Applied Mathematical Modelling 54 (2018) 551–566

Contents lists available at ScienceDirect

Applied Mathematical Modelling

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

A fuzzy α -cut optimization analysis for vibration control of

laminated composite smart structures under uncertainties
Marcos D.F. Awruch a,∗, Herbert M. Gomes b
a
Mechanical Engineering, Federal University of Santa Maria, Av. Roraima 1000, 97105-900 Santa Maria, RS, Brazil
b
Mechanical Engineering, Federal University of Rio Grande do Sul, Rua Sarmento Leite 425, 90050-170 Porto Alegre, RS, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: As a high performance and specific mechanically tailored material, composite structures
Received 16 March 2017 present high structural behavior sensitivity to small variations in geometry or material
Revised 5 September 2017
properties. Some of these uncertainties are intrinsically related to the manufacturing pro-
Accepted 1 October 2017
cess but others are usual uncertainty in material properties. A fuzzy interval analysis
Available online 10 October 2017
methodology is proposed in this paper in order to evaluate the structural behavior of lam-
Keywords: inated composite smart structures under vibration control and uncertain parameters. A
Vibration control Particle Swarm Optimization algorithm is used to search for input parameters in order to
Smart structures minimize and maximize system outputs by an anti-optimization and thus build their en-
Possibilistic theory velopes. Each input parameter has interval values defined by α -cut levels and interpolated
Fuzzy interval analysis by the Fuzzy reasoning. This methodology is applied to a smart structure of laminated
Uncertainties quantification composite material with attached piezoelectric patches and controlled by an optimal con-
Laminated composite trol regulator. System’s interval outputs like natural frequency, mechanical vibration, and
electric control input energy are investigated taking into account uncertainties in the com-
posite and piezoelectric material properties, ply angles and layer thickness. Finally, it is
concluded that these uncertainties may affect the control performance since it was found
significant modifications in the dynamic behavior and therefore this should be accounted
in the design stages.
© 2017 Elsevier Inc. All rights reserved.

1. Introduction

In the past few years, substantial attention has been paid to active vibration control of smart and lightly damped flexible
structures in several fields of civil, mechanical and aerospace engineering. The applications extend from vibration mitigation
on tower structures and motion control of robotic systems to satellite solar panels and energy harvesting devices. In order
to satisfy precision control and lightweight requirements, smart materials such as piezoelectric and shape memory alloys
are frequently integrated into laminated composite structures as sensors or actuators. The use of piezoceramic material is a
field with ongoing investigation and application and its advantages include low-power consumption, fast response time, a
wide variety of shapes and sizes and easy implementation.
An important step related to any structural modelling is the uncertainties identification and quantification present in
the actual system, taking it into account in the analysis stage. The two major types of uncertainties are classified as
aleatory and epistemic uncertainties. The aleatory uncertainty is generally modelled as probabilistic variables with statistical

∗
Corresponding author.
E-mail addresses: marcos.awruch@ufsm.br (M.D.F. Awruch), herbert@mecanica.ufrgs.br (H.M. Gomes).

https://doi.org/10.1016/j.apm.2017.10.002
0307-904X/© 2017 Elsevier Inc. All rights reserved.
552 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

information assumed based on experimental data (loads, boundary conditions, material property variability). On the other
hand, epistemic uncertainty comes from the lack of data information (statistical parameter uncertainty), imperfect mod-
elling, simplified assumptions or lack of awareness of the phenomena being modelled.
Laminated composite structures are known for their challenges during the fabrication processes that might generate un-
certainties in some properties or even defects such as interlaminar voids, fiber misalignment, residual stresses, variation
in ply thickness, just to name a few [1,2]. To deal with the uncertainty in a project, it is possible to use a probabilistic
approach, but in that case, it is desirable to have enough and reliable information about the random variables, such as
mean values, moments and distribution types [3]. Usually, in many engineering applications, there are not enough mea-
surements or knowledge of some parameters, or even they are measured with not sufficient accuracy. When statistical data
cannot be obtained or the information is imprecise, a possibilistic approach is preferable. Possibilistic methods deal with
the extreme scenarios, or the problem’s output envelopes, giving neither information about their probabilistic distribution
nor their correlations. The mathematical framework using this methodology includes Dempster–Shafer evidence theory [4].
Moreover, it is well proven the efficacy in founding extreme bounds by interval analysis in comparison with simple Monte
Carlo simulations [4–6].
According to Potter [7], even for statistical data given by suppliers (e.g. mass/unit area of a prepreg composite), the
measured probability density function does not meet usual specifications (showing even bimodal shape). According to Roy et
al. [8], the development of the Design Allowable Database requires an evaluation of multiple batches of composites with the
associated construction of very large mechanical and other physical property databases. In addition, to quantify uncertainty
in the structural performance, (1) numerous subcomponents may have to be fabricated requiring expensive tooling, and (2)
one must then perform expensive tests on these elements to determine their long-term performance. Furthermore, interval
values can be easily treated numerically. For convergence of statistics using MC simulations framework, several numerical
simulations are necessary that is opposite to the most accurate and less numerically processing and time-consuming results
obtained using a Fuzzy or Possibilistic framework. This is the main motivation behind this research.

2. Brief bibliographical review

The problem that arises in designing composites components with associated uncertainty is well posed by Antonio [9]. A
comprehensive survey of the techniques used for sensitivity evaluation and the uncertainty propagation is presented in this
paper but limited to sensitivity-based methods. It is suggested that the effects of uncertainty in mean values of mechanical
properties of composites are very sensitive for reliability index evaluations.
There are some new studies where the focus in the analysis of uncertainties in composites, such as [10], makes use of
the first order reliability method (FORM) with a high fidelity shear deformable laminated model, considering uncertainties
associated with fiber orientation and ply thickness. Lopez et al. [11] present a comparison between the FORM and the poly-
nomial chaos representation for the reliability analysis, where loads, strength properties and fiber orientation angles for each
layer are considered as random variables. In Goyal and Kapania [12] paper, angles and properties of the laminate are con-
sidered as having an uncertainty degree, so the reliability problem is analyzed based on these uncertainties. There are also
experimental studies trying to take into account sources of uncertainties, as Lekou et al. [13] study where they try to ob-
tain the estimation of measurement uncertainties in the properties of composite materials. A good survey on methodologies
used for uncertainty evaluation in composite structures can be found in Sing and Grover [14].
Zhang et al. [15] propose an interval Monte Carlo method for structural reliability evaluation and report that the Interval
Probability of Failure obtained using the interval MC tends to be wider than that from a Bayesian approach, which incorpo-
rates the probability information into the confidence bounds. Chowdhury and Adhikari [16] proposed a High-Dimensional
Model Representation (HDMR) metamodel for a dynamic analysis with fuzzy uncertain variables and validated the results
that are compared to a direct Monte Carlo Simulation. This approach presents an accurate estimation of the system response
and also a lower computational effort than a direct complete simulation. Chakraborty and Sam [17] made a relevant study
on methods transforming possibilistic variables to equivalent probabilistic variables, enabling a hybrid uncertainty approach
to different problems.
The vibration control performance of structures under the presence of uncertainty is an important field of study, for
example in earthquake isolation systems in civil engineering [18,19] or in dynamic control in aerospace applications [20],
among others. There are few studies related to the performance of piezoelectric controllers for composite materials under
uncertainty properties. Most of them are related to the use of Robust Control (H∞ norm) and LMI (Linear Matrix Inequali-
ties). Koroishi et al. [21] present a robust control applied to a piezoelectric actuator bonded to a composite structure based
on a Linear Matrix Inequalities applied to a Linear Quadratic Regulator. The inclusion of uncertainty in the analysis is per-
formed by introducing a variation of 10% in the dynamic matrix of the structural model (but unfortunately, not assigning
the sources of such uncertainty). The statistics of 100 samples are generated using Monte Carlo Simulation associated to
Latin Hypercube sampling. Envelope curves are presented in order to evaluate the performance of the LMI/LQR controller.
In the case of structural uncertainty, it is reported that the conventional robust control outperformed the proposed LMI/LQR
control. Trevilato [22] uses a similar approach based on LMI, but in his work, the way the uncertainty is introduced in the
model is different: the ply angle orientation and the stiffness of the finite elements attached to the clamped end of a can-
tilever beam are defined as the only sources of uncertainty. The tested control techniques are the traditional LQR and the
 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566 553

H∞ robust control. The expected better performance of the H∞ control compared to the LQR in the presence of uncertainties
was reported.
Choi [23] investigates the effects of actuator uncertainty in the vibration control of a smart beam structure using a
Robust Control incorporating the hysteretic behavior of the actuator and parameter variations for the system uncertainty. It
was reported that the formulation was experimentally verified and presented good performance. Marinova et al. [24] simu-
late damage in a smart composite beam using the Linear Fractional Transformation (LFT) implemented in H∞ control theory.
The LFT assumes a feedforward using a Gain matrix of uncertainty, bounded to the unitary norm that relates the system
outputs to a new control force that comes from a different control channel of the system. The uncertainty is treated as dif-
ferential mass, damping and stiffness matrices in the LFT framework. Based on the simulated numerical results a suboptimal
H∞ control law is implemented and they conclude the vibration suppressions are effective in different loading situations.
In a different way, Moutsopolou et al. [25] modelled uncertainties in the whole mass, stiffness and damping matrices as
variations around mean values and an H∞ criterion, using μ analysis to take into account the worst scenario for uncertain
disturbances and noise in the system. Unfortunately, there is no assignment for the source of the uncertainty present in the
variations of the structural matrices.
In this work it is applied a fuzzy α -cut optimization methodology, which can be described as a multiple interval analysis
process, associated with a heuristic algorithm working as anti-optimization in order to search the output boundaries for an
example case, studying the uncertainty propagation. No papers were found that deal with control performance of piezo-
laminated composites under uncertainties with possibilistic parameters and a heuristic approach. The PSO (Particle Swarm
Optimization) algorithm is used as the tool for the minimum and maximum boundary search. Herein the finite element
model, based on the First-Order Shear Deformation Theory (FSDT), is a laminated composite plate with embedded piezo-
electric actuators and sensors controlled by Linear Quadratic Regulator/Gaussian (LQR/LQG) controllers. Multiple sources of
uncertainties are considered such as thickness, ply orientation and material proprieties for the composite and piezoceramic
components. The analyzed output performance parameter is the integral over time of the kinetic and potential energy of
the controlled beam, as well as the electric control input energy. The spatial displacement along time and the first natural
frequency envelopes are shown as well.

3. Mathematical modeling

3.1. Finite element model for composite and piezoelectric material

The composite laminated plate is modeled using the classical FSDT (First Order Shear Deformation Theory) [26] where
the Kirchhoff assumptions are relaxed and constant shear deformation is assumed along the ply thickness, so moderately
thick plates and shells can be modelled. Other hypotheses are also considered such as the perfect bonding between layers;
the assumption the resin between plies is infinitesimally thin; and that each layer has a uniform thickness. Furthermore,
the modifications in the angle-ply configurations could induce bending–stretching and bending–twisting coupling effects
depending on the generated asymmetry of the laminate [27]. The displacements field has the following form:

ux (x, y, z, t ) = ux0 (x, y, t ) + zβx (x, y, t )

uy (x, y, z, t ) = uy0 (x, y, t ) + zβy (x, y, t )
uz (x, y, z, t ) = uz0 (x, y, t ) (1)

where ux , uy , uz are the displacements field and ux0 , uy0 , uz0 , β x , β y are the unknowns functions. In this case, ux0 , uy0 , uz0 ,
are the displacements of a point on the plane z = 0 and β x , β y represents a rotation of a transverse normal about y and x
axis, respectively. The linear strains associated to the displacements field are thus obtained as

∂ ux ∂ ux0 ∂ βx
εx = = +z
∂x ∂x ∂x
∂ u y ∂ u y0 ∂ βy
εy = = −z
∂y ∂y ∂y
εz = 0
   
∂ ux ∂ uy ∂ u x0 ∂ u y0 ∂ βx ∂ βy
γxy = + = + +z +
∂y ∂x ∂y ∂x ∂y ∂x
∂u ∂u
γxz = z0 + βx , γyz = z0 + βy (2)
∂x ∂y
A triangular three node finite element with 6 DoF per node (including drilling degree of freedom) is implemented for
FSDT that is also used as a framework for the structural piezoelectric lamina (TCGC-T9 element, see [28]). Mindlin plate
theory with Lagrange interpolation functions enhanced with drilling effect are used to approximate the displacements field.
554 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

So, these main assumptions lead to

⎡ ∂ ux0 ⎤ ⎡ ⎤
⎡ε ⎤ ∂x
∂ βx
x ∂ uy0 ∂x
⎢ ⎥ ⎢ ∂ βy ⎥
⎢ εy ⎥ ⎢ ⎥ ⎢ ∂y ∂y ⎥
⎢γyz ⎥ = ε0 + zκ = ⎢ ∂ uz0 + βx ⎥ + z⎢ 0 ⎥ (3)
⎣ ⎦ ⎢ ∂∂uxz0 ⎥ ⎢ ⎥
γzx ⎣ ∂ y + βy ⎦ ⎣ 0 ⎦
γxy ∂ ux0 ∂u
+ y0
∂ βx
+
∂ βy
∂y ∂x ∂y ∂x
where κ are the curvatures (derivatives of the rotation angles), ε0 are the mid-plane strains, z is the ply height referred to
the mid-plane.
Piezoelectric elements might behave nonlinearly at high voltages, so it is necessary to remain at low voltage levels and
use the linear constitutive relations defined by IEEE standards [29] which are usually considered as a good representation
of those materials. The constitutive relations are described as
σ = CL ε − eT E
De = eε + ξ E (4)
where σ is the stress vector, De represents the electric displacement vector, ε is the strain vector, E is the electrical field
vector, CL is the laminate elastic tensor and finally, e is the piezoelectric constants matrix and ξ represents the dielectric
constant matrix. As previously mentioned, it is used a finite element applicable for plates and shells that takes into account
the coupling effects of membrane and bending. This finite element is extended by adding an extra DoF for each piezoelectric
layer present in the element [30].
Most piezoelectric materials used as actuators in structural mechanics have a plate aspect and orthotropic mechanical
properties. Furthermore, the nonlinear behavior is only prominent at higher voltages. Therefore, here in this paper, it is as-
sumed linearity for the piezoelectric material structural behavior. It is also assumed that the electric field (E) and the electric
displacement field (D) are uniform and normal to the plate mid-plane, being the polarization orientation perpendicular to
it.



0 0
E= 0 , De = 0 and E3 = −uφ k /hk (5)
E3 D3
where uφ k is the applied voltage on the kth lamina and hk is the piezoelectric kth lamina thickness.
Detailed steps for assumed interpolation functions, the complete integration scheme for the triangular element and
assembling of the stiffness and mass matrices can be found in [28,31].

3.2. Coupled electro-mechanical system of equations

The problem consists of a laminated composite plate with embedded piezoelectric patches placed as actuators and sen-
sors used for the device’s vibration control. The globally coupled equation of motion and electrical charge of this type of
system can be cast as follows [32]:
      
M 0 ü Dd 0 u˙ Kuu Kuφ u Fu
+ + = (6)
0 0 üφ 0 0 u˙ φ Kφ u Kφφ uφ Fφ
where u and uφ are displacement and potential electrical fields, respectively, M is the mass matrix, Dd is the damping
matrix, Fu is the mechanical force vector and Fφ is the electrical charge vector, Kuu is the mechanical stiffness, Kφφ is the
electric stiffness and, finally, Kφ u = Kuφ T are the electro-mechanical coupling stiffness matrices.
Considering the case where the electrical charge Fφ = 0, and dividing the electrical DoF into separate actuating and sens-
ing capabilities, Eq. (6) can be rewritten as
Mü + Dd u˙ + Ku = −KAuφ uAφ + Fu u (7)

−1
uSφ = −KSφφ KSφ u u (8)
where the S and A notation is indicated for sensors and actuators matrices and vectors, respectively, and K = Kuu −
−1
KSuφ KSφφ KSφu .
The potential, kinetic and electrical energies for a given time instant is defined as [33]:
1 T 1
Epot = u Kuu u + uT Kuφ u (9)
2 2
1 T
Ekin = u˙ Mu˙ (10)
2
1 1
Eele = − uTφ Kφ u u − uTφ Kφφ uφ (11)
2 2
 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566 555

4. Optimal modal control

4.1. Modal state space formulation

To reduce the problem order of multiple degrees of freedom, especially for complex structures in finite elements, it is
usual to work with a truncated modal model [34], where only the most important modes are considered in the simulation
(usually lower modes of vibration are the most easily excitable). The transformation to the modal form is performed starting
from the spatiotemporal separation hypothesis:
u (t ) =  η (t ) (12)
where η(t) is the time-dependent modal coordinates vector and  the time-independent modal matrix that can be found
solving the eigenvalue problem:
(K − M ) = 0 (13)
in addition = diag{ωi 2 }, ωi being the natural frequencies of the structure and each column of  the corresponding eigen-
vector.
Taking into account Eqs. (7) and (12), the following equation of motion is obtained in modal coordinates:
Iη̈ +
η˙ + η = T KAuφ φ A + T Fu (14)

where I is the identity matrix, considering orthonormal eigenvectors (I = T M) and

= diag{2ζ i ωi } with ζ i being the
damping ratio of the ith mode (it is assumed a proportional damping).
In the next step, it is convenient to work in the state space model to reduce the second order differential equations,
Eq. (7), to a set of first order ones. Defining the state space vector x as
 
x1 η
x= = (15)
x2 η˙
It is possible to arrive at the following system of equations:
x˙ = Ax + Bφ uφ + BF Fu
(16)
y = Cx
where uφ is the control input force and Fu some external mechanical disturbance. The state space matrices are defined as

0 I
A=
− −


0
Bφ =
 KAuφ
T


0
BF =
T

−1

−KSφφ KSφ u  0
C = −1 (17)
0 −KSφφ KSφ u 

4.2. Modern control formulation

Modern control theory is usually applied in the case of systems with multiple inputs and outputs, using time domain
instead of frequency domain as the classical control theory. The optimal control objective [35] is to determine the control
signal that will make a process be controlled and at the same time optimize a performance index. For the LQR control, the
main objective it is to minimize the performance index J:
1 ∞ T 
J= ∫ x Qx + uTφ Ruφ dt (18)
2 0
where Q is semi positive definite matrix and R is strictly positive definite matrix, both defined by the system designer using
some criterion. Those weighting matrices are defined beforehand in this paper by an optimization looking for a balance
between mechanical energy and control force (applied voltage) that should result in less mechanical potential and kinetic
energy in the cantilever plate, but with control voltages not exceeding the piezoceramic electric threshold potential for
depolarization. Since there is only one channel of control for the presented examples, R = 1, and Q is defined by γ 1 and γ 2
in the matrix:

γ1 I 0
Q= (19)
0 γ2 I
556 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

Assuming control forces that are proportional to the space state vector, Eq. (20), the objective is to find the gain vector
G applied to the state space variables in order to minimize J in full state feedback [36].

uφ = −Gx (20)
It can be shown that the solution to obtain this gain reads:

G = R−1 BT S (21)
where S is the Riccati matrix, defined by the solution of the algebraic Riccati Equation for Q and R matrices:

SA + AT S − SBR−1 BT S + Q = 0 (22)
In this work, it is used an LQR (Linear Quadratic Regulator) or a Linear Quadratic Gaussian (LQG) control, which is
composed partially of a Linear Quadratic Regulator (LQR) and a state vector observer. It is used the Kalman filter as an
observer which allows the state vector estimations based upon the output y and inferring the plant and measurement noises,
assumed as uncorrelated white noises. The feedback information for the LQG control is given by the sensor measurements
and the estimation of the state is built by the Kalman filter.
Since not all states are available to the LQG, the error between the true and estimated states can be defined as e = x − xˆ .
Adding the plant and measurement noise to the system, gives:

x˙ = Ax + Bφ uφ + BF uF + w
(23)
y = Cx + v
Those white noises are uncorrelated, with zero mean and covariance intensity matrices W and V. The estimator state is
given by

x˙ = A
x + Bφ uφ + BF uF + Kob (Cx + v − (C
x)) (24)

where Kob = Sk CT V − 1 is the observer gain, found in a similar fashion as the previous G gain, by the Riccati equation, using
W and V, instead of the weighting matrices Q and R. The closed-loop system is then defined by
     
x˙ A − Bφ G Bφ G x BF I 0 w
= + uF + (25)
e˙ 0 A − Kob C e BF I −Kob v
The control forces are defined in this case based on the estimated states as

uφ = −G
x (26)

5. Fuzzy analysis

5.1. Fuzzy α -cut optimization

Fuzzy Set Theory was first presented by Zadeh [37] in his seminal paper. A fuzzy set defines a class of objects that have
some sort of belongings (pertinence) to a group of objects. Differently, from traditional Set Theory that has crisp borders,
a fuzzy set is composed of a support function (known as membership function) with values that can lie between 0 and 1,
thus being possible to deal with vague or incomplete information. In 1978, Zadeh [38] released the Possibilistic Theory using
fuzzy sets that played a similar role to the probability theory for definite stochastic distributions. The α -cut (or α -levels)
methodology divides the fuzzy set into well-defined crisp real intervals with α -levels of belongingness or relevance (μ) and
can be defined generically for a fuzzy variable A as

Aαk = {q ∈ A
˜| μ(q ) > αk } (27)

where A˜ is a fuzzy set and q is an element of A ˜ , α k is the α -level cut k that represents some sort of uncertainty (α -
cuts ranges from 0, complete uncertainty, to 1, complete certainty). This interval can also be defined as Aαk = [qα k , q̄α k ],
where qα k and q̄α k represent the minimum and maximum values for the Aαk set, respectively. Fig. 1 shows the graphical
representation of a fuzzy set A˜ .
In order to study and analyze input or model parameter uncertainties modeled as fuzzy sets, it is applied the fuzzy
α -cut optimization methodology (Möller et al. [39]). This methodology consists in multiple interval anti-optimization anal-
yses along membership functions for the output that is itself a fuzzy set composed of fuzzy numbers, as expected. So,
 
˜= f Q
Z ˜ 1, Q
˜ 2, . . . , Q
˜n (28)

where Q ˜ j means the fuzzy inputs, n is the number of inputs, f some linear or non-linear deterministic function and Z ˜ a
fuzzy output set. This can also be understood as uncertainty propagation in terms of intervals, since the uncertainty of the
inputs and parameters will be propagated to the outputs.
 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566 557

˜ and its sub sets Aαi and Aα for i to k α -cuts.

Fig. 1. Fuzzy set A k

Fig. 2. Uncertainty propagation for two input and one output problem at defined α -cut level.

The fuzzy α -cut optimization consists in finding for each pertinence level, the minimum and maximum value for the
desired output. Therefore, the problem consists of multiple optimization direction searches, two (minimization and maxi-
mization) for each α -cut that the fuzzy set was discretized. To solve Eq. (28), the problem can be stated as
Find Qi∗αk to minimize or maximize z jαk = f (q1 , q2 , . . . , qn )
  (29)
subjected to Qiαk = qiα k , qiα k
where zj is a specific desired output and for a different j output, another combination of optimal Qi∗α might arise.
k
Once defined the desired discretization for the α -levels, the uncertainty propagation is performed for each α -level cut.
Fig. 2 represents the proposed problem for a two input and one output problem Q˜1 , Q˜2 and Z˜ :

5.2. Particle swarm optimization

The particle swarm optimization (PSO) is a heuristic optimization algorithm inspired by the observation of the social
behavior of beings such as fish schooling, insects swarming and birds flocking. Kennedy and Eberhart [40] introduced this
algorithm in 1995. The method is based on the social influence and social learning, so the exchange of information between
individuals may lead them to solve complex problems. This algorithm was chosen due to simplicity in implementation and
558 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

Table 1
Composite laminate and PZT piezoceramic properties.

Graphite epoxy PZT [29]

Stacking sequence [45°/−45°/−45°/45°]

Density ρ = 1600 kg/m3 ρ = 7600 kg/m3
Layer thickness hc = 0.5 mm hp = 0.25 mm
Young’s moduli E1 = 172.5 GPa E2 = 6.9 GPa E1 = E2 = 63.0 GPa
Shear moduli G12 = G13 = 3.45 GPa G23 = 1.38 GPa G12 = G13 = G23 = 24.6 GPa
Poisson’s ratio ν 12 = 0.25 ν 12 = 0.28
Piezoelectric constant e31 = e32 = 10.62 C/m2
Electrical permissivity ξ 33 = 15.50 nF/m
Electrical potential limit ±200 V

reported robustness in finding the global optimum. The PSO is used as the main engine for the fuzzy α -cut optimization
analysis. As stated by Li et al. [41], it involves a number of particles, which have a defined position and velocity, and they
are initialized randomly in a multidimensional search space of a cost function (a modification of the objective function to
handle with constraints in the problem). Each particle represents a potential solution for the problem and the measure of
suitability is the cost function value. The set of particles is generally referred as “swarm”. These particles “fly” through the
multidimensional space. They have three essential reasoning capabilities: inertia, the memory of their own best position
and knowledge of the global or neighborhoods best position.
The basic particle parameters, position, and velocity, are updated throughout each iteration by the following equations:
    
vki,+1
j
= χ ωvk
i + λ1 r 1 xlbest
k
i, j − x k
i, j + λ2 r 2 xgbest
k
j − x k
i, j

xki,+1
j
= xki, j + vki,+1
j
(30)

where vki,+1
j
e vki, j are the updated velocity and actual one, respectively, of particle i with respect to design variable j. In the
same way xki,+1
j
and xki, j are the particle position. xlbestik is the best position found so far from the self-historical path by
particle i while xgbest k is the best position found so far in the swarm until that iteration k. r1 e r2 are random numbers
between 0 and 1, while λ1 e λ2 are cognitive parameters that represent the confidence of the particle on its own result
or on the swarm best result (both set as 2.0 in this paper). The χ parameter is used to avoid the divergence behavior in
the algorithm, and it was proposed by Bergh and Engelbrecht [42]. Finally, ω is the inertia factor, introduced in the original
PSO by Shi and Eberhart [43], and means the importance of the past particle velocity on the searching procedure generating
such a momentum in the direction of search.
As a metaheuristic algorithm, PSO makes few or no assumptions about the smoothness or differentiability in the prob-
lem being optimized and thus it is not prone to get stuck in local minima. Furthermore, the algorithm present stochastic
operators (random numbers) and this reinforces this virtue. The good behavior and robustness of the algorithm in searching
very large spaces of candidate solutions are referred in the literature and benchmarked with several optimization problems
(Taherkhani and Sababaksh [44]). The use of metaheuristic PSO as optimization engine in the α -cut level analysis (anti-
optimization) is one of the interesting innovative propositions of the paper since it provides extra security against local
minima.

6. Numerical examples

Being a multidisciplinary field, smart materials presents a variety of uncertainty sources, from materials properties to
circuitry and controlling features. This analysis deals with the composite material and piezoceramic properties uncertainties
and the resulting variability, or uncertainty propagation in the structural behavior. In this example, it is studied a cantilever
composite plate. The cantilever beam is the most used structure for experimental and numerical benchmark/validation of
control strategies. According to Kim et al. [45], this setup is widely used in energy harvesting. Raaja et al. [46] used a
piezoelectric cantilever beam as a MEMS accelerometer to investigate SHM applications. Since the idea of this paper is to
deal with many sources of uncertainties in the parameters of the controlled composite structure, it was chosen a classical
and well-behaved test case to take better account of uncertainties effects.
In the example, it is studied a cantilever composite plate consisted of 4 layers with 30 cm in length and 4.5 cm in width,
embedded with piezoceramic patches (lead zirconate titanate, PZT) to control free vibration due to a suddenly applied
force of 0.5 N for a short period (5 ms) at all free nodes at the tip of the beam. The piezoelectric patch is positioned
at the fixed end and next to the lateral edge as shown in Fig. 3. This position choice is based on the bending and twisting
vibration modes (present in the first 4 modes) that should be controlled and that was previously defined by an optimization
procedure.
The nominal properties of the composite material plate and piezoelectric layer properties are given in Table 1. The piezo-
ceramic electric threshold potential for depolarization is ±200 V. In this configuration, the first 4 vibration modes have been
 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566 559

Fig. 3. Finite element mesh of the cantilever composite plate with the location of the piezoelectric patches (dark elements) and the applied forces.

Fig. 4. Fuzzy sets for layer orientation of 45° (left) and −45° (right).

selected to be controlled and for simulation purposes, the modal FE model including the first 10 vibration modes is used. It
was used a weakly modal damping ratio for all modes with value ζ = 2%.

6.1. Nodal displacement envelope

In this example, it is considered an uncertainty in the fiber orientation for each layer, independently; therefore, there are
four uncertain parameters. It is established that those inputs are modeled as fuzzy sets in a symmetrical triangular shape
and it is discretized in five α -levels, plus the nominal one (α = 1). For each level, there is an interval varying around the
nominal value of ±1.25°, ±2.50°, ±3.75°, ±5.00° and ±6.25° (Fig. 4). For this example, it is used the LQR controller and the
PZT patches are placed only on the upper face of the cantilever, located on the darker elements of Fig. 3. The weighting
matrix Q was optimized (PSO) for the minimum mechanical energy in the nominal case, respecting the PZT limitation of
±200 V. The values used in this example are γ 1 = 2.810 × 106 and γ 2 = γ 1 = 1.184.
The intention is to build a displacement envelope along time (minimum and maximum possible values) for the free
tip corner node displacement (1 uncertain output). This type of analysis might be computationally expensive since two
optimizations are performed for each time step and for each α -level in order to obtain the fuzzy intervals. Figs. 5–9 show
the envelopes for different α -levels and built with a discretization of 125 time intervals for 1 s simulation. In this example,
with 24 individuals, each upper or lower limit for each α -cut, comprising 125 single independent optimizations resulted
around 40,0 0 0 simulations.
Analyzing envelopes in Figs. 5–9, it is possible to note that at the beginning of the transient behavior, most configurations
have a similar path and the interval is narrow. However, as time goes by that interval becomes smooth indicating that there
is not an increase in the displacement interval (it closely accommodate the periodic motion) but mostly in the uncertainty
phase (several periodic displacement histories may occur). That effect is due to different vibration phases for different ply
configurations that fulfill the wave gaps of the envelope presented at the beginning. It is interesting to note that the stacking
sequence for each instant of time is different, so there is not a single case that would follow exactly one of the boundaries.
Overlapping the envelopes for different α -levels and adding, in red, the nominal curve (α = 1, the tip of the triangle) it
results in Fig. 10.
560 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

Fig. 5. Displacement envelop for α = 0 (±6.25°).

Fig. 6. Displacement envelop for α = 0.2 (±5.00°).

Fig. 7. Displacement envelop for α = 0.4 (±3.75°).

Fig. 8. Displacement envelop for α = 0.6 (±2.50°).

Fig. 9. Displacement envelop for α = 0.8 (±1.25°).

 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566 561

Fig. 10. Overlapped envelopes for different α -levels (red is the nominal response, darker areas representing high α -levels, light lines representing lower
α -levels). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 11. Fuzzy sets of displacement for some time instants.

Fig. 12. Displacement variability interval (the difference between lower and upper boundaries along time).

Analyzing each time instant, getting vertical cuts through the curves of Fig. 10, it is possible to see the fuzzy sets for
those outputs, depicted in Fig. 11.
Another study that is possible to be performed is the displacement variability interval for each time instant (distance
between boundaries), in order to define where is the critical time instant with the highest degree of displacement uncer-
tainty. Fig. 12 shows these results for ply angle intervals ±1.25° and ±6.25°.

6.2. Frequency domain envelope

Following the previous example, considering the same structure and uncertainty, the objective here is to build the fre-
quency response for the referred node displacement. The Power Spectral Density (PSD) envelope was built around the first
natural frequency (13.28 Hz for the nominal structure), and the algorithm performed the minimization and maximization of
the PSD amplitude in a range of 5 Hz and 20 Hz, discretized in 0.5 Hz width spectral lines. The overlapped plot for different
α -cuts and the nominal value (red) are shown in Fig. 13. Frequencies are shown as fuzzy sets are presented in Fig. 14, for
some frequency steps.
In a different way, if the interested output is some natural frequency itself, the optimization is performed to find the
extreme cases for each α -level on that specific vibration mode. In the case of the first natural frequency, the resulted fuzzy
set is shown in Fig. 15.
562 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

Fig. 13. Overlapped envelopes for different α -levels. (For interpretation of the references to color in this figure, the reader is referred to the web version
of this article.)

Fig. 14. PSD tip displacement variability as fuzzy set intervals for some discrete frequencies.

Fig. 15. Fuzzy set for the first natural frequency.

6.3. Multiple uncertainty sources and energy analysis

Emulating a more realistic scenario where there might be different small sources of uncertainties in the project, this
study analyzes the uncertainty propagation influence in system’s output (considered as mechanical and electrical energies
related to the vibration energy and used controlling forces). In this numerical example, it will be used the LQG. The PZT
sensors are attached to the lower face of the plate collocated to the actuators as shown in Fig. 3. The used LQG parameters
are γ 1 = 6.353 × 106 and γ 2 = 4.478. Since a Kalman filter is necessary to predict the system states, plant and sensors
noises, W and V need to be defined. Their spectral densities are hard to obtain in real cases, so those matrices are treated
as parameters of the controller like Q and R matrices. Here, it is used W = 10−2 Bφ BTφ and V = 1.
The uncertain fuzzy input parameters are modeled as symmetrical triangles where the lowest α -level (α = 0) have a ±3%
variability from the nominal value and, the fibers orientation for each layer, a variability f ±3°, defined in Table 2. Therefore,
there are 22 uncertain parameters in total, considering PZT actuator, sensor, and the composite Graphite-Epoxy properties.
The fuzzy sets are discretized in four α -levels plus the nominal one.
 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566 563

Table 2
Uncertain parameters for α = 0.

Minimum Maximum

Fibers Ɵ1 and Ɵ3 42° 48°

Fibers Ɵ2 and Ɵ4 −48° −42°
G-E layer thickness 0.485 mm 0.515 mm
G-E E1 167.33 GPa 177.68 GPa
G-E E2 6.69 GPa 7.11 GPa
PZT thickness 0.2425 mm 0,2575 mm
PZT E1 and E2 61.11 GPa 64.69 GPa
PZT e31 and e32 10.30 C/m2 10.94 C/m2
PZT ξ 33 15.03 nF/m 15.97 nF/m

Fig. 16. Time response for the LQG (red) and not controlled (blue). (For interpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

Fig. 17. Control voltages applied to the actuators. Depolarization limit of 200 V is not exceeded.

For the nominal structure, the controlled and free vibration plots are presented in Fig. 16 and the force control is shown
in Fig. 17. The integration over time of the mechanical energy (sum of kinetic and potential energies) in this case resulted
in Emec = 6.35 × 10−5 J s and the electrical input energy resulted in Eele = −1.18 × 10−4 J s.
The analyzed outputs are presented in Figs. 18 and 19 as fuzzy sets and shown as a percentage of variation with respect
to the nominal value. Those results were obtained using the PSO algorithm, as proposed previously in the paper and by
Monte Carlo simulation with different population sizes. For the PSO algorithm, with 80 individuals, each α -cut limit anti-
optimization took 8–10 iterations to converge (change in the objective function) leading to a total number of structural
analysis around 60 0 0 simulations. First, it is easy to notice that even with the higher attempt and computational effort,
with 10 0,0 0 0 simulations, the Monte Carlo simulations could not obtain the extreme values in an accurate way like in the
representation of the PSO fuzzy set. It is interesting to observe that the fuzzy set results are not symmetrical, like the fuzzy
sets for the inputs, and have a higher variability for the critical case, which are the maximum energies values. Although
the inputs were established with relatively small variability, even for lower α -levels, the propagation is considerably high,
reaching up to 21% of mechanical energy increase (related to the value in the nominal structure) and up to 46% more
electrical control energy.
564 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

Fig. 18. Fuzzy number for mechanical energy variation using PSO and Monte Carlo (MC) simulations.

Fig. 19. Fuzzy number for electrical input energy variation using PSO and Monte Carlo (MC) simulations.

7. Conclusions

Smart materials with piezoelectric control are applied in different fields of engineering. The uncertainty propagation can
be used to estimate the final envelope for structural behavior.
In the best of the author’s knowledge, there are no papers that deal with control performance of piezo-laminated com-
posites under uncertainties with possibilistic parameters and a heuristic approach. The proposed methodology allows ac-
counting several uncertainty parameters that are classically treated in a Monte Carlo framework. In this paper, this high-
cost approach (MC) is faced with a heuristic (PSO) anti-optimization approach that shows advantages in finding extreme
values for the uncertain variables. The obtained results using this technique encompass those obtained by traditional MC
simulations.
In this paper, the extreme values for displacements, spent energy, natural frequencies, etc., could be found using the
interval analysis together with a heuristic algorithm approach resulting in accurate solutions and computational time-savings
when compared to simple Monte Carlo simulations. In the last example, even with 10 0,0 0 0 simulations for the uncertain
variables, Monte Carlo technique was unable to be as accurate as the results from the PSO interval search, especially for
lower α -cuts levels. The process to obtain that type of envelope can be computationally expensive in the case of fuzzy
interval analysis, since a high number of optimizations are necessary. On the other hand, when a specific uncertain output
variable is sought, the simulation might be affordable and easily performed, depending only on the size of the problem being
simulated. Considering a low variability in the uncertain parameters, it is shown that the system’s output values presented
substantial uncertainty as their fuzzy sets representation were widely spread for low α -levels. It is important to note that
this methodology is capable to deal with a large number of uncertain parameters problems, and being the main bottleneck,
only the performance of the optimization algorithm.
The importance of the displacement envelope presented in this paper is of concern for robotic applications in obstacle
avoidance. The natural frequency modification analyzed is important since it may affect the interaction behavior like in flut-
ter problems or resonance like in harvesting problems, so their variability should be taken into account in the design stage.
Regarding the variability of the electrical input energy, a substantial increase in this parameter might affect applications
where low consumption control systems are necessary like in aerospace applications. Finally, as indicated by the last exam-
ple, taking into account the vibration energy uncertainty by the mechanical energy presented a useful metric tool to check
the effectiveness of the controller.
 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566 565

In view of practical applications, this fuzzy α -cut optimization method is not intrusive and may well adapt to several
existing codes and software, allowing the technique to deal with a wide range of problems. Material property uncertainties
and fabrication defects are a reality and may cause significant modifications in the dynamic behavior of structures and
controllers. Therefore, the authors conclude that, in these cases, it is always advisable to investigate the design performance
under the presence of uncertain parameters and especially in extreme case scenarios.

References

[1] K. Umesh, R. Ganguli, Material uncertainty effect on vibration control of smart composite plate using polynomial chaos expansion, Mech. Adv. Mater.
Struct. 20 (2013) 580–591.
[2] J. Lucero, F. Hemez, T. Ross, et al., Uncertainty Quantification of Composite Laminate Damage With The Generalized Information Theory, Final Report
of the FY05 ASC-V&V-Engineering Analysis Level-5 Milestone. Los Alamos National Laboratory, 2006.
[3] E. Acar, R.T. Haftka, T.F Johnson, Tradeoff of uncertainty reduction mechanisms for reducing weight of composite laminates, J. Mech. Des. 129 (2007)
266–274.
[4] A.P. Dempster, Upper and lower probabilities induced by a multivalued mapping, Ann. Mater. Stat. 38 (1967) 325–364.
[5] D.G. Kendall, Foundations of a theory of random sets, in: E. Harding, D.G. Kendall (Eds.), Stochastic Geometry, Wiley, New York, 1974, pp. 322–398.
[6] Ferson, S., Kreinovich, V., Ginzburg, L. et al. Constructing probability boxes and Dempster–Shafer structures, Tech. Rep. SAND2002-4015, Sandia National
Laboratories, 2003.
[7] K.D. Potter, Understanding the origins of defects and variability in composites manufacture, in: Proceedings of the ICCM17 International Committee
on Composite Materials Conference 27–31 July, 2009.
[8] Roy, A. Kumar, S. Li, A nano-micro-macro-multiscale model for progressive failure prediction in advanced composites, in: P.W.R Beaumont, C. Soutis
(Eds.), The Structural Integrity of Carbon Fiber Composites Chapter 7, Springer Verlag, 2017.
[9] C.A.C. Antonio, Design with composites: material uncertainty in designing composites component, in: Luigi Nicolais, Assunta Borzacchiello (Eds.),
Wiley Encyclopedia of Composites, second Ed., John Wiley & Sons, Inc., 2012, pp. 1–12.
[10] P.D. Gosling, Faimun, O. Polit, A high-fidelity first-order reliability analysis for shear deformable laminated composite plates, Compos. Struct. 115 (2014)
12–28.
[11] R.H. Lopez, L.F.F. Miguel, I.M. Belo, J.E.S. Cursi, Advantages of employing a full characterization method over FORM in the reliability analysis of lami-
nated composite plates, Compos. Struct. 107 (2014) 635–642.
[12] V.K. Goyal, R.K. Kapania, Dynamic stability of uncertain laminated beams subjected to sub tangential loads, Int. J. Solids Struct. 45 (2008) 2799–2817.
[13] D.J. Lekou, T.T. Assimakopoulou, T.P. Philippidis, Estimation of the uncertainty in measurement of composite material mechanical properties during
static testing, Strain 47 (2011) 430–438.
[14] B.N. Singh, N. Grover, Stochastic methods for the analysis of uncertain composites, J. Indian Inst. Sci. 93 (4) (2013).
[15] H. Zhang, R.L. Muller, R.L. Muhanna, Interval Monte Carlos method for structural reliability, Struct. Safety 32 (2010) 183–190.
[16] R. Chowdhury, S. Adhikari, Fuzzy parametric uncertainty analysis of linear dynamical systems: a surrogate modelling approach, Int. j. numer. methods
eng. 70 (4) (2007) 405–422.
[17] S. Chakraborty, C.S. Palash, Probabilistic safety analysis of structures under hybrid uncertainty, Int. J. Numer. Methods Eng. 70 (4) (2007) 405–422.
[18] S.K. Mishra, B.K. Roy, S. Chakraborty, Reliability-based-design-optimization of base isolated buildings considering stochastic system parameters sub-
jected to random earthquakes, Int. J. Mech. Sci. 75 (2013) 123–133, doi:10.1016/j.ijmecsci.2013.06.012.
[19] B.K. Roy, S. Chakraborty, S.K. Mishra, Robust optimum design of base isolation system in seismic vibration control of structures under uncertain
bounded system parameters, J. Vib. Control 20 (5) (2012) 786–800, doi:10.1177/1077546312466577.
[20] Jeong, D.-Y., Kang, T., Dharmayanda, H.R., Budiyono, A. H-infinity attitude control system design for a small-scale autonomous helicopter with nonlinear
dynamics and uncertainties, J. Aerospace Eng., Vol 25, No. 4. doi:10.1061/(ASCE)AS.1943-5525.0 0 0 0176.
[21] E.H. Koroishi, D.A.L. Molina, A.W. Faria, V.S. Junior, Robust optimal control applied to a composite laminated beam, J. Aerosp. Technol. Manag. São José
dos Campos 7 (1) (2015) 70–80.
[22] T.R.B. Trevilato, Active control of smart composite structures in the presence of uncertainty (in Portuguese). Master’s Thesis, Federal University of
Uberlândia, Graduate Program in Mechanical Engineering, Brazil, 2012.
[23] S.-B. Choi, Vibration control of a smart beam structure subjected to actuator uncertainty: experimental verification, Acta Mech. 181 (2006) 19–30.
[24] D.G. Marinova, G.E. Stavroulakis, E.C. Zacharenakis, Robust control of smart beams in the presence of damage-induced structural uncertainties, in:
Proceedings of PhysCon, St. Petersburg, Russia, 2005, pp. 339–344.
[25] A.J. Moutsopoulou, G.E. Stavroulakis, A.T. Pouliezos, Model uncertainties in smart structures, in: M. Papadrakakis, M. Fragiadakis, V. Plevris (Eds.),
Proceedings of Third ECCOMAS Thematic Conference on Computational Methods in Structural Dynamics and Earthquake Engineering, Corfu, Greece,
25-28 May 2011.
[26] J.N. Reddy, Mechanics of Laminated Composite Plates and Shells: Theory and Analysis, CRC Press, 2004.
[27] R.M. Jones, Mechanics of Composite Materials, CRC Press, 1998.
[28] Y.-Q. Long, S Cen, Z-F. Long, Advanced Finite Element Method in Structural Engineering, Springer, 2009.
[29] IEEE std, ANSI/IEEE Std 176-1987: Standard on Piezoelectricity, IEEE, New York, 1988.
[30] L.A. Isoldi, A.M. Awruch, P.R.F. Teixeira, I.B. Morsch, Geometrically nonlinear static and dynamic analysis of composite laminates shells with a triangular
finite element, J. Braz. Soc. Mech. Sci. Eng. 30 (2008) 84–93.
[31] Y.-Q. Long, B. Xiaoming, L. Zhifei, X. Yin, Generalized conforming plate bending elements using point and line compatibility conditions, Comput. Struct.
54 (4) (1995) 717–723.
[32] S.L. Schulz, H.M. Gomes, A.M. Awruch, Optimal discrete piezoelectric patch allocation on composite structures for vibration control based on GA and
modal LQR, Comput. Struct. 128 (2013) 101–115.
[33] V. Balamurugan, S. Narayanan, Shell finite element for smart piezoelectric composite plate/shell structures and its application to the study of active
vibration control, Finite Elem. Anal. Des. 37 (2001) 713–738.
[34] C.M.A. Vasques, J.D. Rodrigues, Active vibration control of smart piezoelectric beams: comparison of classical and optimal feedback control strategies,
Comput. Struct. 84 (22) (2006) 1402–1414.
[35] D.S. Naidu, Optimal Control Systems, CRC Press, 2002.
[36] A. Preumont, Vibration Control of Active Structures: An Introduction, Springer, 2011.
[37] L.A. Zadeh, Fuzzy sets, Inf. Control 8 (1965) 338–353.
[38] L.A. Zadeh, Fuzzy sets as a basis for a theory of possibility, Fuzzy Sets Syst. 1 (1978) 3–28.
[39] B. Möller, W. Graf, M. Beer, Fuzzy structural analysis using α -level optimization, Comput. Mech. 26 (20 0 0) 547–565.
[40] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proceeding of 1995 IEEE International Conference on Neural Networks, 1995, pp. 1942–1948.
[41] L.J. Li, Z.B. Huang, F. Liu, Q.H. Wu, A heuristic particle swarm optimizer for optimization of pin connected structures, Comput. Struct. 85 (2007)
340–349.
[42] F. Bergh, A.P. Engelbrecht, A study of particle swarm optimization particle trajectories, Inf. Sci. 176 (2006) 937–971.
[43] Y. Shi, R. Eberhart, A modified particle swarm optimizer, IEEE World Congress Comput. Intell. (1998) 69–73.
566 M.D.F. Awruch, H.M. Gomes / Applied Mathematical Modelling 54 (2018) 551–566

[44] M. Taherkhani, R. Safabakhsh, A novel stability-based adaptive inertia weight for particle swarm optimization, Appl. Soft Comput. 38 (2016) 281–295,
doi:10.1016/j.asoc.2015.10.004.
[45] J. Kim, S. Park, W. Lim, J. Jang, T.H. Lee, Design optimization of PZT-based piezoelectric cantilever beam by using computational experiments, J. Electron.
Mater. 45 (8) (2016), doi:10.1007/s11664- 016- 4497- 2.
[46] B.P. Raaja, R.J. Daniel, K. Sumangala, A simple analytical model for MEMS cantilever beam piezoelectric accelerometer and high sensitivity design for
SHM (structural health monitoring) applications, Trans. Electrical electron. Mater. 18 (2) (2017) 78–88.