International Journal of Power Electronics and Drive Systems (IJPEDS)

Vol. 13, No. 4, December 2022, pp. 2529~2540

ISSN: 2088-8694, DOI: 10.11591/ijpeds.v13.i4.pp2529-2540  2529

Design variable structure fuzzy control based on deep neural

network model for servomechanism drive system

Mohamed A. Shamseldin1, Abdel Halim M. Bassiuny2, Abdel Ghany M. Abdel Ghany3

1
Department of Mechanical Engineering, Faculty of Engineering, Future University in Egypt, Cairo, Egypt
2
Department of Mechatronics Engineering, Faculty of Engineering, Heliopolis University, El Salam City, Egypt
3
Department of Electrical Engineering, Higher Engineering Institute, Thebes Academy Cairo, Egypt

Article Info ABSTRACT

Article history: This paper presents a new scheme for variable structure (VS) fuzzy PD
controller. The rule base of the fuzzy PD controller is tuned online. The
Received Dec 3, 2021 purpose of the proposed controller is to track accurately a preselected
Revised Jul 18, 2022 position command for the servomechanism system. Therefore, this study
Accepted Aug 1, 2022 establishes a model using a black-box modeling approach; simulations were
performed based on real-time data collected by LabVIEW and processed
using MATLAB. The input signal for the servomechanism driver is a
Keywords: pseudo-random binary sequence that considers violent excitation in the
frequency interval. The candidate models were obtained using linear least
Deep neural network squares, nonlinear least squares, and deep neural network (DNN). The
Fuzzy logic validation results proved that the identified model based on DNN has the
Harmony research smallest mean square errors. Then, the DNN identified model was used to
Model reference design the proposed control techniques. A comparison had been executed
Servomechanism between the VS fuzzy PD control, the conventional PD control, and the
Variable structure fixed structure fuzzy PD control. The experimental results confirm the
proposed VS fuzzy PD control can absorb the nonlinear behavior of the
system. The speed regulation test, it reduces the rise time from 50% to 56%.
While continuously changing in speed, it has the smallest tracking error
(0.412 inches).
This is an open access article under the CC BY-SA license.

Corresponding Author:
Mohamed A. Shamseldin
Department of Mechanical Engineering, Faculty of Engineering, Future University Egypt
End of 90th St. Fifth Settlement, New Cairo, Egypt
Email: [email protected]

1. INTRODUCTION
The one-stage servomechanism system is the basic unit of most CNC machines [1]. These types of
machines need accurate motion control to produce high-quality products [2], [3]. Also, the high speed and
high accuracy position applications have been rapidly and broadly developed to facilitate the complex
automatic processes in the industrial field [4], [5]. Generally, the servomechanism systems don’t have an
accurate estimated parameter for friction and backlash models which leads to system uncertainty. The friction
causes high steady-state error while the backlash reasons undesired vibration in the system [4]. The models
that can describe the friction phenomenon in servomechanism systems suffer from parameter uncertainty [6].
So, the control task for this operation faces a big challenge [7]. There are many design schemes for single-
axis controllers such as conventional PID control, self-tuning proportional–integral–derivative (PID) control,
fuzzy PID control, adaptive control, and sliding mode control [8]. The advantage of self-tuning and sliding
mode techniques is that the tracking performance is good under the different operating points of position
commands [9], [10]. In several applications, the sliding mode control has been applied such as microgripper

Journal homepage: http://ijpeds.iaescore.com

2530  ISSN: 2088-8694

position/force control as presented in [11] and the position control of X-Y stage servomechanism as
demonstrated in [12].
On the other side, the disadvantages are that the chattering phenomenon can be an obstacle to
improve the control accuracy, and in some cases, it causes instability in the system [13], [14]. The fuzzy logic
system is a suitable technique to treat the uncertainty and nonlinear problems of complex systems [15], [16].
Sometimes, the fixed rule base structure of the fuzzy system cannot deal with system uncertainty and
nonlinearity so, several previous studies had attempted to use different techniques to develop an adaptive
mechanism that can tune online the rule base or normalizing gains as in [17], [18]. However, the
disadvantage of these mechanisms takes a long time to adjust accurately the rule base and need high
processing, so it cannot execute practically [19], [20].
It is known that the model reference adaptive control (MRAC) is highly efficient adaptive control
mechanism where it forces the overall system to follow the behavior of a designed model reference [21]. The
selected model reference can be a first or second-order system according to the point of view of the designer
[22], [23]. Experimental identification is a well-recognized methodology to obtain a precise process model,
often intended for control but also other purposes. The nature of the input signal for the identification has a
great effect on the accuracy of the model. But, in many cases, the input signal cannot be easily selected
concerning plant behavior constrictions. pseudo-random binary sequences (PRBS) are often used as violent
excitation signals for system identification, due to it has a finite length and can be synthesized frequently
with simple generators while presenting favorable spectra for identification [24].
The family of candidate models for system identification can be classified along with several
different aspects. The linear and nonlinear theory has been well developed and investigated for real
applications throughout recent years. The linear structure is used to simplify the analysis where the
parameters are constant and do not vary throughout a simulation, such as the autoregressive with exogenous
inputs (ARX) model. In contrast, a non-linear model presents dependent parameters that are permitted to vary
throughout a simulation run, and its use becomes necessary where interdependencies between parameters
cannot be considered insignificant [25].
Lately, deep learning has become attractive and has significant attention from a wide range of
engineering applications. Compare to traditional neural networks, the vital features of deep learning are to
have more hidden layers and neurons and to improve learning performance [26]. Using these features,
complex and large problems that could not be solved with traditional neural networks can be resolved by
deep learning algorithms [27]. Therefore, deep learning has been subjected to various applications including
pattern recognition and classification problems; for example, handwritten digit recognition, speech
recognition, human action recognition, and so on. However, to the best knowledge of the authors, no result
has been published in the system identification and automatic control field [28]. Thus, this paper focuses on
presenting the applying possibility of deep learning in system identification areas.
This study investigates the performance of a one-stage servomechanism system using several
advanced control techniques. Firstly, identified models using linear least squares, nonlinear least squares, and
DNN had obtained and selected the best between them. Both the conventional PD and fixed rule base fuzzy
PD controller had investigated but still, the performance is not acceptable. Therefore, a new variable structure
(VS) fuzzy PD controller had designed to tune the rule base online using an adaptive mechanism with fast
calculation to can apply practically. The parameters of the adaptive mechanism had obtained based on the
selected model reference. The optimal parameters of the proposed control techniques had been determined
using harmony search (HS) optimization technique. The paper is prepared as follows, firstly, the
experimental setup is presented. Secondly, system identification is explained. Thirdly, the proposed
controller techniques are demonstrated. Fourthly, the experimental results are illustrated. Finally, the
conclusion is discussed.

2. EXPERIMENTAL SETUP
This section demonstrates the main parts of the one-stage servomechanism system prototype. Also,
it shows the open-loop performance of the servomechanism system. Figure 1 illustrates the main components
of one stage servomechanism experimental setup which consists of a DC motor electro-mechanical module.
The stroke of the stage ranges from 0 to 9 Inches.
The DC motor has a nominal speed of 1800 rev/min, and an armature voltage of 90 V dc. The
optical encoder is an add-on that provides position feedback signal (200 pulses per revolution). Two
magnetic limit switches detect when the sliding block reaches the start or end position. The DC motor drive
controls the DC motor electro-mechanical module. This versatile drive also allows an external signal to
control the motor speed. A data acquisition card (DAQ) NI USB-6009 is connected to the computer that is
used to perform the control algorithms. The main idea of the program has been designed to make the NI

Int J Pow Elec & Dri Syst, Vol. 13, No. 4, December 2022: 2529-2540
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2531

DAQ 6009 generates an analog output signal (-5 to 5 V) to the linear amplifier. Also, the analog output signal
from the optical encoder has been collected at the same time. The speed of the DC motor will fluctuate when
the generated signal change continuously. The positive signal will cause the DC motor speed to fluctuate in
the forward direction, while the DC motor will fluctuate in the reverse direction through the negative voltage
ranges. The shaft of the optical encoder is coupled with the lead screw shaft to measure the speed and
position of the stage as in Figure 2.

Figure 1. The one stage servomechanism Figure 2. The block diagram of the experimental setup
experimental setup servomechanism system

Figure 3 demonstrates the PRBS output signal from the NI DAQ card. It can be noted that the signal
has a variable frequency at an acceptable range. Then, this signal will be input to the DC motor driver. Figure 4
shows the corresponding linear speed of the stage measured by the optical encoder. It is clear that the linear
speed was fluctuating highly due to the violent excitation of the servomechanism system. Therefore, the stage
position increases in positive ranges of the input signal entering to DC motor driver while the actual stage
position decreases in negative ranges of the input signal. The input/output data will be collected and stored in
an excel sheet file and then this data will be used to develop an identified model for the experimental setup.

Figure 3. The random output signal of the NI DAQ Figure 4. The output linear speed of the
card servomechanism table

3. SYSTEM IDENTIFICATION
This section investigates an accurate model for a one-stage servomechanism system based on
measured input/output data. Three system identification techniques will be used. The first technique uses
linear least squares. The second technique is the nonlinear least squares. Lastly, the third technique develops
a DNN-identified model.

3.1. Linear least squares model

The general linear transfer function can be expressed as follows:
Design variable structure fuzzy control based on deep neural network model … (Mohamed A. Shamseldin)
2532  ISSN: 2088-8694

𝑦(𝑠) 𝑘
= (1)
𝑢(𝑠) 𝑏𝑛.𝑆 𝑛 +𝑏𝑛−1 .𝑆 𝑛−1 +⋯+𝑏0

In (1), 𝑘, 𝑏𝑛 , … , 𝑏0 are the estimated parameters of the approximate transfer function (T.F). The y(s) stands
for linear stage speed while u(s) is the input voltage to the DC motor drive. The accuracy of the transfer
function improves significantly when the system degree is increased. However, often there is a restriction
that increasing order cannot make the model accurate sufficiently [29]. Therefore, it is necessary to explicitly
add the nonlinearities into the system [30].

3.2. Nonlinear least squares model

As the servomechanism systems suffer from uncertainty and complex nonlinear dynamics, a
nonlinear ARX (NLARX) model structure has been developed to model such systems [31]. The NLARX
model consists of several complicated nonlinear functions that can model the backlash and friction in
servomechanism systems significantly. An NLARX model can be understood as an extension of a linear
model. A linear SISO ARX model has this structure [32].

𝑦(𝑡) + 𝑎1 𝑦(𝑡 − 1) + 𝑎2 𝑦(𝑡 − 2) + ⋯ + 𝑎𝑛𝑎 𝑦(𝑡 − 𝑛𝑎) = 𝑏1 𝑢(𝑡) + 𝑏2 𝑢(𝑡 − 1) + ⋯ + 𝑏𝑛𝑏 𝑢(𝑡 − 𝑛𝑏 + 1) + 𝑒(𝑡) (2)

Where u is input, y is output, 𝑛𝑎 is the number of past output terms and 𝑛𝑏 is the number of past input terms.
This structure can be extended to create a nonlinear form where instead of the weighted sum that represents a
linear mapping, the NLARX model has a more flexible nonlinear mapping function [33].

𝑦(𝑡) = 𝑓(𝑦(𝑡 − 1), 𝑦(𝑡 − 2), 𝑦(𝑡 − 3) … . , 𝑢(𝑡), 𝑢(𝑡 − 1), 𝑢(𝑡 − 2), … ) (3)

Where f is a nonlinear function (to simulate the behavior of friction and backlash that exist in
servomechanism systems). Inputs to f are model regressors.

3.3. Deep neural network model

There are two main types of ANN, the first type is the shallow neural network and the second type is
the deep neural network. Deep learning has several layers of hidden units and it also permits many more
factors to be used before over-fitting occurs. Thus, for deep learning, a deep architecture is used. The neural
network structure in this study contains three layers the first layer is the input layer which receives the input
PRBS signal to the DC motor driver. The second layer is the hidden layer that contains several hidden
neurons and receives data from the input layer. The third layer is the output layer which presents the
corresponding linear stage speed. 𝑥𝑖 = (𝑥1 , 𝑥2 , 𝑥3 , … , 𝑥𝑛 )𝑡 Input vector applied to the layer, the whole of the
hidden neuron input ‘𝑗 ’ is:

𝑛𝑒𝑡𝑗ℎ = ∑𝑛𝑖=1 𝑤𝑖𝑗 𝑋𝑖 + 𝜃𝑗ℎ (4)

With 𝑋𝑖 = (𝑋1 , 𝑋2 , 𝑋3 , … , 𝑋𝑛 )𝑡 (5)

such as 𝑖 = 1,2, … , 𝑁ℎ

𝑥𝑖 is the input vector applied to the layer, and 𝑤𝑖𝑗 is the weights of (𝑖) input neuron connection, and 𝜃𝑗ℎ
represent the bias of hidden layer neurons. The neurons of the hidden layer can be written as follows:

𝑦𝑗ℎ = 𝑓(∑𝑛𝑖=1 𝑤𝑖𝑗 𝑋𝑖 + 𝜃𝑗ℎ ) (6)

3.4. Candidate identified models validation

The obtained input/output data was used to develop a set of three candidate-identified models to
select the best between them. Three models had been obtained, the first identified model is a linear second-
order system while the second identified model is an NLARX model and lastly, the DNN identified model.
The linear second-order system has the following specifications.

𝑦(𝑠) 27.05
= (7)
𝑢(𝑠) 𝑠 2 +13.91𝑠+78.45

By matching with in (1), the estimated parameters of the linear approximation transfer function are 𝑘 =
27.05, 𝑛 = 2, 𝑏𝑛 = 1, 𝑏𝑛−1 = 13.91 𝑎𝑛𝑑 𝑏0 = 78.45. On the other hand, the obtained discrete-time ARX

Int J Pow Elec & Dri Syst, Vol. 13, No. 4, December 2022: 2529-2540
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2533

model has the following structure.

𝐴(𝑧)𝑦(𝑡) = 𝐵(𝑧)𝑢(𝑡) + 𝑒(𝑡) (8)

Where:
𝐴(𝑧) = 1 − 0.7564 𝑧 −1 − 0.09098 𝑧 −2
𝐵(𝑧) = 0.03327 𝑧 −1 + 0.01742 𝑧 −2
By corresponding with in (2), the polynomial orders are 𝑛𝑎 = 2 and 𝑛𝑏 = 2. Also, the used nonlinear
function is a wave net with 25 units. Figure 5 demonstrates the linear speed of the actual experimental setup
and the candidate models. It is noted that the identified model based on DNN can simulate the behavior of an
actual experimental setup compared to other identified models. Table 1 demonstrates the mean square error
of candidate models. It can be noted that the identified model based on DNN has a minimum error compared
to other identified models.

Figure 5. The linear speed of one stage table servomechanism for actual experimental setup and identified
models

Table 1. Mean square error of candidate models

No. System identification method Mean square error
1 Identified Model (Second Order T.F.) 0.1973
2 Identified System (Nonlinear ARX model) 0.05912
3 Identified NARX neural network model 0.00071

4. VARIABLE STRUCTURE FUZZY PD CONTROL

4.1. VS fuzzy PD control
There are two types of adaptive fuzzy controllers. The first type tunes the rules of the fuzzy system
so, it is called self-organizing controllers or variable structure controllers [34]. In the second type of fuzzy
controller, the scaling factors are modified online and can be called self-tuning controllers [35]. The previous
studies showed that the first type of adaptive fuzzy controller is more effective than the second type [36].
But, the first type of adaptive fuzzy controller is more complicated than the second type where the first type
needs building the fuzzy system by the designer without using the fuzzy toolbox for any software program
[37]. The proposed technique has an adaptive mechanism to tune the centroid of rule base membership online
based on the optimal model reference adaptive system. Figure 6 illustrates the general structure of variable
structure fuzzy PD control.
The variable structure fuzzy PD control input/output relationship can be described as follows:

𝑢(𝑠) = 𝐺𝑈. ∆𝑢(𝑠) (9)

∆𝑢(𝑠) = 𝜃 𝑇 ʓ(𝑥) (10)

𝐸(𝑠) = 𝐺𝐸. 𝑒 (11)

𝐶𝐸(𝑠) = 𝐺𝐶𝐸. 𝑐𝑒(𝑠) (12)

Design variable structure fuzzy control based on deep neural network model … (Mohamed A. Shamseldin)
2534  ISSN: 2088-8694

𝑘𝑑
𝐺𝐸 = 1, 𝐺𝐶𝐸 = , 𝐺𝑈 = 𝑘𝑝 (13)
𝑘𝑝

Figure 6. Variable structure fuzzy PD (VSFPD) control

Where u(s) is the output fuzzy controller, where 𝜃 𝑇 is the centroid vector of the output membership
function, ʓ(𝑥) is the vector of the fuzzy basis function, 𝑥 ∈ [−1,1], and GU is the scaling factor for the fuzzy
PD controller. Inputs of the fuzzy controller are scaled using, 𝐺𝐶𝐸 scaling factors also known as fuzzy gains,
𝑒(𝑠) is the error between the reference position and the actual position of the stage and 𝑐𝑒(𝑠) is the change
of this error as shown in Figure 7. Here, we introduce a reference model in the structure of the fuzzy
controller to generate the model error given by:

𝑒𝑚 (𝑠) = 𝑦(𝑠) − 𝑦𝑚 (𝑠) (14)

In (7), the 𝑦(𝑠) is the actual system output, the 𝑦𝑚 (𝑠) stands for the desired performance and 𝑒𝑚 (𝑠)
considers the difference between the actual process value and the expected value of output. The reference
model can be a first or second-order system. The model reference transfer function contains the desired
response of the system such as the desired damping ratio, the desired natural frequency, the desired rise time,
the desired settling time, and the desired overshoot. If the order of the reference model is a stable first-order
system as the following transfer function.

𝑦𝑚 (𝑠) 𝑘𝑚
=𝑡 (15)
𝑢𝑐 (𝑠) 𝑚 𝑠+1

Where 𝑢𝑐 is the reference position, 𝑦𝑚 is the output of the reference model, 𝑘𝑚 represents the DC gain of the
system ratio between the input signal and the steady-state value of output and 𝑇𝑚 is the time constant that
measures how quickly a first-order system responds to a unit step input. The MIT rule is the original
approach to MRAC. The name is derived from the fact that it was developed at the Instrumentation
Laboratory (now the Draper Laboratory) at MIT. To adjust parameters in such a way that the loss function is
minimized [38].

21
𝑗(𝜃) = 𝑒𝑚 (16)
2

To make j small, it is reasonable to change the parameters in the direction of the negative gradient of
j, that is,

𝑑𝜽 𝜕𝑗 𝜕𝑒𝑚
= −𝛾 = −𝛾𝑒𝑚 (17)
𝑑𝑡 𝜕𝜽 𝜕𝜽

where 𝛾 stands for the adaptation gain while 𝜃 is the centroid vector of the output membership function.
Figure 8 demonstrates the adaptive output membership functions for VSFC. The linguistic labels of the
outputs are {NB(𝜃1 ), NM(𝜃2 ), NS(𝜃3 ), ZE(𝜃4 ), PS(𝜃5 ), PM(𝜃6 ), PB(𝜃7 )}. The output memberships centers
are not fixed and change continuously through a certain range to prevent the overlapping between the
memberships. Table 2 summarizes the adaptive rule base of VS fuzzy system.
From the adaptive output of the membership function can outcome the following equations.

Int J Pow Elec & Dri Syst, Vol. 13, No. 4, December 2022: 2529-2540
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2535

𝜃(0) = [ −1 − 0.66 − 0.33 0 0.33 0.66 1] (18)

𝜃 = [𝜃1 𝜃2 𝜃3 𝜃4 𝜃5 𝜃6 𝜃7 ] (19)

The used defuzzification technique is the center of gravity.

∑𝑛
𝑗=1 𝑢(𝑢𝑗 )𝑢𝑗
𝑢(𝑛𝑇) = ∑𝑛
= 𝜃 𝑇 ʓ(𝑥) (20)
𝑗=1 𝑢(𝑢𝑗 )

Where 𝑢 (𝑢𝑗 ) membership grad of the element 𝑢𝑗 , 𝑢(𝑛𝑇) is the fuzzy control output, 𝑛 is the number of
discrete values in the universe of discourse.

Figure 7. The input of membership functions Figure 8. The adaptive output membership functions
(error and change of error)

Table 2. The adaptive fuzzy rule base

E/CE NB NM NS ZE PS PM PB
NB NB(𝜃1 ) NB(𝜃1 ) NB(𝜃1 ) NB(𝜃1 ) NM(𝜃2 ) NS(𝜃3 ) ZE(𝜃4 )
NM NB(𝜃1 ) NB(𝜃1 ) NB(𝜃1 ) NM(𝜃2 ) NS(𝜃3 ) ZE(𝜃4 ) PS(𝜃5 )
NS NB(𝜃1 ) NB(𝜃1 ) NM(𝜃2 ) NS(𝜃3 ) ZE(𝜃4 ) PS(𝜃5 ) PM(𝜃6 )
ZE NB(𝜃1 ) NM(𝜃2 ) NS(𝜃3 ) ZE(𝜃4 ) PS(𝜃5 ) PM(𝜃6 ) PB(𝜃7 )
PS NM(𝜃2 ) NS(𝜃3 ) ZE(𝜃4 ) PS(𝜃5 ) PM(𝜃6 ) PB(𝜃7 ) PB(𝜃7 )
PM NS(𝜃3 ) ZE(𝜃4 ) PS(𝜃5 ) PM(𝜃6 ) PB(𝜃7 ) PB(𝜃7 ) PB(𝜃7 )
PB ZE(𝜃4 ) PS(𝜃5 ) PM(𝜃6 ) PB(𝜃7 ) PB(𝜃7 ) PB(𝜃7 ) PB(𝜃7 )

4.2. Harmony search (HS) optimization

The HS is a phenomenon-mimicking algorithm inspired by the improvisation process of musicians
[39]. Each controller variable corresponds to each musician [40]. The range of the parameter’s value
corresponds to the musical instrument’s pitch range [41]. The solution matrix corresponds to the musical
harmony at a certain time. In this paper, the HS optimization technique was used to obtain the optimal values
of VS fuzzy PD controller based on MRAS (𝜸 is the adaptation gain vector, 𝑘𝑚 is the DC gain, 𝑇𝑚 is the
desired constant time and the scaling factors of the fuzzy system). The initial population of Harmony
Memory (HM) is chosen randomly through a preselected range for each controller parameter. HM consists of
harmony memory solution (HMS) matrix as in (21).

𝐺𝑈 (1,1) 𝐺𝐸 (1,2) 𝐺𝐶𝐸 (1,3) 𝛾1 (1,4) … 𝛾7 (1,10) 𝐾𝑚 (1,11) 𝑇𝑚 (1,12)

𝐺𝑈 (2,1) 𝐺𝐸 (2,2) 𝐺𝐶𝐸 (2,3) 𝛾1 (2,4) … 𝛾7 (2,10) 𝐾𝑚 (2,11) 𝑇𝑚 (2,12)
𝐻𝑀 = .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
(21)
. . . . . . . .
[𝐺𝑈 (𝐻𝑀𝑆,1) 𝐺𝐸 (𝐻𝑀𝑆,2) 𝐺𝐶𝐸 (𝐻𝑀𝑆,3) 𝛾1 (𝐻𝑀𝑆,4)
… 𝛾7 (𝐻𝑀𝑆,10)
𝐾𝑚 (𝐻𝑀𝑆,11) 𝑇𝑚 (𝐻𝑀𝑆,12) ]

Similarly, the HM matrix has been established for other proposed controller parameters (classical
PD control and fixed structure fuzzy control). The objective function corresponds to the audience’s aesthetics
to be minimized or maximized. The used objective function is shown in (22).
1
𝑓= (22)
(1−𝑒 −𝛽 )(𝑀𝑝 +𝑒𝑠𝑠 )+𝑒 −𝛽(𝑡𝑠 −𝑡𝑟)

Where 𝑀𝑝 is the overshoot of system response, 𝑒𝑠𝑠 is the steady-state error, 𝑡𝑠 is the settling time and 𝑡𝑟 is the
rise time. Also, this objective function can compromise the designer demand by the weighting parameter
Design variable structure fuzzy control based on deep neural network model … (Mohamed A. Shamseldin)
2536  ISSN: 2088-8694

value (β). The parameter is set larger than 0.7 to reduce overshoot and steady-state error. If this parameter is
set smaller than 0.7 the rise time and settling time will be reduced.
The HS optimization program contains several steps. The first step, initialize the HS parameters. In
the second step, generate random values for the HM matrix and determine the fitness function corresponding
to each solution vector. The third step, Improvise a new harmony from the HM matrix. The fourth step,
replace the worst solutions with the best solutions. In the fifth step, repeat step 2 and step 4 until the
termination criterion is satisfied. Table 3 demonstrates the obtained parameter values using HS optimization
for the proposed controllers.

Table 3. The parameter values of the proposed controllers

Controller Parameter Value
PD controller 𝑘𝑝 26.1677
𝑘𝑑 1.7924
Fuzzy PD controller 𝐺𝐸 0.087
𝐺𝐶𝐸 0.0456
𝐺𝑈 3.01
VS Fuzzy PD controller 𝑘𝑚 1.01
𝑇𝑚 0.102
𝐺𝐸 0.092734
𝐺𝐶𝐸 0.00054574
𝐺𝑈 40
𝛾1 0.04
𝛾2 0.045
𝛾3 0.056
𝛾4 0.078
𝛾5 0.089
𝛾6 0.0233
𝛾7 0.0123

5. EXPERIMENTAL RESULTS
This section demonstrates the motion behavior of one stage servomechanism system using the
variable structure fuzzy PD control. Also, the performance of the proposed controller technique had been
investigated by comparing it with the PD control and the fixed structure fuzzy PD control. Two tests had
been applied to the experimental setup. In the first test, the position command was adjusted at a constant
value while the second test was executed at the variable position command.

5.1. Constant position command

In this test, the position command was adjusted at a constant value (7 inches) and the stage began its
motion from the position reference (zero inches). Figure 9 shows the position responses of the stage using
three various control techniques. It can be noted that the variable structure fuzzy PD control has a minimum
rise time and minimum settling time compared to other control techniques. Also, the variable structure fuzzy
PD control has no overshoot approximately as demonstrated in Table 4.
Figure 10 demonstrates the corresponding linear speed responses of the stage through the
experiment. It can be obvious that the variable structure fuzzy PD control has a high speed at the rise time.
Also, it is decaying rapidly at moment 2 seconds to compensate for the error in the stage position. Then, the
stage speed stables at zero approximately.

Figure 9. The stage position responses of proposed Figure 10. The stage speed responses of proposed
controllers at constant position command controllers at constant position command

Int J Pow Elec & Dri Syst, Vol. 13, No. 4, December 2022: 2529-2540
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2537

Figure 11 demonstrates the proposed controller's output through the experiment. The output signal
in the case of the VS fuzzy PD control referred to the output after the gain 𝐺𝑈. It is clear that all controllers
give the maximum value at the beginning of the test which makes the stage accelerate rapidly to reach the
desired position. Figure 12 illustrates the online change in the output membership centers through the
constant position command test. It is noted that the adaptive mechanism tunes (𝜃1, 𝜃2, 𝜃3, and 𝜃4) where the
fuzzy output value locates through these memberships so, the other membership centers (𝜃5, 𝜃6, and 𝜃7) are
not changed.

Table 4. The performance at constant position command test

Controller type Rise time (sec) Settling time (sec) Overshoot (%)
The HS PD control 2.7123 3.5678 0.2123
The fuzzy PD control 1.6512 2.6565 2.0702
The VS fuzzy PD control 1.2344 1.8343 0.1778

Figure 11. The stage position responses of proposed Figure 12. The change in membership centers during
controllers at constant position command the constant position command test

5.2. Variable position command

The position command in this test changes continuously to investigate the robustness of the
proposed controller technique and its ability to track a complex position command profile. Figure 13
illustrates the position responses of the stage by several control techniques and under variable position
command (half-sine wave). It is obvious that the variable structure fuzzy PD control can follow accurately
the complicated trajectory compared to other control techniques. Also, the fixed structure fuzzy PD control
has acceptable tracking accuracy but it has a high error at the beginning of tracking while the harmony search
(HS) PD control has a high deviation from the position command profile. Moreover, the variable structure
fuzzy PD control has the smallest mean square error throughout the experiment as demonstrated in Table 5.
Figure 14 demonstrates the corresponding linear speed responses of the stage for control techniques. It can be
noted that the stage speed through the variable structure fuzzy PD control has a high value in the first seconds
of the experiment and then the stage speed decreases gradually to track the position command profile.

Figure. 13. The stage position responses of each Figure 14. The stage speed responses of control
control technique at variable techniques at variable position command
Design variable structure fuzzy control based on deep neural network model … (Mohamed A. Shamseldin)
2538  ISSN: 2088-8694

Table 5. The mean square error of proposed controllers at variable position reference command
Controller Type Mean square error (inch)
The PD control 0.912
The fuzzy PD control 0.7065
The VS fuzzy PD control 0.412

Figure 15 shows the controller's output through the variable position command test. At the start of
the experiment, the controllers have a high output value. These signals decrease gradually with different
behaviors. The controller's signal declines until the polarity of signals change which leads to the reverse of
the stage motion direction. Figure 16 shows the online change in the output membership centers during the
variable position command test. It is obvious that the adaptive mechanism changes (𝜃1, 𝜃2, 𝜃3, and 𝜃4)
where the fuzzy output value belongs to these memberships so, the other membership centers (𝜃5, 𝜃6, and
𝜃7) do not change until the end of the test.

Figure 16. The change in membership centers during Figure 15. The control techniques output at variable
the variable position command test position command

6. CONCLUSION
A new variable structure technique has been developed to tune the rule base of fuzzy logic control
online based on optimal model reference adaptive system (MRAS). This work investigates the robustness of
the proposed technique by applying it to one stage servomechanism system. The controller aims to track
accurately a preselected position reference trajectory in presence of the friction and backlash problems. Also,
the performance of the proposed controller technique has been compared to the classical PD control and the
fixed structure fuzzy PD control to ensure robustness. There are two tests that have been implemented to
investigate each control technique. The first test adjusts the position reference at a constant value while the
second test adjusts the position reference to change continuously with time. The experimental results
demonstrate that the variable structure fuzzy PD control based on an optimal model reference adaptive
system can eliminate the tracking error quickly compared to other control techniques which will save the
machining time cycles.

REFERENCES
[1] P. Zhao, J. Huang, and Y. Shi, “Nonlinear dynamics of the milling head drive mechanism in five-axis CNC machine tools,”
International Journal Advanced Manufacturing Technology, pp. 3195-3210, 2017, doi: 10.1007/s00170-017-9989-6.
[2] C. Abeykoon, “Single screw extrusion control: A comprehensive review and directions for improvements,” Control Engineering
Practice, vol. 51, pp. 69–80, 2016, doi: 10.1016/j.conengprac.2016.03.008.
[3] W. Lee, C. Lee, Y. H. Jeong, and B. Min, “Friction compensation controller for load varying machine tool feed drive,”
International Journal of Machine Tools and Manufacture, vol. 96, pp. 47–54, 2015, doi: 10.1016/j.ijmachtools.2015.06.001.
[4] C. Wang, M. Yang, W. Zheng, X. L.V., K. Hu and D. Xu, “Analysis of limit cycle mechanism for two-mass system with backlash
nonlinearity,” Annual Conference of the IEEE Industrial Electronics Society, 2016, pp. 500-505, doi:
10.1109/IECON.2016.7793734.
[5] F. Wang, C. Liang, Z. Ma, X. Zhao, Y. Tian and D. Zhang, “Dynamic analysis of an XY positioning table,” in International
Conference on Manipulation, Manufacturing and Measurement on the Nanoscale, 2013, pp. 211-214, doi: 10.1109/3M-
NANO.2013.6737416.
[6] S. Yeh and J. Sun, “Feedforward motion control design for improving contouring accuracy of CNC machine tools,” in
International Multi Conference of Engineers and Computer Scientists, 2013.

Int J Pow Elec & Dri Syst, Vol. 13, No. 4, December 2022: 2529-2540
Int J Pow Elec & Dri Syst ISSN: 2088-8694  2539

[7] P. Zhao and Y. Shi, “Robust control of the A-axis with friction variation and parameters uncertainty in five-axis CNC machine
tools,” Journal of Mechanical Engineering Science, 2014, doi: 10.1177/0954406213519759.
[8] B. Feng, D. Zhang, J. Yang, and S. Guo, “A novel time-varying friction compensation method for servomechanism,” Hindawi:
Mathematical Problems in Engineering, pp. 1-16, 2015, doi: 10.1155/2015/269391.
[9] F. Wang, C. Liang, Y. Tian, X. Zhao and D. Zhang, “Design and control of a compliant microgripper with a large amplification
ratio for high-speed micro manipulation,” in IEEE/ASME Transactions on Mechatronics, vol. 21, no. 3, pp. 1262-1271, 2016, doi:
10.1109/TMECH.2016.2523564.
[10] C. Liang, F. Wang, Y. Tian, X. Zhao, and D. Zhang, “Development of a high speed and precision wire clamp with both position
and force regulations,” Robotics and Computer-Integrated Manufacturing, vol. 44, pp. 208–217, 2017, doi
10.1016/j.rcim.2016.09.006.
[11] F. Wang, X. Zhao, D. Zhang, Z. Ma, and X. Jing, “Robust and precision control for a directly-driven XY table,” Journal of
Mechanical Engineering Science, vol. 225, pp. 1107–1120, 2011, doi: 10.1177/2041298310392832.
[12] F. Wang, Z. Ma, W. Gao, and X. Zhao, “Dynamic modeling and control of a novel XY positioning stage for semiconductor
packaging,” Transactions of the Institute of Measurement and Control, vol. 37, no. 2, pp. 177–189, 2014, doi:
10.1177/0142331214541598.
[13] E. Yuliza, H. Habil, R. A. Salam, M. M. Munir, and M. Abdullah, Khairurrijal “Development of a simple single-axis motion table
system for testing tilt sensors,” in Engineering Physics International Conference (EPIC), 2017, vol. 170, pp. 378–383, doi:
10.1016/j.proeng.2017.03.061.
[14] P. Perz, I. Malujda, D. Wilczy, and P. Tarkowski, “Methods of controlling a hybrid positioning system using LabVIEW,”
International Polish-Slovak Conference: Machine Modeling and Simulations, 2017, vol. 177, pp. 339–346, doi:
10.1016/j.proeng.2017.02.235.
[15] M. Chang and G. Guo, “Sinusoidal servocompensator implementations with real-time requirements and applications,” in IEEE
Transactions on Control Systems Technology, vol. 25, no. 2, pp. 645-652, 2017, doi: 10.1109/TCST.2016.2558476.
[16] M. Irfan, M. Effendy, N. Alif, S. Lailis, I. Pakaya, and A. Faruq, “Performance comparison of fuzzy logic and proportional-integral
for an electronic load controller,” International Journal of Power Electronics and Drive System (IJPEDS), vol. 8, no. 3, pp. 1176–
1183, 2017, doi: 10.11591/ijpeds.v8i3.pp1176-1183.
[17] M. A. Abdel Ghany, M. A. Shamseldin, and A. M. Abdel Ghany, “A novel fuzzy self tuning technique of single neuron PID
controller for brushless DC motor,” International Journal of Power Electronics and Drive System (IJPEDS), vol. 8, no. 4, pp.
1705-1713, 2017, doi: 10.11591/ijpeds.v8i4.pp1705-1713.
[18] U. K. Bansal and R. Narvey, “Speed control of DC motor using fuzzy PID controller,” Advance in Electronic and Electric
Engineering, vol. 3, no. 9, pp. 1209–1220, 2013.
[19] I. Barkana, “Simple adaptive control-A stable direct model reference adaptive control methodology-brief survey,” International
Journal of Adaptive Control and Signal Processing, vol. 28, no. 7–8, pp. 567–603, 2013, doi: 10.1002/acs.2411.
[20] T.H. Chiew, Z. Jamaludin, A.Y. Bani Hashim, L. Abdullah, N.A. Rafan and M. Maharof “Second order sliding mode control for
direct drive positioning system,” Journal of Mechanical Engineering and Sciences, vol. 11, no. 4, pp. 3206–3216, 2017, doi:
10.15282/jmes.11.4.2017.23.0289.
[21] P. Jain, “Design of a model reference adaptive controller using modified MIT rule for a second order system,” Advance in
Electronic and Electric Engineering, vol. 3, no. 4, pp. 477–484, 2013.
[22] N. Kumar, “Application of fractional order PID controller for AGC,” International Journal of Automation and Computing, vol. 15,
pp. 84–93, 2018, doi: 10.1007/s11633-016-1036-9.
[23] S. Pankaj, J. S. Kumar, and R. K Nema, “Comparative analysis of MIT rule and lyapunov rule in model reference adaptive control
scheme,” Innovative Systems Design and Engineering, vol. 2, no. 4, pp. 154–163, 2011.
[24] S. Y. Hwang, G. Y. Park, D. H. Kim, and K. S. Jhan, “Efficient implementation of a pseudorandom sequence generator for high-
speed data communications,” ETRI Journal, vol. 32, no. 2, pp. 222–229, 2010, doi: 10.4218/etrij.10.1409.0047.
[25] B. J. Moore, T. Berger and D. Song, “Validation of a convolutional neural network model for spike transformation using a
generalized linear model,” Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC),
2020, pp. 3236-3239, doi: 10.1109/EMBC44109.2020.9176458.
[26] G. Sun, X. Ren and D. Li, “Neural active disturbance rejection output control of multimotor servomechanism,” IEEE Transactions
on Control Systems Technology, vol. 23, no. 2, pp. 746-753, March 2015, doi: 10.1109/TCST.2014.2336595.
[27] B. Siegel, “Industrial anomaly detection: a comparison of unsupervised neural network architectures,” IEEE Sensors Letters, vol.
4, no. 8, pp. 1-4, Aug. 2020, Art no. 7501104, doi: 10.1109/LSENS.2020.3007880.
[28] Manuel Baltieri and Christopher L. Buckley, “PID control as a process of active inference with linear generative models,” Entropy,
2019, doi: 10.3390/e21030257.
[29] A. A. El-samahy and M. A. Shamseldin, “Brushless DC motor tracking control using self-tuning fuzzy PID control and model
reference adaptive control,” Ain Shams Engineering Journal, 2016, doi: 10.1016/j.asej.2016.02.004.
[30] M. A. Shamseldin, M. A. Eissa, and A. A. El-samahy, “Practical implementation of GA-based PID controller for brushless DC
motor,” International Middle East Power System Conference (MEPCON’15), 2015.
[31] T. H Chiew, B. Hashim, A. Y, and N. A, Raffan “Identification of friction models for precise positioning system in machine tools,”
Procedia Engineering, vol. 53, pp. 569–578, 2013, doi: 10.1016/j.proeng.2013.02.073.
[32] I. B. Tijani, R. Akmeliawati, A. Legowo, and A. Budiyono, “Nonlinear identification of a small scale unmanned helicopter using
optimized NARX network with multiobjective differential evolution,” Engineering Applications of Artificial Intelligence, vol. 33,
pp. 99–115, 2014, doi: 10.1016/j.engappai.2014.04.003.
[33] E. Cajueiro, R. Kalid and L. Schnitman, “Using NARX model with wavelet network to inferring the polished rod position,”
International Journal of Mathematics and Computers in Simulation, vol. 6, no. 1, pp. 66-73, 2012.
[34] M. R. Stankovi, M. B. Naumović, S. M. Manojlović, and S. T. Mitrović, “Fuzzy model reference adaptive control of velocity servo
system,” FACTA Universitas: Electronics Energetics, vol. 27, pp. 601–611, 2014, doi: 10.2298/FUEE1404601S.
[35] M. A. Shamseldin, A. A. EL-Samahy and A. M. A. Ghany, “Different techniques of self-tuning FOPID control for Brushless DC
Motor,” Eighteenth International Middle East Power Systems Conference (MEPCON), 2016, pp. 342-347, doi:
10.1109/MEPCON.2016.7836913.
[36] Vasadi Srinivasa Rao, P. D. Srinivas and Ch. Ravi Kumar “Load Frequency Control (LFC) of Three Area Interconnected Power
System using Adaptive Neuro Fuzzy Interface System,” International Journal Scientific Engineering and Technology Research,
vol. 4, no. 43, pp. 9482–9488, 2015.
[37] A. L. Elshafei, K. A. El-Metwally, and A. A. Shaltout, “A variable-structure adaptive fuzzy-logic stabilizer for single and multi-
machine power systems,” Control Engineering Practice, vol. 13, no. 4, pp. 413–423, 2005, doi:
10.1016/j.conengprac.2004.03.017.
Design variable structure fuzzy control based on deep neural network model … (Mohamed A. Shamseldin)
2540  ISSN: 2088-8694

[38] P. Swarnkar, S. Jain, and R. K. Nema, “Effect of adaptation gain in model reference adaptive controlled second order system,”
Engineering Technology and Applied Science Research, vol. 1, no. 3, pp. 70–75, 2011, doi: 10.48084/etasr.11.
[39] M. Omar, M. A. Ebrahim, A. M, and F. Bendary, “Tuning of pid controller for load frequency control problem via harmony search
algorithm,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 1, no. 2, pp. 255–263, 2016, doi:
10.11591/ijeecs.v1i2.pp255-263.
[40] M. Omar, A. M. A. Ghany, and F. Bendary, “Harmony search based PID for multi area load frequency control including boiler
dynamics and nonlinearities,” WSEAS Transactions on Circuits and Systems, vol. 14, pp. 407–414, 2015.
[41] M. A. Ebrahim and F. Bendary, “Reduced size harmony search algorithm for optimization,” Journal of Electricarl Engineering,
pp. 1–8, 2016.

BIOGRAPHIES OF AUTHORS

Mohamed A. Shamseldin received M.Sc. degree in system automation and

engineering management and the Ph.D. degree in mechatronics engineering both from the
Helwan University, Cairo, Egypt, in 2016 and 2020, respectively. In 2018, he was a Lecturer
with the Faculty of Engineering, University of Central Lancashire, UK. Currently, he is an
assistant Professor with the faculty of engineering and technology of Mechatronics
Engineering, Future University in Egypt, Cairo, Egypt. His research interests include electric
machine drives, mechatronics, robot control, system automation, and electric vehicle. He can
be contacted at email: [email protected].

Abdel Halim M. Bassiuny was born in Egypt. He is professor at the department

of mechanical engineering, faculty of engineering, Helwan University, Egypt. His research
interests include fuzzy control, neuro-fuzzy, and adaptive control system theories and their
applications, machine tool dynamics and control, condition monitoring and mechatronics.
Dr.Bassiuny was granted a DAAD scholarship, during the interval 1986-1989, at the Group of
Automatic Control and Technical Cybernetics, Wuppertal University, Germany. He can be
contacted at email: [email protected].

Abdel Ghany M. Abdel Ghany was born in Cairo. He received his B.Sc. and
M.Sc. degrees in 1980 and 1987 from the Electrical Power System and Machines department,
Helwan University, Cairo, Egypt. From 1989 to 1994, he got his Ph.D. in Computer
Controlled Systems from the Institute of Control and Systems Engineering, Technical
University of Wroclaw Poland. From 1994 to 1999, he worked as an assistant professor at the
department of Electrical Machines and Power System, Helwan University, Cairo Egypt. In
2002, he was promoted to Professor at the Department of Electrical Power Systems and
Machines Department, University of Helwan, Cairo, Egypt. Dr. Abdel Ghany shared in the
Economical Lighting of Helwan industrial plant as a part of the Supreme Council of Egyptian
Universities Projects. He can be contacted at email: [email protected].

Int J Pow Elec & Dri Syst, Vol. 13, No. 4, December 2022: 2529-2540