Artificial Neural Network-based Control of a Switched Reluctance
Motor for a High-precision Positioning System
IMED MAHMOUD1,3,*, ADEL KHEDHER2,3
1University of Monastir, Higher Institute of Applied Science and Technology of Mahdia, Street
Rejiche 5121 Mahdia, TUNISIA
2University of Sousse, National Engineering School of Sousse, BP 264 Sousse Erriadh 4023,
TUNISIA
3Laboratory of Advanced Technology and Intelligent Systems (LATIS)
Abstract: — This paper presents a approach to control a switched reluctance motor (SRM) in the context of a
high-precision positioning system using artificial neural networks (ANNs). The SRM is known for its
robustness and simplicity, making it suitable for various applications, including positioning systems where
precision is paramount. Traditional control methods often struggle to achieve the desired level of accuracy due
to the non-linear and dynamic nature of the SRM. In this study, we propose an advanced control strategy
leveraging the adaptive learning capabilities of ANNs. The neural network is trained to capture the intricate
relationships between the motor's inputs and outputs, allowing for precise control in real-time. By measuring
the electromagnetic torque and phase currents, the neural network is able to estimate the rotor position,
facilitating the elimination of the rotor position sensor. The training data set of the neural network consists of
magnetization data for the SRM with the electromagnetic torque and current as inputs and the corresponding
position as outputs in this set. With a sufficiently large training data set, the artificial neural networks (ANN)
can be correlated for appropriate network architecture.
Key-words: — Switched Reluctance Motor (SRM), Rotor Position, artificial neural networks, and
electromagnetic torque.
Received: March 15, 2024. Revised: August 19, 2024. Accepted: September 17, 2024. Published: October 18, 2024.
1. Introduction
Switched reluctance motors (SRM) are a type of
stepper motor that are receiving increased attention
for high-precision positioning applications. SRMs
have several advantages over other motors:
• Simple and robust structure with no windings or
permanent magnets on the rotor. This makes them
reliable and suitable for harsh environments.
• Low cost due to the simple construction and
absence of rare earth materials.
• Good torque-to-weight ratio and torque density,
enabling high-precision motion control.
However, SRMs also possess some challenges for
effective control:
• Highly nonlinear torque-speed and torque-current
characteristics. The torque produced depends on the
relative positions of the stator and rotor poles.
• Parameter variations due to temperature changes,
aging effects, and load disturbances. This affects the
motor's performance over time.
Conventional control methods for SRM include
closed-loop PID control and open-loop voltage
control. These methods struggle to adapt to the
motor's nonlinearities and variations.
Artificial neural networks (ANN) can potentially
solve this problem as they can learn the complex
input-output relationships of the SRM through
training. An ANN controller can be developed based
on experimental motor data to produce optimal
control signals for the inverter driving the SRM.
Previous research has demonstrated the benefits of
ANN control for SRMs, including:
• Improved steady-state and dynamic performance
• Higher precision for speed and position control
• Better adaptation to motor nonlinearities and
parameter variations, [1, 2].
In order to get more accurate characteristics than
those given by analytical modelling, numerical
analysis methods are a potentially effective means
and often produce results that are very close to
reality. In particular, in the field of electromagnetic
structures, the use of finite elements methods allows
a precise characterization of electromagnetic devices
using materials with non-linear characteristics and
complex geometry. These finite elements methods
are the basis for powerful electromagnetic
calculation software known as computer-aided
design. Recent work has often combined these
methods with non-conventional modelling
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
136
Volume 6, 2024
techniques. Among these modelling techniques,
artificial neural networks are particularly
noteworthy, as they have shown their power in
modelling non-linear systems, [3, 4]. In addition, the
works published in the recent literature propose
various contributions, mainly focused on improving
the performance of SRM’s used in propulsion mode
and not in positioning mode. Obviously, the strong
distortion of the angular characteristics is a handicap
for the use of this actuator for positioning purposes.
This handicap is all the more pronounced the more
demanding the positioning requirements are in terms
of accuracy. It is in this perspective that our research
work, developed in this paper, is situated. It consists
mainly in proposing control approaches for the use
of this actuator in positioning.
The main contributions are:
1. An ANN-based position estimator that
learns the nonlinear relationship between stator
current and rotor position of the SRM. This ANN
(ANN1) is trained offline using FEA data and
estimates the rotor position based on the supplied
current.
2. An ANN-based controller that generates
the optimal stator current required driving the rotor
to a desired position. This ANN (ANN2) is trained
online during motor operation to minimize position
error.
3. An integrated control approach that
combines ANN1 and ANN2 to achieve high-
precision positioning of the SRM. ANN1 estimates
the current position based on the stator current,
while ANN2 generates the next current command
based on the position error.
This paper is divided into two distinct parts. In
the first section, a finite element study is carried out
to characterize the SRM in order to determine its
electromagnetic properties. This electromagnetic
study is based on the CAD environment "Magnet
2D". The second part is dedicated to the
development of a control approach, using both the
database generated by the finite element method
(FEM) and a cascade of estimators based on
artificial neural networks, in order to correct the
asymmetry of the machine through the control and
to achieve precise positioning of the actuator,
operating at constant load or fluctuating by stages.
2. Electromagnetic characteristics
of the SRM
The SRM performance analysis, both electric
and magnetic, depends on its geometric construction
and materials used. It is almost impossible to
determine exact mathematical equations that take
into account all these influential parameters. In
this way , it is able to give useful results to
calculate the electric machine performance.
Figure 1 presents all the dimensions of the SRM.
Some significant mechanical parameters of the three
topologies are shown in Table 1.
Table. 1. Mechanical and electrical parameters of
the SRM 8/6 considered
Fig. 1- Switched reluctance machine dimensions
The SRM 8/6 exhibits a symmetric and
homogeneous structure and geometry, enabling us
to analyze and obtain the properties of other phases
based on the analysis of a single phase. To
understand the characteristics of the SRM, we can
model the system from aligned and unaligned
positions.
Figure 2 shows the magnetic spectrum of the SRM
at the aligned position where the rotor and stator
poles are perfectly aligned. Figure 2(a) depicts the
flux linkage which is maximum in this position
Parameters
Symbol
Rotor pole angle
𝛽𝑟
Stator pole angle
𝛽s
Stator external diameter
𝐷𝑠
Rotor diameter
𝐷r
Air gap length
g
Stator pole height
Hr
Rotor pole height
Hs
Stator yoke
Ys
Rotor yoke
Yr
Shaft diameter
D0
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
137
Volume 6, 2024
indicating high inductance. Figure 2(b) shows the
high density of Mesh distribution in the overlapped
region of rotor and stator poles. Figure 2(c)
illustrates the magnetic field distribution which is
strongly concentrated in the pole overlap region.
Figure 3(d) displays the induction vector which is
uniformly distributed in this region of maximum
inductance aiding development of torque.
Fig.2- Magnetic spectrum of SRM at aligned
position
(a)- Flux linkage (b)- Mesh distribution
(c)- Magnetic field distribution (d)- Induction vector
Fig.3- Magnetic spectrum of SRM at unaligned
position
(a)- Flux linkage (b)- Mesh distribution
(c)- Magnetic field distribution (d)- Induction vector
Figure 3 shows the magnetic spectrum of SRM at
the unaligned position where rotational
misalignment exists between rotor and stator poles.
Figure 3(a) presents the flux linkage which is now
minimum in this position of low inductance. Figure
3(b) reveals sparse distribution of flux lines between
the poles. Figure 3(c) depicts the magnetic field
spreading wider instead of being focused in the
overlap region. Figure 3(d) shows the induction
vector spreading unevenly outside the region of
highest inductance not contributing to torque. These
spectral representations provide a visual insight into
variation of magnetic quantities affecting operation
from aligned to unaligned position, helping better
understand SRM electromagnetic behavior.
Understanding the SRM requires a detailed
analysis of the torques, flux linkage and inductances
for different rotor positions and different values of
stator excitation currents based on FEA. Therefore,
an FEA simulation tool was used to solve the
magnetic circuit to determine the magnetic fields
and electromagnetic quantities of each machine.
This simulation tool allows us to obtain a data set
that fully characterizes the magnetic and
electromagnetic states of the SRMs. Examples of
this data are shown in Figure 4.
010 20 30 40 50 60
-20
-10
0
10
20
Rotor position [°]
Torque [Nm]
(a)
010 20 30 40 50 60
0
0.002
0.004
0.006
0.008
0.01
0.012
Rotor position [°]
Inductance [H]
(b)
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
138
Volume 6, 2024
010 20 30
0
0.2
0.4
0.6
0.8
1
Current [A]
Flux linkage [Wb]
(c)
Fig.4- Electromagnetic performances of SRM
(a)- Torque profile (b)- Inductance profile (c)-
Flux linkage profile
The evolution of the magnetic flux depends
essentially on the level of saturation of the magnetic
circuit. In fact, for a constant position, such as that
of alignment, it is observed that the more the
excitation current increases, the more the flux
variation is limited, figure 4c. The value of the
inductance can be determined by using the
calculation results of the magnetic flux as a function
of the rotor position at constant excitation levels.
The results shown in Figure 4b show that the
inductance of a stator phase varies inversely with
the excitation current in the vicinity of the alignment
position (30°), while in the vicinity of the opposition
position (0°) the influence of the current on this
inductance is very limited. For a fixed position, it
can be seen that the influence on the inductance
decreases as the saturation level increases. The real
angular characteristics determined by the finite
element method are distorted and far from being
sinusoidal, which shows the inadequacy of the
analytical method based on the simplifying
assumptions adopted. Thus, in Figure 4a we can see
the presence of some oscillations at the levels of
these characteristics.
3. Neural approach to SRM
position control
Knowing the angular characteristics of the
SRM allows it to be used for positioning. The
results obtained show that the angular characteristics
describing the evolution of the torque as a function
of the rotor position of the variable reluctance
machine are clearly affected by distortions.
Therefore, the use of this type of actuator in
positioning applications cannot be envisaged
without the development of powerful control
approaches that allow the appropriate adjustment of
the stator excitations, taking into account both the
load to be positioned and the level of distortion
affecting the characteristics [5-6-7]. In order to
develop this control strategy, and after having
carried out a detailed characterization of the
machine using CAD, we resorted to non-
conventional control techniques, known as
intelligent, to develop a cascade of control blocks
based on artificial neural networks.
Several works prove that multilayer perceptrons
are the most widely used neural networks today, [5-
6-7] they are able to realize nonlinear associations
between input and output. The architecture of this
type of neural network is shown in Figure 5. Each
neuron has an activation function, which can be
sigmoid, bipolar sigmoid, log-sigmoid, etc. The
weights on the connections can be determined by
the back-propagation algorithm during the training
process and then used to calculate the outputs.
Fig.5- Architectural graph of a multilayer network
Error back-propagation in a multilayer network is
supervised learning. The input is presented for
which the output is determined. The set of synaptic
weights determines the operation of the neural
network. The neurons outputs of the output layer are
compared with the model values which are the
desired outputs and the error of each is calculated as
clearly shown in Figure.6. The most commonly used
function that we have adopted in this work is the
squared error function. This function is defined for
each example (n) a number of behavioural examples
(N) as inputs to the network, associated with the
same number (N) of desired outputs as follows:
2
1
1
( ) ( ) ( )
2
K
kk
k
E n d n y n

(1)
For all examples we consider the mean square error
as follows:
1
1()
N
moy
n
E E n
N
(2)
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
139
Volume 6, 2024
Fig.6- Multi-layer network learning with error
determination and back- propagation
The realization of the learning phase is strongly
related to the relevant choice and number of
examples that must be made available to the
network. These examples must be sufficiently
representative of the evolution of these angular
characteristics for the reconstruction to succeed in
completing this important phase [8-9-10-11].
010 20 30
0
20
40
0
20
40
60
80
Current [A]
Rotor position [°]
Torque [Nm]
Fig.7- training database of artificial neural network
(ANN)
For this purpose, we used a numerical
interpolation technique available in the Matlab
environment, which led us to develop a computer
program based on cubic interpolation. The response
surface shown in Figure.7 is a graphical
representation of the database describing the
evolution of the torque as a function of the rotor
position for the whole operating range of the
machine.
3.1. Design and development of the
position estimation network (ANN1)
In order to estimate the stopping position of the
considered SRM when the coupled load and the
stator excitation are known, we preceded to the
creation of a multilayer neural network (ANN1)
using predefined functions in the MATLAB
environment. The inputs of this network are the
torque exerted by the load and the excitation
current, while its output is the angular position of
the rotor. This network consists of a single hidden
layer of 13 neurons and an output layer of a single
neuron. The activation function chosen for the
hidden layer neurons is that of the hyperbolic
tangent of the sigmoid, while for the output layer
neuron the activation is provided by the linear
function. Through repeated learning and the use of
the previously developed database, we have ensured
that this network is capable of estimating the stop
position over the entire working range of the
machine, regardless of the level of stator excitation
and the magnitude of the torque imposed by the
coupled load [12-13-14-15-16].
In order to verify the effectiveness of the
estimation provided by the developed neural
network across the entire operating range of the
considered SRM, we proceeded by conducting a
learning test. This test involved reconstructing
several other examples that were not included in the
database presented to the network during the
training phase and comparing the calculated results
by the network with the expected results. They
display, for an excitation current vector ranging
from 20A to 36A with a step of 0.5A, the evolution
of the electromagnetic torque based on the target
positions and the positions estimated by ANN1,
Figure 8. . For all these examples, the error did not
exceed 0.8%.
15 20 25 30
0
20
40
60
80
100
Rotor position [°]
Torque [Nm]
22.5 23 23.5
38
40
42
44
46
48
Rotor position [°]
Torque [Nm]
Fig.8-effectiveness of ANN1 for unlearned
responses
3.2. Design and development of the
current estimation network (ANN2)
In order to give the rotor a well defined stop
position, for an SRM with a given angular
Desired output
Estimated
output
Desired output
Estimated
output
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
140
Volume 6, 2024
characteristic, it is necessary to modulate the stator
excitation according to the resistance force imposed
by the coupled load. This is the basic idea behind
the design of the second multilayer neural network,
ANN2, whose objective is to determine the
amplitude of the excitation current required to reach
the target position. Therefore, the inputs or
attributes for this network can only be the load
torque and the target position, while the output or
class is none other than the excitation current. The
ANN2 is composed of a single hidden layer
structured around 14 neurons and an output layer
composed of a single neuron. We have chosen the
hyperbolic tangent function of the sigmoid for the
activation of all the neurons of this hidden layer,
while the activation of the neuron of the output layer
is provided by the linear function.
Similarly to the previously designed ANN1
position estimation network and in order to show
that the network has learned the characteristics
presented in the learning base and that the
performance obtained is satisfactory, we have
presented in the same Figure 9, for different stable
positions, the evolution of the torque as a function
of the target currents and the currents calculated by
the ANN2 network. In fact, for several positions
considered by successions of steps of 0.75° and
delimited by the terminals 15° and 30°, the
evolution of the torque is plotted by triangular
patterns as a function of the target intensities and by
star patterns as a function of the intensities
estimated by the ANN2. These characteristics are
determined with a gradual current variation of 1A.
The results obtained show a satisfactory agreement
with the error between the values of the target
intensities and the intensities calculated by the
designed RMC2 network not exceeding 0.16%.
20 25 30 35
0
20
40
60
80
100
Current [A]
Torque [Nm]
Fig.9- Comparison of target currents and
calculated currents by ANN2.
The results presented in Figure 9, show that the
developed ANN2 is able to accurately estimate the
appropriate level of stator excitations, allowing to
give the rotor the target position, taking into account
the coupled load.
3.3. Simulation Validation of the
Proposed Control Approach
In this section, we propose to perform numerical
simulation tests to verify the effectiveness of the
proposed control approach. For this purpose, we
have used the first network ANN1 to simulate the
behavior of the switched reluctance machine
through its angular characteristics and we have
inserted the network ANN2 to calibrate the stator
excitations as a function of both the set position and
the magnitude of the coupled load, Figure 10.
The tests carried out consist of applying a well-
defined resistive torque to the machine each time
and successively varying the position setpoint. The
ANN2 neural network then estimates the amount of
stator current that must be applied to the machine so
that its rotor stops at the target position. To verify
this target position by simulation, we have
represented the machine by the neural network
ANN1, which describes the electromagnetic
behavior of the machine through its angular
characteristics.
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
141
Volume 6, 2024
Fig.10- General overview of the proposed control approach
0 5 10 15 20 25 30
0
20
40
60
80
100
Rotor position [°]
Torque [Nm]
22.5 23 23.5
38
40
42
44
46
48
Position de rotor[°]
Couple[Nm]
Fig.11-Comparison between the reference and the
achieved positions
The result presented in Figure 11 highlight the
effectiveness of the proposed control approach for
using SRMs as positioning actuators and show that
control can provide effective solutions to
significantly mitigate natural machine
imperfections.
4. Conclusion
The choice of SRM is based on its many
advantages, such as excellent performance in
extreme environments, simple rotor structure,
robustness, no coils, no permanent magnets, no
brushes, high overload capability, low
manufacturing, repair and maintenance costs, and
operation in a wide power range. The problem
discussed is how to overcome the constraint of
torque ripple for its best use as an electric vehicle
drive motor in underground mines to replace highly
polluting diesel vehicles.
In this paper, we have proposed a control
approach for the operation of the SRM as a
positioning actuator. This approach is based on
artificial intelligence control techniques and in
particular on artificial neural networks. The results
obtained show the potential power of the proposed
control approach for the exploitation of variable
reluctance machines in the positioning domain.
Overall, the presented research demonstrates the
potential of artificial neural networks in enhancing
the control of switched reluctance motors for high-
precision applications, opening new avenues for
advancements in positioning system technology.
References
[1] T. Zeineb, M. Imed, A. Rui Esteves, K. Adel,”
Fuzzy Super-Twisting Sliding Mode Controller
for Switched Reluctance Wind Power Generator
in Low-Voltage DC-microgrid Applications”,
Energies 2024, 17, 1416.
[2] E. Sanches and J. Santisteban, “Comparative
Study of Conventional, Fuzzy Logic and Neural
PID Speed Controllers with Torque Ripple
Minimization for an Axial Magnetic Flux
Switched Reluctance Motor”, Engineering,
2014, 6, 655-669
[3] S. Vikas Wadnerkar, G. Tulasi Ram Das and
A.D. Rajkumar, “Performance Analysis of SRM
Drive Using ANN Based Controlling of 6/4
Switched Reluctance Motor”, Journal of
Theoretical and Applied Information
Technology, pp. 334-338, 2009.
[4] R.A. Gupta and S.K Bishnoi,., “Sensorless
Control of Switched Reluctance Motor Drive
with Fuzzy Logic Based Rotor Position
Estimation”, International Journal of Computer
Applications, Vol. 1, No. 22, 2010.
[5] H. Ibrahim and Al-Bahadly, “Examination of a
Sensorless Rotor-Position Measurement Method
for Switched Reluctance Drive,IEEE
Transactions on Industrial Electronics,Vol. 55,
No. 1, pp. 288-295, 2008.
[6] A. M. El-Naggar, A. A. Zaki and M. A. El-
Kholy, "Artificial Neural Network-Based
Control for Speed Control of Switched
Reluctance Motor", International Journal of
Electrical, Computer, Energetic, Electronic
and Communication Engineering, Volume 14,
Number 2, 2020.
ANN2
ANN1
LW[1,1]
b [2]
LW[2,1]
θ ref [°]
Cem [Nm]
I es
2
14
1
b [1]
LW[1,1]
b [2]
LW[2,1]
2
13
1
b [1]
θ es [°]
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
142
Volume 6, 2024
[7] H. S. Kim, H. Y. Jung, J. H. Kim and H. J. Kim,
"Position Control of Switched Reluctance Motor
Using Neural Network-Based SMC for High-
Precision Applications", Energies, Volume 13,
Number 21, pp5719, 2020.
[8] A. B. Khairi, M. R. M. A. Rahman, M. N. M.
Nasir, M. Z. A. A. Kadir and M. S. A. Rahman,
"Switched Reluctance Motor Control Using an
Artificial Neural Network-Based Controller for
Positioning Applications", Journal of Advanced
Research in Dynamical and Control Systems,
Volume 12, Number 7, 2020.
[9] S. Khorrami-Sina, M. Eghtesad and M.
Parniani, "Position Control of Switched
Reluctance Motor Using an Artificial Neural
Network-Based Controller", IEEE Transactions
on Industrial Electronics ,Volume, 68, Number
4, 2021.
[10] M.R.A.Ghani, N.Farah, and M. R. Tamjis.
"Vector control of switched reluctance motor
using fuzzy logic and artificial neutral network
controllers", International Conference on
Electrical, Electronics, and Optimization
Techniques (ICEEOT), pp.4412-4417, (2016).
[11] [11]. K. B. Naouel. "Contribution à la
modélisation par la méthode des réseaux de
réluctances (MRR) d’une machine à réluctance
variable", Thèse de Doctorat, U.S.T.O Oran
2015.
[12] A. F. Reza, M. R. Amin, M. Ferdows and
M. M. Rahman, "Position Control of Switched
Reluctance Motor Using Neural Network-Based
Control Technique", International Journal of
Electrical and Computer Engineering (IJECE),
Volume 12, Number 6, pp5505-5517, 2022.
[13] D. Zhou, Y. Li, M. Yan and Q. Wang,
"Hybrid Neural Network-Based Control of
Switched Reluctance Motor for High-Precision
Positioning System", IEEE Transactions on
Industrial Informatics, Volume 17, Number 2,
pp1312-1323, 2021.
[14] P.V.Rajesh, M.Balamurugan, N.Ramaiah.
"Artificial neural network based on direct torque
control of four phases switched reluctance
motor", SSRN Electronic Journal, 2019.
[15] K. S. Kumar, K. R. Sreejaya, M. Usha and
B. S. Anand, "Speed Control of Switched
Reluctance Motor using Artificial Neural
Network Controller", International Journal of
Power Electronics and Drive Systems, Volume
11, Number 3, 2020.
[16] F. Zhang, Y. Zhang, Y. Li and Y. Zhao,
"Position Control of Switched Reluctance Motor
Based on Adaptive Neural Network", IEEE
Access, Volume 8, pp 144098-144108, 2020.
Engineering World
DOI:10.37394/232025.2024.6.14
Imed Mahmoud, Adel Khedher
E-ISSN: 2692-5079
143
Volume 6, 2024
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The authors equally contributed in the present
research, at all stages from the formulation of the
problem to the final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflicts of interest to declare
that are relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US