15th International Conference on DEVELOPMENT AND APPLICATION SYSTEMS, Suceava, Romania, May 21-23, 2020

Sensorless Control of PMSM using FOC Strategy

Based on Multiple ANN and Load Torque Observer

Marcel NICOLA1 Claudiu-Ionel NICOLA1,2, Marian DU1

1 2
Research Department Department of Automatic Control and Electronics
National Institute for Research, Development and Testing in University of Craiova
Electrical Engineering – ICMET Craiova, Romania
Craiova, Romania [email protected], [email protected]
[email protected]

Abstract—This article presents the sensorless control system complexity relative to the stochastic ones, but they have lower
of a Permanent Magnet Synchronous Motor (PMSM), where the accuracy as compared to them.
speed controller consists of a Mutliple - Artificial Neural
Networks (M-ANN) and the rotor speed is provided by a Sliding A particular role in the modern control systems is given by
Mode Observer (SMO). The performance of the proposed control their implementation using the ANN [15-17]. Among the main
system is presented in comparison with the classic Field Oriented advantages of the ANN we can mention the robustness and the
Control (FOC) system. The implementation of a load torque implementation of very complex systems by simply knowing
observer allows the selection of a specially trained corresponding the evolution in time of their input and output. This article
ANN to obtain optimum performance over the load torque presents the sensorless control system of a PMSM using a M-
variation ranges. The PMSM equations, the speed and load ANN for the speed control, a load torque estimator and a SMO-
torque observers equations, the main blocks and control type rotor speed and position estimator. The load torque can be
structures, their parameterizations and the results of the considered from the point of view of the PMSM automatic
numerical simulations obtained are presented. The good results control system as disturbance parameter. Thus, a key element
obtained as a result of the numerical simulations involving the in the overall control structure proposed in this article is the
use of the usual Simulink and Stateflow blocks, which can be presence of a load torque estimator. By knowing the range of
implemented in embedded systems recommend the real-time the usual load torque variation, a number of ANNs will be
implementation of the proposed PMSM control system. trained on significant intervals, depending on the application,
to reproduce the control achieved with FOC-type structure with
Keywords—State observers; Artificial neural networks; Motor
drives; Permanenet magnet motors; Automatic control PI-type controllers tuned for each of these sub-intervals of the
load torque variation. By selecting the corresponding ANN,
according to the estimated value of the load torque, superior
I. INTRODUCTION performances will be achieved compared to the use of a FOC
It is undeniable that there has been a growing interest in the structure by tuning the control parameters for the entire range
PMSM lately. This is explained by the fact that the PMSM has of the load torque variation. The rest of the paper is organized
a number of advantages over other types of motors: low inertia, as follows: Section II describes the mathematical model of the
high load torque density, reduced dimensions, but also faster PMSM and the FOC-type control structure. The mathematical
cooling, considering that the losses are concentrated in the models and the structures of implementation of the speed and
stator and the harmonic content is lower [1-4]. load torque estimators in the Matlab/Simulink environment are
presented in Section III. Section IV presents the ANN
Naturally, a series of PMSM control systems have been structures which selectively perform the PMSM control, based
developed. Thus, we can mention the classic FOC and Direct on the estimated load torque, and the results of the numerical
Torque Control (DTC) control structures, but also modern simulations. The final section summarizes the conclusions and
adaptive, robust and predictive type control structures [5-8]. ideas for the continuation of the topics addressed in this article.
The intelligent PMSM control can be achieved by using fuzzy,
ANN, genetic or Particle Swarm Optimization (PSO) type
control algorithms [9, 10]. In order to eliminate the II. PMSM AND FOC STRATEGY - MATHEMATICAL MODEL
disadvantages of using position and speed transducers, a By using the usual notations in the rotor reference frame
number of estimators were achieved, among which we can and by following [3, 17], the mathematical model of the
mention: estimators based on the SMO method, Model PMSM in d-q frame can be written as follows:
Reference Adaptive System (MRAS) and Luenberger observer
for the deterministic approach and Kalman-type estimators for
a stochastic type approach [11-14]. It is obvious that the ª u q º ª R q + ρ Lq ω e L d º ª iq º ªω e λ 0 º
(1)
«u » = « − ω L +
deterministic estimators have the advantage of reduced
¬ d¼ ¬ e q R d + ρ L d »¼ «¬ i d »¼ «¬ ρλ 0 »¼
This paper was developed with funds from the Ministry of Research and
Innovation as part of the NUCLEU Program: PN 19 38 01 03.

978-1-7281-6870-8/20/$31.00 ©2020 IEEE

32
Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:46:51 UTC from IEEE Xplore. Restrictions apply.
 where: Rq, Rd and Lq, Ld are the stator resistances and III. ROTOR SPEED AND LOAD TORQUE ESTIMATION
inductances in the d-q reference frame of the PMSM, e is the The speed estimator equations are obtained based on the
electrical angular velocity of the rotor, 0 represents the flux PMSM model given by equations (1)-(4), and by using the
linkage,  is the differential operator, while id, iq and ud, uq are transformation matrix in (5), the equations in the - frame
the stator voltages and currents in the d-q reference frame. (using Park inverse transformation) are obtained.
The flux on the d-q axes are given by:

§ cos θ e − sin θ e ·
λ q = Lq iq T = ¨¨ ¸ (5)
(2) © sin θ e cos θ e ¸¹
λ d = Ld id + λ 0
Based on these, the currents i, i and the back-EMF e and
By indicating the electromagnetic torque developed by the e are given by the relations (6) and (7) respectively:
PMSM as Te the following relations regarding the PMSM
dynamics can be expressed:
iα = id cos(θ e ) − iq sin(θ e )
(6)
iβ = id sin(θ e ) + iq cos(θ e )
Te = n p {λ d i q − λ q id }
3
2 (3)
dω
Te = T L + B ω + J
dt dλα
eα = = −λ0ωe sin (θ e )
where: np is the number of pole pairs, B is the viscous
dt (7)
dλ β
friction coefficient, J is the rotor inertia and TL is the load eβ = = −λ0ωe cos(θ e )
torque. dt
If we consider the simplifications Ld=Lq=L, Rd=Rq=Rs, and In the - frame, the PMSM equations can be synthetized
e=np· where  is the angular velocity of the rotor, then the as in the following relation:
PMSM model becomes:

diα R 1 1
§ Rs · § ud · = − s iα − eα + uα
¨ − n pω 0 ¸ ¨ ¸ dt L L L
§ id · ¨ L ¸ § id · ¨ L ¸ (8)
¨ ¸ ¨ n λ ¨ ¸ u (4) diβ
− p 0 ¸ ¨ iq ¸ + ¨ q ¸
R R 1 1
¨ iq ¸ = ¨ − n p ω − s
¸ ¨ ¸ = − s i β − eβ + u β
¨ ω ¸ ¨ L L ¨ ¸ L dt L L L
© ¹ Kt B ¸© ω ¹ ¨ TL ¸
¨ 0 − ¸ ¨− ¸
© J J ¹ © J ¹ By estimating the back-EMF e and e, by using a SMO,
the position and the speed of the rotor can then be estimated.
where: Kt is the torque constant. The equations of the observer are given in the following
The classic FOC-type control structure of the PMSM is relation [17]:
shown in Fig. 1 [3]. The FOC-type control structure shown in
Fig. 1 contains a cascade control system where the outer loop is
diˆα R 1 1
intended for the control the rotor speed of the PMSM motor. = − s iˆα + uα − kH (iˆα − iα )
The current reference id is set to zero for torque maximization dt L L L (9)
reasons, and the reference iq for the inner current control loop diˆβ R 1 1
is prescribed by the PI-type speed controller. The SVPWM = − s iˆβ + u β − kH (iˆβ − iβ )
(Space Vector Pulse Width Modulation) block ensures the dt L L L
transition of the IGBTs (power transistors) according to the where: k is the observer gain, and function H is of the
control inputs u and u and a switching table [3]. sigmoid type:

2
H ( x − y) = −a ( x − y )
−1 (10)
1+ e
where: a is a positive constant, and the sigmoid function
defined by (10) will take values between -1 and 1 for a=4 and
will replace the sign function of the classic approach of the
SMO, resulting in the reduction of the chattering phenomenon.
Fig. 1. Block diagram of the FOC strategy for the PMSM The sliding vector is defined as:

33
Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:46:51 UTC from IEEE Xplore. Restrictions apply.
 S n = [ sα sβ ]T = [iˆα − iα iˆβ − iβ ]T = [iα iβ ]T (11) The implementation of the speed observer in
Matlab/Simulink is shown in Fig. 2. The equations (9)
describing the observer are implemented as the subsystem on
and Lyapunov function is chosen as follows [17]: the left side of the block diagram in Fig 2. The sigmoid
function is used, instead of the sign function, to smooth the
inherent oscillations in a SMO. In the final part of the SMO
1 1
V = S nT Sn = sα2 + sβ2
2 2
( ) (12) implementation, an intermediate filter block of the back-EMF
e and e is also added as low-pass filters with a cut-off
frequency of 500Hz.
Parameter k is chosen such that V < 0 , ensuring the SMO
stability when the system trajectory is on the sliding surface,
Sn=0. The system of equations of the errors of currents is
defined as follows:


iα = iˆα − iα = − s iα + eα − kH (iα )
R 1 1
L L L (13)

iβ = iˆβ − iβ = − s iβ + eβ − kH (iβ )
R 1 1
L L L

By calculating V , the following relation is obtained [17]:

Fig. 2. Block diagram of rotor speed and position SMO implementation in
Matlab/Simulink
R
L
( 1
L
) [ ]
V = − s iα2 + iβ2 + (eα − k )iα H (iα ) + (eβ − k )iβ H (iβ ) < 0 (14)
Fig. 3 presents the results of the numerical simulations of
the evolution in time of quantities provided by the SMO: rotor
And to fulfill the condition V < 0 , the observer gain is speed and position, back-EMF e and e and further id and iq
chosen as follows: currents. A very good estimation of the parameters provided by
the SMO is noticed. For constructive reasons, the PMSM has a
special start, so that up to 100ms the operation is done in the
(
k ≥ max eα , eβ ) (15) open loop, after a predefined current sequence. For this reason,
the rotor speed is not estimated during this time period. The
nominal parameters of the PMSM are given in Table I.
The sliding mode is obtained on the surface:

[ sα sβ ]T = [ sα sβ ]T ≈ [0 0] (16)

Thus, from the relations (14)-(16), the estimates for the

back-EMF e and e are obtained in the following form:

eˆα = kH (iα ) = −λ0ωˆ e sin θ e

(17)
eˆβ = kH (iβ ) = λ0ωˆ e cosθ e

The rotor speed and position estimates can be obtained

from the relation (17), in the form of the following relations: Fig. 3. Simulation of the back-EMF, rotor position and rotor speed time
evolution from the SMO

The relations (3) are used for the estimation of the load
eˆα2 + eˆβ2 torque, and the implementation in Matlab/Simulink is
ωˆ e = (18)
λ0 presented in the block diagram in Fig. 4. On the left side of the
block diagram there is a subsystem which implements the
relations (3), and on the right side there is a block of
adjustment of the output to get a good estimate of the load
t
torque. Fig. 5 shows the results of the numerical simulations of
θˆe (t ) = ³ ωˆ e (t )dt + θ 0 (19) the evolution in time of the load torque estimate for a
t0 prescribed sequence of the following form: [0.1 0.25 0.5 0.75 1
1.25]s [1 2 3 4 3 3 2]Nm (to which a uniformly distributed
where: 0 is the initial electrical position of the rotor. noise with an amplitude of 0.4Nm is added).

34
Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:46:51 UTC from IEEE Xplore. Restrictions apply.
 2 z −1 (21)
s=
Ts z + 1

The sampling period Ts used in the Matlab/Simulink

numerical simulations is of 2 s. In Equations (20) and (21), the
continuous and discrete variables are denoted as s and z
respectively.

Fig. 4. Block diagram of the load torque observer implementation in

Matlab/Simulink

Fig. 5. Time evolution simulation of the estimated load torque

Fig. 6. Sensorless control of the PMSM based on M-ANN using rotor speed
SMO and load torque observer
IV. SENSORLESS CONTROL OF THE PMSM USING
MULTIPLE ANN – NUMERICAL SIMULATION
The PMSM control system proposed in this article replaces
the PI-type speed controller in the FOC-type structure
presented in Section II, with a control system based on a series
of ANNs (see Fig. 6), selected according to the load torque
value. In the electric drives, the load torque is an external
signal, considered as a disturbance from the point of view of
the speed control system, so that it is imperative to use a load
torque observer TL. Also, since the control system is sensorless,
a rotor speed observer will be used as presented in Section III.
Depending on the application, the load torque variation range is
divided into sub-intervals, and on each of these sub-intervals,
the speed controller is optimally tuned and the speed error Fig. 7. Simulink subsystem for the selection of the ANN
values (the PI-type speed controller input) and iq_ref (the PI-type
speed controller output) are saved as two vectors. After this The block diagram of an ANN structure is presented in Fig.
stage, each ANN is trained with these vectors. Depending on 8. The Deep Learning toolbox was used for the implementation
the load torque value (provided by the observer), the specially of the ANN in Matlab. The selected ANN is of the Neural
trained ANN is selected, obtaining superior performance over a Fitting type with two Layers (Hidden Layer with 150 neurons
classic FOC structure, where the PI-type speed controller and Output Layer with 1 neuron). For the training of each
performs the speed control over the entire range of the speed ANN, the sample length of the input/output vectors (the
reference and load torque. input/output of the PI-type speed controller) is of 125224
The numerical simulations were performed in the samples. Of these, 87565 samples were used for training,
Matlab/Simulink environment. The implementation of the 18784 samples for validation and 18784 samples for testing.
ANN selection according to the load torque is presented in Fig. The Levenberg-Marquardt method was chosen for the training
7. The implementation is based on the Stateflow toolbox in process of each ANN.
Matlab.
The PI type controller has the transfer function:

1 (20)
H PI ( s) = K p + K i
s Fig. 8. Block diagram of the ANN structure

For discretization, the Tustin substitution of the following The performance of the training of an ANN corresponding
form is used: to a sub-interval of the load torque is presented in Fig. 9 and

35
Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:46:51 UTC from IEEE Xplore. Restrictions apply.
10. The value close to 1 of the regression parameter R in Fig. 9 11. If the load torque has a dynamic given by the sequence: [0
(which represents a measure of the approximation between 0.05 0.1 0.25 0.5 0.75 1 1.25]s [1 3 5 4 4 2 2 3]Nm, to which
output and target) is a guarantee that the ANN has been a uniformly distributed noise of the type described above is
successfully trained. added, the best tuning of the PI controller provides the results
obtained in Fig. 12. It can be observed that the profile cannot
be tracked with good results in the 0-100ms range when the
load torque has a fast variation, but also shortly after the 750ms
time, when the load torque has a relatively high value at a
speed of 1200rpm. By selecting five equal sub-intervals of the
load torque variation, 5 ANNs were trained under the
conditions in which the optimal tunings of the PI-type speed
controller were achieved for each of these sub-intervals. In this
way, it can be seen in Fig. 13 that higher performance was
obtained compared to a PI controller tuned for the entire load
torque variation range. The value of the speed ripple is defined
as follows [5]:

¦ (ω (i) − ωref (i))

1 N (22)
ωrip = 2

N i =1

where: N represents the number of samples,  and ref

represent the rotor speed and respectively, the prescribed
reference of speed.
Fig. 9. Regression R – after the training of the ANN
By calculating the speed ripple for the system response in
Fig. 12, a value of 16.273rpm is obtained. Using the M-ANN
based controller for the same speed and load torque profile
defined above, a reduction of the speed ripple by
approximately 25% is obtained.
Also, the speed profile is tracked with good fidelity by the
M-ANN based controller in the areas specified above where
Fig. 10. Training state of the ANN - performance the classic controller provided modest results. Regarding the
parametric robustness of the M-ANN-based controller in Fig.
For the numerical simulation in the Matlab/Simulink 14, the results of the numerical simulation for the speed and
environment, a PMSM with the nominal parameters given in load torque profiles described above are presented, under the
Table I was selected. conditions where the J parameter (combined inertia of rotor
and load) has a 50% increase (from 0.8e-3 kg·m2 to 1.2e-3
kg·m2). A good response of the M-ANN-based control is also
TABLE I. PMSM - NOMINAL PARAMETERS
noticed in the case of the parametric variation.
Parameter Value Unit
Stator resistance - Rs 2.875
q and d inductance - Lq and Ld 0.0085 H
Combined inertia of rotor and load - J 0.8e-3 kg·m2
Combined viscous friction of rotor and
0.005 N·m·s/rad
load - B
Flux induced by the permanent magnets
0.175 Wb
of the rotor in the stator phases - 0
Pole pairs number - P 4 -

In Fig. 11-14, the speed profile required to compare the

performance of the M-ANN-based control systems to the FOC-
type control system is described by the following sequence: [0
0.25 0.5 0.75 1 1.25]s [300 600 900 1200 900 600 ]rpm. For
PMSM used in the numerical simulations, the nominal load
torque is of 3Nm. For a load torque profile with a relatively
low dynamic described by the sequence: [0 0.25 0.5 0.75
1.25]s [1 2 3 3 2]Nm (to which a uniformly distributed noise
with an amplitude of 0.4Nm is added), after an optimal tuning
of the PI-type speed controller (using the Ziegler-Nichols Fig. 11. Simulation of the PMSM time evolution with FOC strategy and
tuning method), a good response is obtained like the one in Fig. low/medium variation of the load torque

36
Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:46:51 UTC from IEEE Xplore. Restrictions apply.
 REFERENCES

[1] V. Utkin, J. Guldner, J. Shi, Sliding mode control in electromechanical

systems, second edition. Automation and Control Engineering, Taylor &
Francis, 2009.
[2] Y. Zhao, “Position/speed sensorless control for permanent-magnet
synchronous machines,” Ph.D. dissertation, Electrical Engineering,
Univ. Nebraska – Lincoln, Lincoln, 2014.
[3] B. K. Bose, Modern power electronics and AC drives, Prentice Hall,
Knoxville, Tennessee, USA, 2002.
[4] S. Hussain and M. A. Bazaz, “Comparative analysis of speed control
strategies for vector controlled PMSM drive,” International Conference
on Computing, Communication and Automation (ICCCA), Noida, India,
2016, pp. 1314-1319.
[5] B. Ning, S. Cheng, and Y. Qin, “Direct torque control of PMSM using
Fig. 12. Simulation of the PMSM time evolution with FOC strategy and high sliding mode backstepping control with extended state observer,” in
variation of the load torque Journal of Vibration and Control , vol. 24, no.4, pp. 694-707, Feb. 2018.
[6] P. M. Nicolae, M. C. Nitu and D. Constantin, “Direct flux vector control
for dual-three phase induction motor drives,” 4th International Youth
Conference on Energy (IYCE), Siofok, Hungary, 2013, pp. 1-5.
[7] C. Nicola, M. Nicola, S. Popescu, and M. Du , “Power Factor
Correction and Sensorless Control of PMSM Using FOC Strategy,”
International Conference on Electromechanical and Energy Systems
(SIELMEN), Craiova, Romania, 2019, pp. 1-6.
[8] M. Nicola and C. Nicola, “Sensorless Predictive Control for PMSM
Using MRAS Observer,” International Conference on
Electromechanical and Energy Systems (SIELMEN), Craiova, Romania,
2019, pp. 1-7.
[9] M. Mutluer and O. Bilgin, “Design optimization of PMSM by particle
swarm optimization and genetic algorithm,” 12 International Symposium
on Innovations in Intelligent Systems and Applications, Trabzon,
Turkey, 2012, pp. 1-4.
Fig. 13. Simulation of the PMSM time evolution with M-ANN and high [10] W. A. A. Salem, G. F. Osman, and S. H. Arfa, “Adaptive Neuro-Fuzzy
variation of the load torque Inference System Based Field Oriented Control of PMSM & Speed
Estimation,” Twentieth International Middle East Power Systems
Conference (MEPCON), Cairo, Egypt, 2018, pp. 626-631.
[11] K. Kyslan, V. Šlapák, V. Fedák, F. urovský, and K. Horváth, “Design
of load torque and mechanical speed estimator of PMSM with unscented
Kalman filter - An engineering guide,” 19th International Conference on
Electrical Drives and Power Electronics (EDPE), Dubrovnik, Croatia,
2017, pp. 297-302.
[12] N. Henwood, J. Malaizé and L. Praly, “A robust nonlinear Luenberger
Fig. 14. Simulation of the PMSM time evolution with M-ANN and high observer for the sensorless control of SM-PMSM: Rotor position and
variation of the load torque and 50% increase of J parameter magnets flux estimation,” 38th Annual Conference on IEEE Industrial
Electronics Society (IECON), Montreal, Canada 2012, pp. 1625-1630.
[13] A. Aissa, K. Ameur, and B. Mokhtari, “MRAS for Speed Sensorless
V. CONCLUSIONS Direct Torque Control of a PMSM Drive Based on PI Fuzzy Logic and
This article presents the sensorless control system of a Stator Resistance Estimator,” in Transaction on Control and Mechanical
PMSM, where the speed controller consists of a M-ANN, the Systems, vol. 2, no. 7, pp. 321-326, Jul. 2013.
rotor speed is provided by a SMO-type observer, and better [14] B. Liu, “Speed control for permanent magnet synchronous motor based
on an improved extended state observer,” in Advances in Mechanical
performances was obtained compared to the classic FOC-type Engineering, vol. 10, no.1, pp. 1-12, Nov. 2018.
control system based on PI-type speed controller. A key [15] S. K. Suman, M. K. Gautam, R. Srivastava and V. K. Giri, “Novel
element is the implementation of a load torque observer, which approach of speed control of PMSM drive using neural network
allows the selection of a specially trained corresponding ANN controller,” International Conference on Electrical, Electronics, and
for optimum performance over the load torque variation Optimization Techniques (ICEEOT), Chennai, China, 2016, pp. 2780-
ranges. The PSMS equations, the speed and load torque 2783.
observers equations, the main blocks and control structures, [16] J. Lee and J. Ha, “On-Line Switch-Open Fault Detection of PMSM
Using Artificial Neural Network,” 10th International Conference on
and the results of the numerical simulations obtained are also Power Electronics and ECCE Asia (ICPE 2019 - ECCE Asia), Busan,
presented. Following the good results obtained from the Korea (South), 2019, pp. 1-6.
numerical simulations and the use of the usual Simulink and [17] M. S. Wang and T. M. Tsai, “Sliding Mode and Neural Network Control
Stateflow blocks (which can be implemented in embedded of Sensorless PMSM Controlled System for Power Consumption and
systems), future approaches will present the real-time Performance Improvement,” in Energies, vol. 10, no. 11, pp. 1-15, Nov.
implementations of the control system of the proposed PMSM 2017.
using hardware dedicated platforms with DSP controllers.

37
Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:46:51 UTC from IEEE Xplore. Restrictions apply.