Horizontal Curves. Int. J. Environ. Res. Public Health 2022, 19, 2025. [CrossRef] [PubMed]

384
Mathematics 2022, 10, 3081

17. Yin, Q.; Stiharu, I.; Rakheja, S. Kinetostatic analysis of a beam-axle suspension with Panhard rod restraining linkage. Int. J. Veh.
Des. 1998, 19, 108–123.
18. Mayr, S.; Wagner, W. Development of a front axle suspension for special purpose tractors. In Proceedings of the Conference:
Agricultural Engineering 2004, Dresden, Germany, 7–8 October 2004.
19. Sonsino, C.M.; Streicher, M. Optimization of cast iron safety components of commercial vehicles by material selection and
geometry under consideration of service loadings. Mater. Test.-Mater. Compon. Technol. Appl. 2009, 51, 428–436.
20. Torbic, D.J.; O’Laughlin, M.K.; Harwood, D.W.; Bauer, K.M.; Bokenkroger, C.D.; Lucas, L.M.; Ronchetto, J.R.; Brennan, S.; Donnell,
E.; Brown, A.; et al. NCHRP 774 Report: Superelevation Criteria for Sharp Horizontal Curves on Steep Grades; The National Academie
Press: Washington, DC, USA, 2014.
21. Esveld, C. Modern Railway Track; Delft University, MRT—Productions: Zaltbommel, The Netherlands, 2001.
22. Rajamani, R. Vehicle Dynamics and Control; Mechanical Engineering Series; Springer: Berlin/Heidelberg, Germany, 2012.
23. Aparicio, F.; Vera, C.; Díaz, V. Teoría de los Vehículos Automóviles; Sección de Publicaciones, ETSII, Universidad Politécnica de
Madrid: Madrid, Spain, 1995.
24. Mustafa Siddiqui, O. Dynamic Analysis of a Modern Urban Bus for Assessment of Ride Quality and Dynamic Wheel Loads.
Master’s Thesis, Concordia University, Montreal, QC, Canada, 2000.

385
 MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel. +41 61 683 77 34
Fax +41 61 302 89 18
www.mdpi.com

Mathematics Editorial Office

E-mail: mathematics@mdpi.com
www.mdpi.com/journal/mathematics
MDPI
St. Alban-Anlage 66
4052 Basel
Switzerland
Tel: +41 61 683 77 34
www.mdpi.com ISBN 978-3-0365-5924-7