Implementation of adaptive radial basis function

neural network based sliding mode control on

magnetorheological elastomer isolator
Kiran Katari1 , Umanath R P2*† , Gangadharan K V1†
1 Department of Mechanical Engineering, National Institute of
Technology, Srinivasa nagar, Mangalore, 575025, Karnataka, India.
2* Department of Mechanical and Manufacturing Engineering, Manipal

Institute of Technology, Manipal, Udupi, 576104, Karnataka, India.

*Corresponding author(s). E-mail(s): umanr@hotmail.com;

Contributing authors: kiran.katari@hotmail.com; kvganga@nitk.edu.in ;
† These authors contributed equally to this work.

Abstract
This article presents a novel approach for effectively controlling vibrations in mag-
netorheological elastomer-based semi-active vibration isolators. The controller’s
ability to respond to variations in magnetic field, frequency, and amplitude of
stimulation is crucial for achieving optimal vibration control. To address this
challenge, a combination of a radial basis function neural network and a sliding
mode controller (RBSM controller) is proposed and implemented under vari-
ous scenarios on the magnetorheological elastomer (MRE) isolator, with control
performance as the primary criterion. To implement the controller, a mathemat-
ical model is required. To estimate the unknown properties of the mathematical
model, a neural network is employed. The neural network is integrated with a
sliding mode controller, which acts as the primary controller for regulating vibra-
tions. Single and multi-amplitude frequency excitation simulations are conducted
to assess the performance and adaptability of both the sliding mode and RBSM
controllers. Furthermore, the Lyapunov stability of the RBSM controller is eval-
uated. The simulation results demonstrate that the RBSM controller effectively
mitigates vibrations. Additionally, when the sliding mode controller is utilized in
conjunction with the neural network, chattering - an undesired phenomenon asso-
ciated with sliding mode control - is significantly reduced. This combination of
RBSM control and neural network estimation proves to be a promising strategy
for vibration control in magnetorheological elastomer-based semi-active systems.

1
 Keywords: Magnetorheological Elastomer, Vibration isolator, Hybrid controller,
Sliding mode controller, Radial basis function neural network, Dynamic properties
characterization.

1 Introduction
Passive rubber isolators have set dynamic qualities and may reduce structural vibra-
tion over a restricted frequency range. The dynamic features of semi-active isolators,
on the other hand, may be changed to suit the demand of vibration isolation across
a broad frequency range. The semi-active isolator may have variable stiffness and
variable damping, or it can be a mix of the two. Semiactive isolators with variable
stiffness change the natural frequency of the system by adding or deleting stiffness.
The variable damping type semiactive isolator, on the other hand, adjusts the sys-
tem response by adding or eliminating damping qualities. Semi-active isolators with
variable stiffness were generally made of shape memory alloys (SMAs) [1], piezoelec-
tric [2], and magnetorheological elastomers [3]. The variable damping type isolator is
constructed by either changing the aperture of the fluid system [4] or changing the
viscosity of the magnetorheological fluid [5]. M.F. Winthrop et al. [6] studied a variety
of variable stiffness devices for vibration control applications. The research indicated
that magnetorheological elastomer (MRE)-based devices have more promise than their
competitors since they provide bigger stiffness changes (SMAs and piezoelectric).
MRE is a kind of smart material that alters its dynamic characteristics when
subjected to a magnetic field [7]. It is made up of an iron particle that is spread in
a nonmagnetic polymer matrix. MRE is classified as isotropic or anisotropic based
on the arrangement of the filler [8]. When the curing procedure is carried out in the
absence of a magnetic field, isotropic structure develops. Because the curing process
is carried out under the influence of a magnetic field, the anisotropic stricture has an
ordered layout. The amount of change in natural frequency, stiffness, or MR effect
determines the performance of a semi-active isolator. The MR effect can be described
as either absolute or relative. The absolute magnetorheological (MR) effect is defined
as the difference between the maximum shear modulus (Gmax) values obtained in
a magnetic field and the value obtained in the absence of a magnetic field (G0)[9].
The isolator device performance mainly depends on the amount MR effect achieved
by the particular device. There numorous research articles provides the details of
the fabrication, characterization of MREs aiming at enhancing their MR effect have
been widely studied [10] [11] [12] [13]. MR properties vary with conditions such as
preload, excitation amplitude, frequency, and applied magnetic field, as demonstrated
in numerous articles.[14] [15].
The usage of the MRE-based semi-active isolator device needs the employment of
an effective control approach. When an MRE isolator is needed, the controller must
generate a control signal with a frequency range of 0 Hz to the frequency at which
the passive MRE isolator and maximum active MRE isolator characteristics coin-
cide. The controller should be turned off at the remaining higher frequencies, and the
MRE isolator should operate as a passive device. Apart from that controller must be

2
robust and adaptive to the any changes in properties of the MRE isolator. Several
control strategies for MRE-based isolators have been proposed in order to increase
their vibration mitigation efficacy. However, relatively few MRE-based isolator control
techniques have been evaluated under diverse excitation situations and demonstrated
robust and adaptive control performance. The switching law governs on-off control,
which involves a control signal created between the lowest and maximum current. To
address the control need, an on-off control technique is often used [16]. The switching
law governs on-off control, which involves a control signal created between the lowest
and maximum current. Jerkin, the oscillating member, is created by the on-off switch-
ing law [17]. For MRE-based vibration isolation devices, a fuzzy logic controller was
also presented, in which the control rule is built using expert knowledge [18] [19] [20].
The fuzzy controller has the advantage of not requiring sophisticated mathematical
modelling. However, its success is heavily reliant on the predefined fuzzy rules based
on expert knowledge. For MRE-based vibration isolation devices, clippedoptimal and
Lyapunov-based controllers were also presented [21] [22][23]. However, precise dynamic
modelling of the system integrated with MRE-based devices is required for these con-
trollers, which is not accessible in most actual instances. A neural network-based
inverse model for accurate modelling of MRE-based vibration isolation devices was
recently presented; however, the inverse model derived from off-line training cannot
be directly used for a real-time controlled system [24] [25] [26].
Developing an exact dynamic model for model-based controller design is extremely
difficult or impossible. Traditional adaptive control techniques had been devised to
control dynamic systems with faulty information. In general, these algorithms are
predicated on the assumption that the structure of the system model is known, but
the system parameters are unknown. Sliding mode control (SMC), on the other hand,
has proven to be capable of dealing with system dynamic uncertainty and external
disturbance while achieving high robustness. However, it still requires system model
information for control law calculation. Although neural network controllers have been
effectively used in various disciplines of controls engineering to find system dynamic
models for indirect adaptive control, the performance of neural networks (NNs) in
direct closed-loop controllers is difficult to guarantee.
In this current study, a combination of sliding mode control and radial basis neural
network was developed to provide a control input to MRE isolator in order to achieving
the required isolation in the receiver end. Sliding mode controllers frequently require
an adaptation rule to change the non-linear input output mappings. The RBF network
has been widely used to represent the non-linear mappings between the inputs and
outputs of non-linear control systems. The adaptive rule was used in the RBF network
for online adjustment of the weighting of radial basis functions by employing the
reaching condition of a defined sliding surface. The control performance of the RBSM
controller was tested in simulation under single and multi-frequency stimulation at
varied amplitudes of excitation.

3
Fig. 1 (a) Schematic representation of MRE isolator and (b) magnetic field simulation of a cross-
section of the isolator using ANSYS Maxwell 15.

2 Isolator Design
Because of lower starting stiffness levels, the suggested MRE isolator operates in shear
mode, which provides a greater enhancement. Figure 1(a) shows a schematic diagram
of the isolator. The primary isolator body and the shell make up the isolator. The
fixed electromagnet, moveable central core, two ring-shaped MRE components, and
an aluminium piston coupled to the mobile central core comprise the isolator body.
A copper coil (20 AWG gauge wire with 980 turns) is twisted around the bobbin to
form the electromagnet (inner diameter 20 mm, Outer diameter 100 mm and height 40
mm). Low-carbon steel is used to make the core material (fixed and movable steel core)
(relative permeability of 100). Two ring-shaped MRE resilient elements are mounted
at the top and bottom of the isolator between the fixed and movable cores. To facilitate
motion transfer from the shaker to the receiving end, a nonmagnetic piston is mounted
to the moving steel core. In section 3, the ingredients and synthesis process of the MRE
are addressed. To focus the flux lines through the MRE, a fixed steel plate core and a
centre moving cylindrical steel core are used. The flux lines passing through the core
of the electromagnetic coil are intended to be perpendicular to the loading direction
of the MRE resilient element. To investigate the flux route and field distribution in
the MRE area, magnetic field simulations were run in ANSYS Maxwell (Figure 1(b)).
The effective magnetic permeability of MRE is computed using the Maxwell Garnett
mixing rule [27], and it is represented as follows:
µp − µm
µef f = µm + 2ϕµm (1)
µp + µm − ϕ(µp − µm )
where, µm and µp are permeability of the matrix and particle. ϕ is the volume
fraction of the particle. The magnetic flux lines are channelled via the MRE robust
element, as shown in Figure 1a. With a 3 A input current, the suggested electromag-
net design could create a magnetic flux density of roughly 0.38 T via the MRE robust
element. The isolator’s shell is a cylindrical aluminium enclosure with dimensions of
(inner diameter 120 mm and outer diameter 150 mm and height 170 mm). To ease

4
Fig. 2 MRE isolator components

movement of the centre core, two aluminium support rings (inner diameter 100 mm,
outer diameter 120 mm, and height 25 mm) are supplied at the top and bottom. To
securely install the MRE isolator body in the case, two mounting plates (one at the
bottom and one at the top, plate diameter 180 mm and thickness 10 mm) are sup-
plied. The bottom mounting plate secures the isolator assembly to the electrodynamic
shaker.

3 Manufacturing of MRE
The field-induced enhancements are critical for MRE isolators to achieve isolation
over a wide frequency range. Past studies have revealed that silicone rubber as a
matrix for MRE isolator could achieve a high relative MR effect [28]. For the proposed
design, a two-ring shaped MRE resilient elements are prepared (inner diameter 40 mm,
outer diameter 50 mm and thickness 5 mm). The Sample images and other isolator
components images are shown in Figure 2. The MRE is synthesized with the silicone
matrix and CIP fillers. The silicone matrix is procured from Aditya Silicone Rubber
(MurtiSil, Rtv 1010) and the CIP fillers are of type CC (manufacturer: BASF, size
distribution: 3-5 µm). As reported by the past studies, an optimum content of 72%
by a weight ratio of CIP is chosen for the synthesis process [28]. Additionally, Silicone
oil is added as an additive for the ease of the synthesis process. These materials are
mixed thoroughly with a weight ratio of 72% carbonyl iron, 16% silicone rubber, and
12% silicone oil, and the mixture is poured into the mould. The mould was kept
in the vacuum chamber for about 20 min to remove the mixer trapped air bubbles.
After the vacuuming process, the MRE samples are cured for about 24 hours at room
temperature under constant pressure.

4 CONTROL STRATEGIES
Developing a mathematical model for an MRE isolator is a time-consuming process
that required a large number of experimental tests under various conditions. If the
mathematical model for the system is unknown, a radial basis function neural network
sliding mode controller is used to implement control (RBSM controller). To achieve
the desired response, the weights of the neural network are changed such that the
sliding mode controller generates the necessary control input to the isolator to regulate
vibrations in the receiving member.

5
Fig. 3 SDOF of the MRE isolator

4.1 System Description

A single-degree-of-freedom system (SDOF) representation of the MRE isolator model
with inclusion of control force F (t) is considered, as shown in Figure 3. This model
comprises Zener [29] and Bouc-Wen [30] elements representing the viscoelastic and
hysteresis behaviour of the MRE isolator. Both Zener and Bouc-wen elements are
arranged in parallel and the parameter of both modelling elements (k1 , c, k2 and Z)
vary with the magnetic field, the amplitude of excitation, and frequency. The control
output F (t) from sliding mode controller is added to model in order to produce desired
output.
The governing equation of the MRE isolator is

mẍ (t) + (k 1 ) (x (t) − w (t)) + (c) (ẋ (t) − ẏ (t)) + z(t) + F (t) = 0 (2)

(k 2 ) (y (t) − w (t)) = (c)(ẋ (t) − ẏ (t)) (3)

where, Z is hysteretic force. The Bouc-Wen element of the model is represented as,
where, α, β and γ are non-dimensional parameters responsible for the shape and
size of the hysteretic loops. The system viscoelastic parameters are altered by varying
the applied current and excitation amplitude.

4.2 Sliding mode control

In general, there are differences between the response of the mathematical model and
the actual response of the MRE isolator. These inconsistencies are caused by a variety
of factors, including variations in MRE property due to magnetisation and excita-
tion amplitude. To overcome these inconsistencies and establish a successful controller
strategy, a collection of strong controller designs is offered specifically for the MRE
isolator. The sliding mode controller methodology is one of the robust controllers for
the MRE isolator. The sliding mode controller is distinguished by its ability to per-
form efficiently for nonlinear systems while remaining robust in the face of parametric
uncertainties and external disruptions. Two steps are involved in the design of a slid-
ing mode controller [31]. The first step is to mathematically define the sliding surface

6
so that the MRE isolator can be confined over it in order to get the appropriate system
response. The second step is to calculate the expression for switched feedback gains
needed to move state trajectory of the plant to the sliding surface. Finally, using the
Lyapunov stability theory, the developed controller is proved to be stable.
The sliding surface variable s defined as[31]
n
X n
X
s(x) = cT x = ci xi = ci x i + x n (4)
i=1 i=1
where, x is the state vector, the slope of the sliding surface c = [c1 c2 . . . . . . cn−1 ]T
In sliding mode control, parameters c1 c2 . . . . . . cn−1 should be selected so that the
polynomial ln−1 + cn−1 ln−2 + . . . + c2 l + c1 is Hurwitz, where l is Laplace operator.
For the second-order viscoelastic model for MRE isolator rewritten as
n
X n
X
s(x) = cT x = ci xi = ci x i + x n (5)
i=1 i=1
1
ẍ (t) = − (k1 (x (t) − w (t)) + c (ẋ (t) − ẏ (t)) + z(t) + F (t)) (6)
m
Let

f (t) = k1 (x (t) − w (t)) + c (ẋ (t) − ẏ (t)) + z(t) (7)

then ẍ (t) will be
1
ẍ (t) = −
(f (t) + F (t)) (8)
m
Now let the desired output be xd then the error e and error rate ė will be

e = xd (t) − x(t) ė = x˙d (t) − ẋ(t) (9)

For second-order system, design sliding mode function as

s = ė + c1 e, where c1 > 0 (10)

Lyapunov function as
1 2
sV = (11)
2
The derivative of the s variable written as
1
ṡ = ë + c1 ė = − (f (t) + F (t)) + x˙d (t) + c1 ė (12)
m
To guaranty stability, V̇ ≤ 0 Then,
 
1
V̇ = sṡ = s − (f (t) + F (t)) + x˙d (t) + c1 ė = −η sgn(s) ≤ 0 (13)
m
With known f (t), the desired control force is

F (t) = m (ẍd + η.sgn (s) + c1 ė) − f (t) (14)

7
Fig. 4 The architecture of a radial basis function neural network.

With the known desired response ẍd from the viscoelastic model, the sliding model
uses single gains of η and c1 to track the desired response with the generated controller
input to the MRE isolator.

4.3 RBSM controller (Radial basis neural network controller)

Neural networks are made up of neuron clouds analogous to human brain nerve cells
and primary objective of the nerve cell is to communicate and process information from
our senses. The output is created using the trained neural network’s activation function
and the provided input. The multilayer perceptron (MLP) is a form of feedforward
neural network. MLP neural networks have several layers, and the convergence of
optimisation targets is much slower than in radial basis neural networks because they
only have one hidden layer [32]. Despite having only one layer, the RBFNN has been
shown to be a universal approximation. The RBFNN has a few applications, including
function approximation, interpolation, classification, and time series prediction [33].
The RBFNN is a feed-forward neural network with three layers: input, hidden, and
output [34]. Figure 4. depicts the usual architecture of the RBFNN. The input layer
is a vector, while the hidden layer is made up of n neurons. The radial basis function
activates each neuron in the buried layer, which has weights and biases. The weight
vector has the same dimension as the input vector.
The Euclidean distance between the weight vector w and the input vector x is
defined by
The activation function of the neural network is [34]
where, Ω a positive scalar value and j is the number of neurons in the hidden layer.
The neural network to approximate f will be

f (t) = W T h (x) + ε (15)

where, is the approximation error of the neural network; h (x) The Gaussian func-
tion of the neural network; W ideal neural network weights with the variations in

8
 z}|{
viscoelastic model parameters for the input x = [e ė], neural network to output f (t)
will be
z}|{ z}|{T
f (t) = W h (x) (16)
The desired control force is
z}|{
F (t) = m (ẍd + η.sgn (s) + c1 ė) − f (t) (17)
Then, the derivative of s variable is
 z}|{
1 1 e 
ṡ = − f (t) + m (ẍd + η.sgn (s) + c1 ė) − f (t) +x˙d (t)+ c1 ė = − f (t) +(η.sgn (s))
m m
(18)
where
z}|{ z}|{T
fe(t) = f (t) − f (t) = W T h (x) + ε − W h (x) = W
f T h (x) + ε (19)
z}|{
f=W − W
and W
The Lyapunov function defined for the RBSM controller as
1 2 1 fT f
V = s + γW W (20)
2 2
The derivative of V
Due to the approximation error,is limited and sufficiently small. To achieve!V̇ ≤ 0,
˙
z}|{
f T γ W + 1 sh (x) = 0
the design of the controller should be η ≥ ε, so that -W m

The adaptive law of the neural network is

˙
z}|{ 1 1
W =− s h(x) (21)
mγ
Equation 21 calculates estimated adaptive weights of the neural network. In
Equation 16, these weights are replaced to approximate the unknown resistive force
f (t). This f (t) is added to the sliding mode controller output to accomplish the MRE
isolator’s required control performance. The chattering component (sgn(s)) is abol-
ished with the inclusion of resistive force f (t) because the lower value is chosen for
the RBSM controller and compensated by the neural network component.

4.4 Block diagram of the RBSM controller on MRE isolator

Figure 5 depicts a block diagram of an RBSF controller. The MRE isolator is excited by
the excitation w (t) and generates response x (t). The error signal e (t), the difference
between the desired xd (t) and x (t) isolator response, is calculated and fed to the
sliding mode control and neural network. The sliding mode controller generates the
sliding surface s (t) with the error and error rate signals. The adaptive rule uses a
sliding surface and the Gaussian function h (t) to evaluate neural network weights.

9
Fig. 5 Block diagram of the RBSM control scheme

The neural network adjusts the weights and the property function f (t) is computed.
Finally, the control output of the sliding mode is added to the property function f (t),
and the control output F (t) is sent to the MRE isolator.

4.5 The controller performance in simulation

In this section, the performance of the sliding mode controller and RBSM controller
implemented on the MRE isolator in simulation is measured under the excitation of a
single sine input with varying amplitude and frequency. The performance evaluation
of the controller relies on the analysis of transmissibility plots, as presented in our
published paper [35], where essential data was extracted for assessment purposes.
The simulation amplitudes and frequencies were chosen based on the maximum
isolation condition, which is at the natural frequency of the system and the controller
stopping frequency. A total of ten conditions were chosen for vibration control via
controller. Five of the ten conditions were chosen at the natural frequency of the sys-
tem, 28 Hz, with varying amplitudes of excitation ranging from 0.00125 m to 0.00225
m. The other five conditions were chosen at the controller stopping frequency criteria,
which are 39.5 Hz, 35.16 Hz, 34.52 Hz, 34.25 Hz, and 33.78 Hz. In each set of condi-
tions, the neural network adjusts the weighting of radial basis functions by using the
reaching condition of a defined sliding surface.

Performance of the sliding mode controller

The performance of sliding mode controller is measured on the MRE isolator using sin-
gle sinusoidal excitation at 28 Hz frequency with amplitudes of excitation of 0.00125
m, 0.00150 m, and 0.00175 m. Without taking into account the model, the implemen-
tation of sliding mode control is considered. Because the desired response estimated
from the mathematical model is unknown, it is assumed that the desire is zero for the
sliding mode controller. The performance of the controller is now entirely dependent

10
 Table 1 Performance characteristics of MRE isolator at different frequencies and
amplitude (Data taken from our paper[36])

Transmissibility Transmissibility Amount

Frequency
Input of the SDOF system of the SDOF system of isolation
displacement (m) without with 3 A
(Hz)
current input current input %
28 0.00125 2.47 1.09 55.8%
39.5 0.00125 1.19 1.19 0%
28 0.00150 2.47 1.39 51.8%
35.16 0.00150 1.68 1.68 0%
28 0.00175 2.47 1.45 41.2%
34.52 0.00175 1.77 1.77 0%
28 0.00200 2.47 1.47 40.4%
34.25 0.00200 1.81 1.81 0%
28 0.00225 2.47 1.51 38.8%
33.78 0.00225 1.88 1.88 0%

Fig. 6 With the implementation of sliding mode controller (a), (c) and (e) Response of the MRE
isolator (b), (d) and (f) Controller force generation under 0.00125 m, 0.00150 m, 0.00175 m,0.002 m
and 0.00225 m amplitude of excitation at 28 Hz frequency

11
Fig. 7 Radial basis neural network

on the controller gains c and η. As shown in the figure 6, the sliding mode controller is
turned off for the first two seconds and then turned back on for the next two seconds.
For the first set of amplitude and frequency of excitation, the sliding mode controller
gains c1 and η were set to 15 and 35, respectively. With those gains, the controller
generates control force (Figure 6.9(b)), which is fed into the MRE isolator to isolate
the vibrations from the source by 55.8% (Figure 6(a)), as specified in the table. Simi-
larly, for amplitudes of excitation of 0.00150 m and 0.00175 m, the controller gains c1
and η were chosen to achieve the required isolation of 51.8% (Figure 6(c)) and 41.2%
(Figure 6(e)), respectively, which is 15 and 8, and 8 and 3. The controller output force
for those gains is shown in Figure 6(d) and Figure 6(f).
Based on the above controller performance analysis, it is concluded that in the
absence of a mathematical model for the MRE isolator, the sliding mode controller
gains must be changed every time. As a result, in order to effectively implement a
sliding mode controller over a wide frequency range, gains of the controller must be
scheduled for each set of conditions. For that, an addition of radial basis neural network
was chosen for this purpose for controlling the vibration over a wide frequency range.
The detailed controller performance of the RBSM controller is analyzed in the next
section.

Performance of the RBSM controller

The performance of the RBSM controller was investigated at various excitation ampli-
tudes and frequencies. A radial basis neural network was initially created for ten sets
of conditions which are listed in the table. The weight of the neural network is chosen
in such a way that the sliding mode controller produces the required control output for
the desired vibration isolation. The neural network automatically changes its weights
as the amplitude of excitation changes, causing the controller output from the sliding
mode controller to change and achieve the required isolation.
The structure of a radial basis neural network is 2-10-1. (2 neuron input layer-
10 neuron in the hidden layer-1 neuron in the output layer). Figure 7 depicts the
evaluation process of neural network for a given input. 10 scenarios are studied, each
with a different frequency and magnitude of excitation (Ten cases are listed in Table

12
Fig. 8 With the implementation of tuned RBSM controller (a) Response of MRE isolator with
and without controller (b) Control Force of the controller (c) Variation of excitation amplitude with
respect to time (d) Variation of excitation frequency with respect to time

Fig. 9 With the implementation of RBSM controller (a) Response of MRE isolator with and without
controller (b) Control Force of the controller (c) variation of excitation amplitude with respect to
time (d) variation of excitation frequency with respect to time

1 ). Because the input vector size was 10, the hidden layer takes into account 10
neurons. The weight of the linear layer and bias are selected so that only one neuron is
engaged for each error and error rate input. The weights of neural network and biases

13
are chosen
 as     T
0.00125 0.21991 0.83255 −1.469
0.00150 0.26389 0.83255  4.154 
     
0.00175 0.30787 0.83255 −4.823
     
0.00125 0.31023 0.83255  1.187 
     
0.00150 0.33137 b1 = 0.83255 W 2 = 1012  0  b2 = [260944770.7]
1
   
W = 0.00225 −1.552
 0.47755
0.83255
   
0.00200 0.35185 0.83255  3.680 
     
0.00175 0.37956 0.83255 −5.163
     
0.00225 0.39584 0.83255 −9.782
0.00200 0.43039 0.83255 4.966
(22)
With the neural network weights chosen, the RBSM controller was implemented on
the MRE isolator, with sliding mode gains c1 and η set to 130 and 0.1, respectively.
Because the controller gain η is 0.1, the controller output has no chattering effect. The
amplitude of excitation (0.00125 m to 0.00225 m) to the MRE isolator changes every
4 seconds over a 20 second period, as shown in Figure 8 (c). Similarly, the frequency
of sinusoidal excitation changes every 2 seconds over a 20 second period, as shown in
Figure 8 (d). For each set of condition, the controller is turned off for 1 second before
turning back on for the next second. The Figure 6.11(a) depicts the response of the
MRE isolator for each set of conditions with and without controller input (Figure 8(b))
to the MRE isolator. The maximum vibration isolation (55.8%) occurs between 2 and
4 seconds, corresponding to an excitation amplitude of 0.00125 m and a frequency
of 28 Hz. The control force of the controller corresponds to the maximum isolation
at 3 to 4 seconds. With the implementation of control, the isolation of vibration at
natural frequency 28 Hz decreases from 55.8% to 38.8% as the amplitude of excitation
increases from 0.00125 m to 0.00225 m, which is the desired performance from the
controller. The implementation of a controller at the frequencies of excitation of 39.5
Hz, 35.16 Hz, 34.52 Hz, 34.25 Hz, and 33.78 Hz results in the least amount of vibration
isolation (0%), as shown in the Figure 8(a). This is because at those frequencies, the
system already achieves isolation without the need for any control input.
The current control performance analysis for the RBSM controller was only for a
small set of excitation amplitudes and frequencies. Predicting performance in between
frequencies required the use of additional set of data or additional neuron. To test per-
formance in between the frequencies, the predicted frequency and amplitudes must be
closest to the chosen data in order to provide an accurate prediction. This is because
the activation function under consideration is a radial basis function. Considering the
facts, the excitation frequency of 29 Hz is chosen to predict the performance of the
controller for different excitation amplitudes (0.00125 m to 0.00225 m). Figure 9(a)
depicts the performance of the RBSM controller under these conditions. The excita-
tion amplitude changes every 2 seconds (Figure 9(c)), while the excitation frequency
remains constant throughout the simulation (Figure 9(d)). The controller is turned off
for 1 second and then turned back on for 1 second for each set of data (Figure 9(b)).

14
According to the response plot (Figure 9(a)), it is confirmed that for different ampli-
tude of excitation, the desired amount of isolation is achieved with the given control
input to the MRE isolator (Figure 9).

5 CONCLUSIONS
Developing a mathematical model for a nonlinear system is a difficult and time con-
suming undertaking. A model-free radial basis neural network sliding mode controller
is designed to eliminate the burden of constructing a mathematical model for con-
trol implementation. The 10 criteria are selected to evaluate the performance of the
controller. The performance result demonstrates that the developed controller is able
to modify the weights of the neural network such that the sliding mode controller
generates the necessary control output for vibration isolation.

Author contributions
Katari Kiran: Methodology, Software, Formal analysis, Review Editing, Data Cura-
tion. Umanath R Poojary: Conceptualization, Methodology, Software, Formal analysis,
Writing—Original Draft Investigation, Visualization. KV Gangadharan: Resources,
Supervision, Project administration.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research,
authorship, and/or publication of this article.

References
[1] Heinonen, J., Vessonen, I., Klinge, P., Lindroos, T.: Semi-active vibration con-
trol based on shape memory alloy actuators: Analysis and experimental testing.
Rakenteiden Mekaniikka 40(1), 23–38 (2007)

[2] Qiu, J., Ji, H., Zhu, K.: Semi-active vibration control using piezoelectric actuators
in smart structures. Springer (2009). https://doi.org/10.1007/s11465-009-0068-z
. https://link.springer.com/article/10.1007/s11465-009-0068-z

[3] Zhu, G.H., Li, Z.G., Xiong, Y.P., Xiao, L., Li, M.: Dynamic Mechanical Properties
of Magnetorheological Elastomers: Experiment and Modelling. In: IOP Confer-
ence Series: Materials Science and Engineering (2019). https://doi.org/10.1088/
1757-899X/559/1/012001

[4] Wongprasert, N., Symans, M.D.: Experimental Evaluation of Adaptive Elas-

tomeric Base-Isolated Structures Using Variable-Orifice Fluid Dampers. Journal
of Structural Engineering 131(6), 867–877 (2005) https://doi.org/10.1061/(asce)
0733-9445(2005)131:6(867)

15
 [5] Forte, P., Paternò, M., Rustighi, E.: A Magnetorheological Fluid Damper for
Rotor Applications. International Journal of Rotating Machinery 10(3), 175–182
(2004) https://doi.org/10.1155/s1023621x04000181

[6] Winthrop, M.F., Baker, W.P., Cobb, R.G.: A variable stiffness device selection
and design tool for lightly damped structures. Journal of Sound and Vibration
287(4-5), 667–682 (2005) https://doi.org/10.1016/j.jsv.2004.11.022

[7] Davis, L.C.: Model of magnetorheological elastomers. Journal of Applied Physics

85(6), 3348–3351 (1999) https://doi.org/10.1063/1.369682

[8] Lokander, M., Stenberg, B.: Performance of isotropic magnetorheological rub-

ber materials. Polymer Testing 22(3), 245–251 (2003) https://doi.org/10.1016/
S0142-9418(02)00043-0

[9] Wang, Y.F., Wang, G.F.: Study on the mechanical properties of magnetorheo-
logical elastomers. Advanced Materials Research 774-776, 54–57 (2013) https:
//doi.org/10.4028/www.scientific.net/AMR.774-776.54

[10] Li, Y., Li, J., Tian, T., Li, W.: Erratum: A highly adjustable magnetorheologi-
cal elastomer base isolator for applications of real-time adaptive control (Smart
Materials and Structures (2013) 22 (095020)). Smart Materials and Structures
23(12) (2014) https://doi.org/10.1088/0964-1726/23/12/129501

[11] Li, Y., Li, J., Li, W., Samali, B.: Development and characterization of a mag-
netorheological elastomer based adaptive seismic isolator. Smart Materials and
Structures 22(3) (2013) https://doi.org/10.1088/0964-1726/22/3/035005

[12] Kallio, M.: The elastic and damping properties of magnetorheological elastomers
(565), 3–146 (2005)

[13] elastomers, Hafeez, M., Usman, M., Umer, M.A., Hanif, A.: Recent progress in
isotropic magnetors: A review. Polymers 12(12), 1–35 (2020) https://doi.org/10.
3390/polym12123023

[14] Blom, P., Kari, L.: Amplitude and frequency dependence of magneto-sensitive
rubber in a wide frequency range. Polymer Testing 24(5), 656–662 (2005) https:
//doi.org/10.1016/j.polymertesting.2005.04.001

[15] Lejon, J., Kari, L.: Preload, frequency, vibrational amplitude and
magnetic field strength dependence of magnetosensitive rubber.
http://dx.doi.org/10.1179/146580109X12473409436823 38(8), 321–326 (2013)
https://doi.org/10.1179/146580109X12473409436823

[16] Zheng, H.M., Dong, D.D., Zhu, L.H.: Tunable stiffness and damping vibration
control strategy based on magnetorheological elastomer isolator. In: Applied
Mechanics and Materials (2014). https://doi.org/10.4028/www.scientific.net/

16
 AMM.543-547.1461

[17] Yu, Y., Li, Y., Li, J., Nguyen, T.N., Li, S., Erkmen, E.: Vibration control of MRE
isolator-embedded smart building using genetic algorithm. In: Proceedings of 30th
International Conference on Adaptive Structures and Technologies, ICAST 2019
(2019)

[18] Fu, J., Li, P., Wang, Y., Liao, G., Yu, M.: Model-free fuzzy control of a
magnetorheological elastomer vibration isolation system: Analysis and experi-
mental evaluation. Smart Materials and Structures (2016) https://doi.org/10.
1088/0964-1726/25/3/035030

[19] Gu, X., Yu, Y., Li, Y., Li, J., Askari, M., Samali, B.: Experimental study of semi-
active magnetorheological elastomer base isolation system using optimal neuro
fuzzy logic control. Mechanical Systems and Signal Processing (2019) https://
doi.org/10.1016/j.ymssp.2018.10.001

[20] Nguyen, X.B., Komatsuzaki, T., Iwata, Y., Asanuma, H.: Fuzzy Semiactive
Vibration Control of Structures Using Magnetorheological Elastomer. Shock and
Vibration 2017 (2017) https://doi.org/10.1155/2017/3651057

[21] Behrooz, M., Wang, X., Gordaninejad, F.: Modeling of a new semi-active/passive
magnetorheological elastomer isolator. Smart Materials and Structures (2014)
https://doi.org/10.1088/0964-1726/23/4/045013

[22] Du, H., Li, W., Zhang, N.: Semi-active variable stiffness vibration control of
vehicle seat suspension using an MR elastomer isolator. Smart Materials and
Structures 20(10) (2011) https://doi.org/10.1088/0964-1726/20/10/105003

[23] Opie, S., Woosoon, Y.: A tunable vibration isolator using a magnetorheological
elastomer with a field induced modulus bias. In: ASME International Mechanical
Engineering Congress and Exposition, Proceedings (2008). https://doi.org/10.
1115/IMECE2007-41505

[24] Bao Nguyen, X., Komatsuzaki, T., Thi Truong, H.: Novel semiactive suspension
using a magnetorheological elastomer (MRE)-based absorber and adaptive neural
network controller for systems with input constraints. Mechanical Sciences 11(2),
465–479 (2020) https://doi.org/10.5194/ms-11-465-2020

[25] Leng, D., Xu, K., Ma, Y., Liu, G., Sun, L.: Modeling the behaviors of magnetorhe-
ological elastomer isolator in shear-compression mixed mode utilizing artificial
neural network optimized by fuzzy algorithm (ANNOFA). Smart Materials and
Structures (2018) https://doi.org/10.1088/1361-665X/aadfa9

[26] Liao, G., Fu, J., Li, P., Yu, M.: Neural network modeling of magneto-rheological
elastomer isolator. In: Proceedings of the 28th Chinese Control and Decision
Conference, CCDC 2016 (2016). https://doi.org/10.1109/CCDC.2016.7531434

17
[27] Kärkkäinen, K.K.: Effective permittivity of mixtures: numerical validation by the
FDTD method. IEEE Transactions on Geoscience and Remote Sensing (2000)
https://doi.org/10.1109/36.843023

[28] Khairi, M.H.A., Fatah, A.Y.A., Mazlan, S.A., Ubaidillah, U., Nordin, N.A.,
Ismail, N.I.N., Choi, S.B., Aziz, S.A.A.: Enhancement of particle alignment using
silicone oil plasticizer and its effects on the field-dependent properties of mag-
netorheological elastomers. International Journal of Molecular Sciences 20(17),
4085 (2019) https://doi.org/10.3390/ijms20174085

[29] Nguyen, X.B., Komatsuzaki, T., Iwata, Y., Asanuma, H.: Modeling and semi-
active fuzzy control of magnetorheological elastomer-based isolator for seismic
response reduction. Mechanical Systems and Signal Processing (2018) https://
doi.org/10.1016/j.ymssp.2017.08.040

[30] Yu, Y., Li, Y., Li, J.: Forecasting hysteresis behaviours of magnetorheological
elastomer base isolator utilizing a hybrid model based on support vector regression
and improved particle swarm optimization. Smart Materials and Structures 24(3),
35025 (2015) https://doi.org/10.1088/0964-1726/24/3/035025

[31] Liu, J., Wang, X.: Advanced Sliding Mode Control for Mechanical Systems,
(2011). https://doi.org/10.1007/978-3-642-20907-9

[32] Jayawardena, A.W., Fernando, D.A.K., Zhou, M.C.: Comparison of Multilayer

Perceptron and Radial Basis Function networks as tools for flood forecasting.
Technical Report 239 (1996)

[33] Wu, Y., Wang, H., Zhang, B., Du, K.-L.: Using Radial Basis Function Networks
for Function Approximation and Classification. International Scholarly Research
Network ISRN Applied Mathematics 2012 (2012) https://doi.org/10.5402/2012/
324194

[34] Park, J., Sandberg, I.W.: Universal Approximation Using Radial-Basis-Function

Networks. Neural Computation (1991) https://doi.org/10.1162/neco.1991.3.2.
246

[35] Kiran, K., Poojary, U.R., Gangadharan, K.: Fractional-order viscoelastic model-
ing of the magnetic field dependent transmissibility response of MRE isolator:.
https://doi.org/10.1177/1045389X221087172, 1045389–2210871 (2022) https://
doi.org/10.1177/1045389X221087172

[36] Kiran, K., Poojary, U.R., Gangadharan, K.V.: Developing the viscoelastic model
and model-based fuzzy controller for the MRE isolator for the wide frequency
range vibration isolation. Journal of the Brazilian Society of Mechanical Sciences
and Engineering 44(7) (2022) https://doi.org/10.1007/s40430-022-03575-y

18