IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO.

10, OCTOBER 2021 12045

Neural Network Observers and Sensorless Robust

Optimal Control for Partially Unknown PMSM With
Disturbances and Saturating Voltages
Luy Nguyen Tan , Thanh Pham Cong , and Duy Pham Cong

Abstract—This article proposes neural network (NN) based ob- control, the position sensors (optical encoders, or resolvers) are
server schemes and a sensorless robust optimal control scheme commonly used [1], [2]. Due to the drawbacks of using the
for partially unknown permanent magnet synchronous motors sensors such as costs, sizes, and reliabilities of the PMSM drive
with disturbances and saturating voltages. First, an NN-observer
scheme is designed to estimate back-electromotive force (EMF), systems, sensorless control has, therefore, been developed in
for which the mathematical model in rotary or stationary refer- recent years [3], [4], [6]–[11].
ence frames is relaxed. The NN weight tuning law is designed Compared to the model reference [11], extended Kalman fil-
via Lyapunov theory to guarantee that EMF is ultimately uni-
tering observers [7] based on high-frequency injection methods,
formly bounded. Second, to compensate the inexact extraction of
the estimated back-EMF at any speed conditions, disturbances, Luenberger linear observers [6] that give lower convergence
and NN approximation errors, another NN-observer scheme is rates compared with nonlinear observers and are sensitive to
designed to estimate the tracking errors of rotor position and parameter variation, sliding-mode observers (SMOs) [8] have
speed, for which low-pass filters and/or phase-locked loops are some advantages thanks to their robustness with model errors.
not needed. Third, a sensorless saturated robust optimal control
The EKF provides accuracy and flexibility of the estimate and
scheme dealing with general disturbances and saturating voltages
is designed. The scheme includes the augmented feedforward con- its relatively low complexity. In addition, the EKF can be inte-
troller to transform the speed and current tracking problem into grated with SMO or PLL to enhance the robustness of nonlinear
an optimal control problem. Finally, the feedback control law and systems. To compensate phase delay due to low-pass filters in
worst disturbance law are obtained without estimating unknown SMO, an extended SMO version is proposed [9]. Recently, Xu et
internal dynamics. The effectiveness of the proposed schemes is
al. [10] have adopted SMO for a current adaptive observer with
tested through simulations and comparative experiments on a load
drive application with a DSP board TMS320F28379D. estimation of back-electromotive force (back-EMF), and ex-
ploited the back-EMF error to design another adaptive observer
Index Terms—Neural network (NN) observers, robust optimal for the rotor speed and position tracking. Although a continuous
control, saturating voltages, sensorless control.
sigmoid function is employed instead of a signum function to
reduce the chattering phenomenon in the estimated signals, it
I. INTRODUCTION cannot completely overcome the chattering at low-speed ranges.
ERMANENT magnet synchronous motor (PMSM) is one SMO based on phase-locked loops (PLL) [12] is employed to
P of the most powerful actuators often applied to industrial
rotary systems such as electric vehicles and wind generation
enhance the position and speed extractions from the observed
back-EMF values [13], [14]. Qu et al. [13] propose the sensorless
systems. For over decades, researchers have been motivated by control scheme by using two linear extended state observers,
control system design due to its not only control performances which can reject both external and uncertain parameter distur-
enhancement but also cost reduction [1]–[5]. For the feedback bances. The phase delay and chattering reduction are solved
by exploiting PLL. Lin and Zhang [14] improved SMO and
Manuscript received November 17, 2020; revised February 11, 2021; accepted designed a PLL structure with tangent function to guarantee
April 3, 2021. Date of publication April 6, 2021; date of current version June the robust estimation performance in both forward and reverse
30, 2021. This work was supported by the Vietnam National Foundation for Sci- rotation of a rotor. This approach is not only simple but also
ence and Technology Development (NAFOSTED) under Grant 107.04-2019.25.
Recommended for publication by Associate Editor K.-B. Lee. (Corresponding effective. For an optimization approach, Toso et al. [4] develop a
author: Luy Nguyen Tan.) linear observer and control scheme based on the moving horizon
Luy Nguyen Tan is with the Faculty of Electronics Technology, Industrial estimation method. Meanwhile, Sun et al. [3] propose a nonlin-
University of Ho Chi Minh City, Ho Chi Minh 700000, Vietnam (e-mail:
[email protected]). ear observer scheme. The introduced optimization-schemes can
Thanh Pham Cong is with the VietNam Aviation Academy, Ho Chi Minh minimize cost functions, and therefore, are proven to be more
700000, Vietnam (e-mail: [email protected]). more effective than those using extended Kalman filters or SMO.
Duy Pham Cong is with the Faculty of Electrical Technology, Industrial
University of Ho Chi Minh City, Ho Chi Minh 700000, Vietnam (e-mail: The performances of the phase delay and noise reduction rely
[email protected]). on PLL.
Color versions of one or more figures in this article are available at https:
//doi.org/10.1109/TPEL.2021.3071465.
The drawback of the estimation methods mentioned above
Digital Object Identifier 10.1109/TPEL.2021.3071465 is that they completely depend on the known current dynamics

0885-8993 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 12046 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 10, OCTOBER 2021

in rotary or stationary reference frames, and thus, cannot be To the best of authors’ knowledge, observer schemes and
applied to sensorless control with unknown dynamics, including sensorless robust optimal control schemes have not been tackled
unknown stator resistance, inductance, and a set of differential for partially unknown PMSM with disturbances and saturating
equations. Recently, the back-EMF error has been estimated voltages. In this article, we design such a scheme with three key
to reduce the chattering [10], but this method still requires the advantages. First, compared to the most recent studies on sensor-
known values of the stator resistance and inductance. In [15], a less control [3], [4], [7]–[11], [13], which require known current
flux SMO is designed to deal with the drift of the stator resistance dynamics without disturbance, the proposed NN-observers do
and inductance under increasing temperature when a motor runs not require such dynamics and, therefore, can generally be
in real time. In addition, an embedded flux-observer is designed applied to PMSMs with completely unknown current dynamics.
to replace the conventional PLL. However, the nominal values Second, the NN-observers are implemented online in sync with
of the stator resistance and inductance must be priorly known. a controller without requiring any sample dataset for training
Note that, for most practical processes, having an exact model is offline. Third, a robust optimal control method is developed
a difficult or even impossible task. To overcome that, the PMSM for real-time sensorless PMSM with input constraint to further
model is considered as a black box and a system identification improve tracking errors and energy consumption in practice. The
technique based on NN [16]. However, the identification proce- following main contribution of this article includes not only an
dure requires the offline implementation and the sample set. observer approach but also a control approach.
The offline identification is not appropriate in most of the 1) An NN-observer scheme is designed to estimate the back-
industrial processes, thus, the online observer and optimal EMF, in which the known current dynamics in rotary or
control algorithms are investigated for nonlinear systems with stationary reference frames are relaxed. The NN weight
known structures based on fuzzy [17] or completely unknown tuning law for the observer is designed via Lyapunov
structures based on NN [18]. Because of universal approx- theory to guarantee the back-EMF is ultimately uniformly
imation, the NN-observers do not require a priori knowl- bound (UUB).
edge of system dynamics. However, direct application of the 2) Since the extraction of the estimated back-EMF is not
NN-observer to PMSM dynamics faces two main challenges. exact at any speed conditions under disturbances and
The former relates to back-EMF extraction for rotor posi- NN approximation errors, another NN-observer scheme is
tion estimation being inexact at any speed conditions under designed to estimate the tracking errors of rotor position
disturbances of uncertain parameters and NN approximation. and speed. The tracking errors are proven UUB and can
The latter deals with measured current data for the observers be adjusted significantly small by choosing the appropriate
with large error caused by external disturbances and inverter update rate.
nonlinearity. 3) Via IRL, the sensorless saturated robust optimal tracking
It is worth emphasizing that the performance of observers in control (SSROC) scheme dealing with general distur-
closed-loop sensorless systems is also affected by the perfor- bances and saturating voltages is introduced.
mance of controllers. Optimal control is preferred over stability The scheme includes the augmented feedforward controller to
for PMSM, but by using conventional PI controllers with the transform the speed and current tracking problem into an optimal
cascade structures, the optimal regulation techniques cannot be control problem based on back-stepping techniques [23]–[27].
utilized to improve the control performance. Although the works The optimal control law and worst disturbance law are obtained
in [3] and [4] approach optimization, the control schemes still without estimating drift parameters in system dynamics such as
use PI control structures. stator resistance, viscous friction, and torque load. The proposed
Recently, the quadratic optimal control using state-dependent- NN-observer and control schemes are applied to an experimen-
Riccati-equation (SDRE) for PMSM are proposed in [1] and tal application with a DSP board TMS320F28379D of Texas
[2]. As the Riccati solution of SDRE is solved in real time to Instruments.
stabilize the states around equilibrium steady-state points, the The rest of this article is organized as follows. Section II
high computing speed is required; otherwise, the control signals designs the NN-observer schemes. Then, the sensorless robust
are delayed, leading to instability of the closed loop. In addition, optimal control scheme is proposed in Section III, the simula-
as full-state measurements are required, SDRE is not appropriate tions and experiments are given in Section IV, and the conclusion
for sensorless optimal control. is briefly drawn in Section V.
In industrial applications, the control performance of PMSM
is constrained by the saturation inputs due to the voltage limita-
tion of a power converter. Aghili [1] proposes the PMSM optimal II. NEURAL NETWORK-OBSERVER DESIGN FOR SENSORLESS
control with upper bounds of voltages in the dq stationary CONTROL
reference frame. Most recently, inspired by intelligent learning A. Back-EMF NN-Observer Scheme
systems, saturated robust adaptive optimal control based on a
Consider PMSM in the stationary reference frame α–β with
powerful technique of integral reinforcement learning (IRL) [19]
completely unknown dynamics
has been widely developed for simulated nonlinear systems [20],
consensus robots with disturbances [21], and saturated-voltage
power systems [22]. With IRL, the full knowledge of system i̇α,β = Fp (iα,β , uα,β , eα,β ) + dL
(1)
dynamics is not required. yα,β = Cp iα,β

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 TAN et al.: NEURAL NETWORK OBSERVERS AND SENSORLESS ROBUST OPTIMAL CONTROL 12047

where iα,β = [iα , iβ ] , uα,β = [uα , uβ ] , eα,β = [eα , eβ ] ,

iα , iβ are the measured phase currents, uα , uβ are phase volt-
ages, and eα , eβ are back-EMFs, Fp is an unknown continuous
nonlinear smooth function, yα,β = [yα , yβ ] , Cp ∈ I 2×2 is the
identity matrix, dL is disturbance due to parameter uncertainty,
which is unknown but bounded by a positive constant bL .
By adding and subtracting Ap iα,β , dynamics (1) can be
rewritten as
i̇α,β = Ap iα,β + Fp1 (iα,β , uα,β , eα,β ) + dL
(2)
yα,β = Cp iα,β
where Ap is a Hurwitz matrix, the pair (Ap , Cp ) is observable, Fig. 1. Structure diagram of back-EMF NN-observer scheme.
Fp1 (iα,β , uα,β , eα,β ) = Fp (iα,β , uα,β , eα,β ) − Ap iα,β .
The back-EMFs eα and eβ are estimated by are obtained by subtracting (8) from (6)
êα = K1 (yα − ŷα )
êβ = K2 (yβ − ŷβ )
(3) ĩ˙ α,β = (Ap − Kp Cp )ĩα,β + W̃p φp (ξ) + ζ(t) (9)
ỹα,β = Cp ĩα,β
where K1 and K2 are observer gains, ŷα and ŷβ are outputs of
the observer. The observer model for (2) is given by where W̃p = Wp − Ŵp , and ζ(t) = Wp (φp (ξ) − φp (ξ)) ˆ +
εp (ξ) + dL . It can be seen that ζ(t) ≤ 2bW p bφp + bεp +
î˙ α,β = Ap îα,β + Fp1 (îα,β , uα,β ) + Kp (yα,β − Cp îα,β ) + dL bL = bζ , where bζ is a positive constant.
ŷα,β = Cp îα,β
To guarantee Ŵp → Wp , the NN weight-tuning law has to be
(4) designed. Furthermore, the stability of the NN-observer needs
where îα,β = [îα , îβ ] , ŷα,β = [ŷα , ŷβ ] , Kp = diag[K1 , K2 ]. to be ensured in the online update process.
Due to the unknown function Fp1 (.), êα , êβ cannot be ex- We propose a structure diagram of the NN-observer scheme
tracted, but they can be approximated by NN. in Fig. 1 then design and analysis the stability of an online
According to the Weierstrass higher order approximation NN parameter update law for the scheme through the following
theorem [28], [29], Fp1 (iα,β , uα,β ) can be represented as theorem.
Theorem 1: Consider the unknown PMSM model (1) and
Fp1 (ξ) = Wp φp (Vp ξ) + εp (ξ) (5) the NN-observer model (8). Let the NN weight-tuning law be
where ξ = [i   2×nh
, and Vp ∈ Rnh ×4 are designed as
α,β , uα,β ] , Wp ∈ R
the ideal weight matrices, nh is the number of hidden layer  ˆ
˙ −λp Kp−1 ê  −1
α,β Cp Pp Ŵp + Pp Cp Kp êα,β φp (ξ)
neurons. φp (.) is the NN activation function vector includ- Ŵp =
ing hyperbolic tangent functions, and εp is the NN functional Cp 2
approximation error vector. According to NN approximation (10)
property [29], Wp , φp (ξ), and εp (ξ) are bounded such that where λp > 0 is the update rate and Pp is a positive definite
Wp  ≤ bW p , φp (ξ) ≤ bφp , εp (ξ) ≤ bεp for some positive solution of the following Lyapunov equation for some positive
constants bW p , bφp , bεp . definite matrices Qp :
Then, the dynamics of (1) are represented by NN as
(Ap − Kp Cp ) Pp + Pp (Ap − Kp Cp ) = −Qp . (11)
i̇α,β = Ap iα,β + Wp φp (Vp ξ) + εp (ξ) + dL
(6)
yα,β = Cp iα,β . Then, the NN-observer is stable, i.e., the current estimation error
Note that for the simplicity, Vp can be chosen as a unit matrix ĩα,β , the estimated back-EMF êα,β , and the NN approximation
without being tuned, i.e., Vp ∈ I. error W̃p are UUB.
Since the ideal NN weight matrices are unavailable, F1 (ξ) is Proof: Select the Lyapunov function candidate
approximated by 1  1
ˆ = Ŵp φp (ξ)
ˆ L= ĩα,β Pp ĩα,β + Trace{W̃p W̃p }. (12)
F̂p1 (ξ) (7) 2 2
where ξˆ = [î  
α,β , uα,β ] . The NN-observer dynamics of (4) are
Differentiating (12) and using (9), one obtains
given by 1   
L̇ = ĩ (Ap − Kp Cp ) Pp + Pp (Ap − Kp Cp ) ĩα,β
î˙ α,β = Ap îα,β + Ŵp φp (ξ)
ˆ + Kp (Cp iα,β − Cp îα,β ) + dL 2 α,β
ŷα,β = Cp îα,β . ˙  W̃ }. (13)
+ ĩ P W̃ φ (ξ) + ĩ P ζ + Trace{W̃
α,β p p p α,β p p p
(8)
From (1) to (4), one obtains
Denote the state and output estimation errors ĩα,β = iα,β −
îα,β and ỹα,β = yα,β − ŷα,β , respectively. The error dynamics êα,β = Kp Cp ĩα,β . (14)

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 12048 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 10, OCTOBER 2021

˙
˙ = −Ŵ function and its derivative derived from (3), (4), and (6), i.e.,
Applying the tuning law (10) and (14) to (13) with W̃ p p
⎛  ⎞
yields  K1 C1 iα − îα + εiα
êα
1    θ̂e = − arctan = − arctan ⎝  ⎠
L̇ = ĩα,β (Ap − Kp Cp ) Pp + Pp (Ap − Kp Cp ) ĩα,β êβ K C i − î + ε
2 2 2 β β iβ
  d
 ω̂e = θ̂
C C p
+ ĩ  p
α,β Pp ζ + ĩα,β I − Pp W̃p φp (ξ) dt
Cp 2 (21)

ĩα,β Cp Cp Pp  where εiα and εiβ are residual errors. From (21), it can be seen
+ λp Trace{W̃p Ŵp }. (15) that the extraction is not exact at any speed conditions under
Cp 2
the residual errors. To provide the high-performance sensorless
As W̃p Ŵp = W̃p (Wp − W̃p ), one has Trace{W̃p Ŵp } = position and speed for the tracking control scheme in Section III,
Trace{W̃p Wp } −Trace{W̃p W̃p } < W̃p Wp − W̃p 2 . Then, a tracking error observer based on NN is constructed in this
(15) is rewritten as section.
Denote the position tracking error and the speed tracking error
1   
L̇ ≤ ĩα,β (Ap − Kp Cp ) Pp + Pp (Ap − Kp Cp ) ĩα,β as
2
eθ = θd − θ̂e
+ bζ ĩα,β Pp  + bφp ĩα,β Pp W̃p  (22)
eω = ωd − ω̂e
 
+ λp ĩα,β Pp  W̃p Wp  − W̃p 2 . (16) where ω̂ is estimated speed extracted by the estimated back-
EMF, θd is the reference position and ωd is the reference speed.
From (11), we have
It is desirable that the errors equal go to zero.
1 Remark 1: A low-pass filter can be employed for (21) to reject
L̇ ≤ − λmin (Qp )ĩα,β 2 + bζ ĩα,β Pp 
2 disturbance without accessing the system parameters, but it often
causes the phase delay. On the other hand, PLL can be designed
+ bφp ĩα,β Pp W̃p 
  to fast respond to the change of rotor information and simply
+ λp ĩα,β Pp  W̃p Wp  − W̃p 2 (17) implemented [14], [31], but the performance depends on the
gains of a PI regulator and the error of the back-EMF observer via
where λmin (Qp ) is the smallest eigenvalue of Qp . the formerly known current dynamics in the stationary reference
Using the fact that 2ab ≤ a2 + b2 and −a2 + 2ab = −(a − frame α–β. On the other hand, the gains of the PI regulator are
b) + b2 , one has
2 favored for design based on the nominal load torque and the
 moment of inertia [31].
2
1 2 λp Pp  bφp Remark 2: The position error (22) is the tracking error be-
L̇ ≤ − λmin (Qp )ĩα,β  − W̃p − ĩα,β 
2 2 λp tween the estimated position extracted from the back-EMF
 NN-observer, with the completely unknown current dynamics
λ 1 2
+ bζ + b2W p + b Pp ĩα,β . (18) in the stationary reference frame α–β, and the reference trajec-
2 2λp φp tory, and thus, is different from the expression of the error in
Denote μ = bζ + 2λ b2W p + 1 2
≥ 0, the negative semidef- PLL, which contains the error between the actual and estimated
2λp bφp , μ
positions [14], [31].
initeness of L̇ is guaranteed if either of the following conditions As the back-EMF êα and êβ contain disturbances and NN
holds: approximation errors, a model of the tracking error dynamics is
2μPp  introduced by completely unknown dynamics as
ĩα,β  ≥ = bĩαβ , êα,β  ≥ Cp Kp−1 bĩαβ (19)
λmin (Qp )
ėθ,ω = Fe (eθ , eω ) (23)
or
2μ + bφp where eθ,ω = [eθ , eω ] , Fe is the smooth unknown function.
W̃p  ≥ = bW̃ p . (20) By adding and subtracting Ae eθ,ω , the dynamics (23) can be
λp
rewritten as
It can be concluded that if the estimation error ĩα,β or êα,β
ėθ,ω = Ae eθ,ω + Fe1 (eθ , eω ) (24)
exceeds a compact set proportional to bĩαβ or the approximation
error W̃p exceeds bW̃ p , then L̇ < 0. As a result, the errors are where Ae is a Hurwitz matrix, and Fe1 (eθ , eω ) = Fe (eθ , eω ) −
attracted back into the sets. This satisfies UUB condition [30]. Ae eθ,ω .
The observer dynamics is given by
B. Tracking Error NN-Observer Scheme
ê˙ θ,ω = Ae êθ,ω + Fe1 (êθ,ω ) + Ke (eθ,ω − êθ,ω ) (25)
The estimated rotor position and speed θ̂ and ω̂ can be directly
extracted from the estimated back-EMF [8] using an arctangent where êθ,ω = [êθ , êω ] , Ke is the observer gain matrix.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 TAN et al.: NEURAL NETWORK OBSERVERS AND SENSORLESS ROBUST OPTIMAL CONTROL 12049

A. PMSM Dynamics in Reference Frame D–Q

The PMSM mathematical model in reference frame d–q is
given by
B 3np TL 1
ω̇ = − ω+ ψ f iq − + dτ
J 2J J J
Rs 1
i̇d = − id + np ωiq + u d + dd (30)
Ld Ld
Rs np ψf 1
i̇q = − iq − np ωid − ω+ u q + dq
Lq Lq Lq
where in the dq-axis, id and iq are stator currents, ud and uq
are stator voltages denoted as control inputs, which are bounded
by ū due to a voltage limit of a power converter, i.e., |ud | ≤ ū
and |uq | ≤ ū, Rs , Ld , Lq are the stator resistance and induc-
tance, without loss of generality, Ld = Lq = Ls [32], J, B, and
Fig. 2. Structure diagram of tracking error NN-observer scheme.
TL are the moment of inertia, the viscous friction coefficient,
and the load torque, respectively, dτ is external disturbance,
Applying NN to (25), the NN-observer dynamics is given by dd and dq are disturbances caused by uncertainty such that
dL = diag[dd , dq ].
ê˙ θ,ω = Ae êθ,ω + Ŵe φe (eθ,ω ) + Ke (eθ,ω − êθ,ω ). (26) The dynamics with estimation states is given by
If êθ and êω are estimated, the new observed position and B 3np TL 1
ω̂˙ = − ω̂ + ψf îq − + dτ
speed of the rotor are obtained as J 2J J J
1
î˙ d = − îd + np ω̂ îq +
Rs
u d + dd (31)
θ̂ = θd − êθ Ld Ld
(27) 1
î˙ q = − îq − np ω̂ îd −
ω̂ = ωd − êω . Rs np ψf
ω̂ + u q + dq
Lq Lq Lq
We propose a structure diagram of the NN-observer scheme
in Fig. 2 and now design an online parameter update law for the where the estimation currents îd and îq are derived from the
observer. NN-observers and Park transformation as follows:
Theorem 2: Consider the tracking error dynamics (23) and the îd = iα cos θ̂ + iβ sin θ̂
(32)
observer model (25). Let the NN weight-tuning law be designed îq = iβ cos θ̂ − iα sin θ̂.
as
It can seen that dd ≤ bd , dq ≤ bq , and dd , dq ∈ L2 [0, ∞), where
˙
Ŵe = −λe ẽ 
θ,ω Pe Ŵe + Pe ẽθ,ω φe (êθ , êω ) (28) bd and bq are positive constants.
To facilitate the control design, the boundedness and assump-
where λe > 0 is the update rate, ẽθ,ω = eθ,ω − êθ,ω , and Pe is tion are introduced.
a positive definite solution of the following Lyapunov equation Boundedness 1: Denote
for some positive definite matrix Qe : ⎡ 1 ⎤
  0
(Ae − Ke ) Pe + Pe (Ae − Ke ) = −Qe . 3np ψf ⎢ ⎥
(29) g1 (ω̂) = 0 , g2 (ω̂, îd , îq ) = ⎣ Ld 1 ⎦ .
2J 0
Then, the NN-observer is stable, i.e., the position and speed Lq
estimation tracking error ẽθ,ω and NN approximation error W̃e According to the physical nature, there exist g1 (ω̂) ≤ g1 max ,
are UUB. g2 (ω̂, îd , îq ) ≤ g2 max , where Jmax , g1 max , and g2 max are
Proof: This proof is similar to the proof of Theorem 1 and unknown positive constants.
omitted herein. Assumption 1: The reference speed ωd is smooth and
Remark 3: The tracking errors can be reduced as small as bounded.
possible by decreasing λe to reduce bounds of the sets. However, Remark 5: Boundedness 1 and Assumption 1 are practical.
if λe is small the convergence rate may be slow.
Remark 4: Using the tuning-law (28) of the tracking error B. Augmented Feedforward Control
NN-observer with the completely unknown dynamics, identify-
ing the nominal parameters of PMSM is not needed. Inspired by [33], this section designs a virtual feedforward
control scheme to convert the tracking control problem of (31)
into a sensorless saturated robust optimal control problem. The
III. SENSORLESS SATURATED ROBUST OPTIMAL CONTROL
design exploits a back-stepping procedure.
SCHEME
Step 1: From (27) and (30), the speed tracking error dynamics
In this section, by employing the results of the NN-observers, are written as
a sensorless saturated robust optimal control scheme for PMSM B TL 3np ψf 1
is designed. ê˙ ω = −ω̇d − ω̂ − + îq + dτ
J J 2J J

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 12050 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 10, OCTOBER 2021

B TL 3np ψf d 3np ψf  
Fc = [f1 (ω̂), f2 (ω̂, îd , îq ), f3 (ω̂, îd , îq )] are unknown internal
= −ω̇d − ω̂ − + iq + îq − idq
J J 2J 2J dynamics, where
dτ f1 (ω̂) = −BJ ω̂ − J
TL
+
J f2 (ω̂, îd , îq ) = − Ld îd + np ω̂ îq
Rs
(41)
B TL 3np ψf d 3np ψf 1 f2 (ω̂, îd , îq ) = −np ω̂ îd − R
np ψf
= −ω̇d − ω̂ − + iq + êiq + dτ Lq îq − ω̂.
s
Lq
J J 2J 2J J
(33) Then, the tracking control problem for the PMSM subject to
(31) is equivalent to the optimal control problem for system
where êid = îd − idd , êiq = îq − idq . The virtual control inputs
(40) with unknown knowledge of drift dynamics Fc including
idd , idq are defined as the load torque.
idd = id Proof: Choose a Lyapunov function candidate for tracking
(34) dynamics (31)
idq = iq + iaq
1 2 1 1
where id and iq are the optimal control inputs designed in the J1 = êω + ê2id + ê2iq . (42)
following section, and iaq is the augmented feedforward control 2 2 2
input chosen as The differentiation of J1 with respect to time is given by
2J J˙1 = eω ê˙ ω + êid ê˙ id + êiq ê˙ iq . (43)
iaq = ω̇d . (35)
3np ψf
From the dynamics (36) and (39) yield
Using (33) and (34), the speed tracking error dynamics (32) are 
B TL 3np ψf 3np ψf 1
rewritten as J˙1 = êω − ω̂ − + iq + êiq + dτ
J J 2J 2J J
B TL 3np ψf 3np ψf 1 
ê˙ ω = − ω̂ − + iq + êiq + dτ . (36) Rs 1
J J 2J 2J J + êid − îd + np ω̂ îq + u + dd
Step 2: The actual control inputs ud and uq are designed as Ld Ld d

ud = ud + uad Rs np ψf 3np ψf
(37) + êiq −np ω̂ îd − îq − ω̂ − êω
uq = uq + uaq Lq Lq 2J
where ud and uq are the optimal control inputs designed in the 1
+ u + dq . (44)
following section, and uad and uaq are the augmented feedfoward Ld q
control inputs chosen as Equation (44) can be rewritten as
uad = 0 J˙1 = êT (Fc + Gu∗ + Kd) . (45)
3Lq np ψf (38)
=
uaq −
Lq i̇dq êω .
2J On the other hand, the Lyapunov function candidate for the
Note that in design of the strict feedback nonlinear systems using closed-loop dynamics (40) is chosen as
back-stepping [33], i̇dq is derivative of the virtual signal idd . The 1 
current tracking error dynamics (31) are rewritten as J2 = ê ê. (46)
2
Rs 1 The derivative of J2 gives
ê˙ id = − îd + np ω̂ îq + u + dd
Ld Ld d
3np ψf J˙2 = êT (Fc + Gu∗ + Kd) . (47)
˙êi = −np ω̂ îd − Rs îq − np ψf
ω̂ − êω (39)
It can be seen that if u∗ makes (40) stable, by Lyapunov theory,
q
Lq Lq 2J
1 J˙2 must be negative. Comparing (47) with (45), J˙2 < 0 results
+ u q + dq .
Lq in J˙1 < 0, which implies that the tracking errors in (36) and (39)
The equivalent conversion from the tracking control for the are UUB. In other words, the tracking control problem for the
sensorless PMSM with the NN-observers to the optimal tracking sensorless PMSM defined by (31) is converted to the SSROC
control will be shown in the following theorem. for (40).
Theorem 3: Consider PMSM dynamics with states estimated
by NN-observers (31), let the reference speed satisfy Assump- C. Sensorless Saturated Robust Optimal Control Law
tion 1, the virtual and actual control inputs be defined in (34) According to Theorem 4, the control requirement will be
and (37), where the augmented feedforward control inputs are met if law u∗ of the SSROC scheme is solved. This section
designed in (35) and (38), let the sensorless saturated robust provides an equation Hamilton–Jacobi–Issac (HJI) that supports
optimal control input vector u∗ stabilize the closed-loop system the solution gradient for the law. Since the analytic solution of
ê˙ = Fc + Gu∗ + Kd (40) HJI is not available, NN is used to approximate it.
The structure of NN-based SSROC for the sensorless PMSM
where ê = [êω , êid , êiq ] , u∗ = [i∗d , i∗q , u∗d , u∗q ] , G = diag is proposed in Fig. 3. To design the control inputs u∗d and u∗q in
[g1 (ω̂), g2 (ω̂, îd , îq )], K = diag[1/J, 1, 1], d = [dτ , dd , dq ] , the presence of unknown internal dynamics, voltage constraints,

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 TAN et al.: NEURAL NETWORK OBSERVERS AND SENSORLESS ROBUST OPTIMAL CONTROL 12051

  
∂Vc∗ (ê)
+ Fc + Gu∗ + Kd∗ = 0. (54)
∂ê
The HJI equation is high-order nonlinear differential and con-
tains the unknown function Fc . It, therefore, does not have
analytic solution. In the rest of the section, NN is going to be
employed to approximate the solution.
By the Weierstrass higher order approximation theorem [28],
[29], Vc∗ (ê) is approximated by
Fig. 3. Sensorless saturated robust optimal control structure for PMSM with Vc∗ (ê) = Wc φc (ê) + εc (ê) (55)
partially unknown dynamics, saturation voltages, and disturbances.
where φ(e) : R3 → Rn is the vector of n smooth activation
and disturbances, the H∞ tracking performance value function functions, εc (e) is approximation error, Wc ∈ Rn is the ideal
is defined as NN weight vector. As Wc is unknown, Vc∗ (ê) is approximated
 ∞ by
Vc (ê, u, d) = r(ê(τ ), u(τ ), d(τ ))dτ (48) V̂c (ê) = Ŵc φc (ê) (56)
t

where r(ê, u, d) = ê Qc ê + U (u) − γ 2 d d, Qc is the positive- where Ŵc = [Ŵc1 , Ŵc2 , . . ., Ŵcn ] is the approximation pa-
definite matrix, u is an approximation of u∗ at each given rameter vector.
time. γ > 0 is the prescribed disturbance attenuation level. If Assumption 2: φc (ê) ≤ bφc , ∇φc (ê) =
the closed-loop dynamics (7) are stable with a minimum gain ∂φc (ê)/∂ê ≤ b∇φc , ∇εc (ê) = ∂εc (ê)/∂ê ≤ b∇εc ,
γ ∗ > 0, it remains stable with γ > γ ∗ [34]. εc (ê) ≤ bεc , where bφc , b∇φc , bεc , b∇εc are positive constants.
Inspired by [28] and [35] for the constrained inputs, the Utilizing V̂c (ê) for the law (52) and the worst-case disturbance
function U (u) is chosen as law (53), we obtain that
 
 u 1 −1  ∂ V̂ (ê)
U (u) = 2ū tanh (s/ū)Rds (49) û = −ū tanh(N̂ ), N̂ = R G (57)
0
2ū ∂ê
 
where R is a positive definite diagonal matrix. Note that 1 ∂ V̂ (ê)
| tanh(.)| ≤ 1 and since tanh (s/ū) is the monotonic odd ˆ
d = − 2K 
. (58)
2γ ∂ê
function, U (u) is nonnegative.
For optimal control, the Hamiltonian is defined as The HJI equation (54) is then approximated by
  
∂Vc (ê) Ĥ = Ŵc ∇φc (ê) Fc + Gû + K dˆ + ê Qc ê + U (û)
H ê, u, d, = r(ê, u, d)
∂ê
 
− γ 2 dˆ d.
ˆ (59)
∂Vc (ê)
+ (Fc + Gu + Kd) . 1
∂ê According to IRL [19], the residual error EĤ = ε ε needs
(50) 2 Ĥ Ĥ
to be minimized, where
Then, the optimal value function satisfies the Nash solution from    
the zero-sum game theory [36] as εĤ = Ŵc ∇φc (ê) Fc + Gû + K dˆ dτ
t−ta
 ∞    
Vc∗ (ê(t)) = min max r(ê(t), u(t), d(t))dt. (51) + ê Qc ê + U (û) − γ 2 dˆ dˆ dτ (60)
u d 0
t−ta
Applying the stationary condition for (50), i.e., ∂H(.)/∂u = 0 where ta > 0 is a given time interval.
and ∂H(.)/∂d = 0, the saturated optimal control law u∗ and the The integral of the first term of (60) is computed as
worst-case disturbance law d∗ are obtained as      

1 −1  ∂Vc∗ (ê) ∇φc (ê) Fc + Gû + K dˆ dτ = ˙
∇φc (ê)êdτ
u∗ = −ū tanh(N ∗ ), N ∗ = R G (52) t−ta t−ta
2ū ∂ê
 = φc (ê(t)) − φc (ê(t − ta )) = Δφc (ê(t)). (61)
∗ 1  ∂Vc∗ (ê)
d = − 2K . (53) To relax the persistent excitation condition [30] and make
2γ ∂ê
the parameter convergence faster, the concurrent learning tech-
Then, substituting (52) and (53) into (50), the HJI equation is nique [37] is employed, i.e., the following squared function of
given by P residual error vectors from the time tl in the past is minimized

∂V ∗ (ê) 1 P 1 P 
H ê, u∗ , d∗ , c
∗
= ê Qê + U (u∗ ) − γ 2 d∗T d∗ Ep = EĤ = e (tl )eĤ (tl ) (62)
∂ê 2 l=1 2 l=1 Ĥ

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 12052 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 10, OCTOBER 2021

where TABLE I
NOMINAL PARAMETERS OF PMSM
 tl
eĤ (tl ) = Ŵc Δφc (ê(tl )) + r̂(τ )dτ (63)
tl −ta

where ˆ l ).
r̂(tl ) = ê (tl )Qc ê(tl ) + U (û(tl )) − γ 2 dˆ (tl )d(t
Δφc (ê(tl )) and r̂(tl ) are recorded and stored in sets
{Δφc (ê(tl ))}Pl=1 , {r̂(tl )}l=1 . To guarantee the execution
P

of the concurrent learning technique, the following condition

should be satisfied.
Condition 1 (see [37]): At least P past stored-data vectors
in the set {Δφc (ê(tl ))}P l=1 must be linearly independent, i.e.,
rank[Δφc (ê(t1 )), Δφc (ê(t2 )), . . ., Δφc (ê(tP ))] = P .
˙ ∂E
Based on the gradient descent rule, Ŵc = −αc1 ∂ ŴĤ −
 ∂EĤ (tl )
c
The parameters of the rotor position NN-observer scheme are
αc2 P l=1 ∂ Ŵc , where αci > 0, i = 1, 2 are the update rates,
set up as
the online parameter tuning law for the approximator is designed      
as 300 0 1 0 1 0
Ap = − , Kp = , Qp = .
0 300 0 1 0 1

˙ (65)
Ŵc = −αc1 Δφc (ê) Δφ
c (ê)Ŵc
Then, the solution of Lyapunov equation (11) is
     
+ ê Qc ê + U (û) − γ 2 dˆ dˆ dτ 0.0017 0
Pp = . (66)
t−ta 0 0.0017

P   tl The parameters of the tracking error NN-observer scheme are
− αc2 Δφc (ê(tl )) Δφ
c (ec (tl ))Ŵc + r̂(τ )dτ . set up as
l=1 tl −ta    
0.01 0 −2 1 0.1
(64) Ae = − , Ke = 10 ,
0 0.01 0.1 0.01
 
1 0
D. Stability and Convergence Analysis Qe = (67)
0 1
In this section, stability and convergence of the proposed
Then, the solution of Lyapunov equation (29) is
control scheme are analyzed.  
Theorem 4: Consider the observed speed and current track- 25.12 −2.48
Pe = . (68)
ing error dynamics (31) with unknown uncertain dynamics, −2.48 49.75
saturating voltages, and general disturbances, let Assumptions
For each NN, the initial weights W{p,e} (0) ∈ R5 are set to
1 and 2, Boundedness 1, Condition 1 be guaranteed, let the
be random within [0, 1], activation functions φ{p,e} = 2/(1 +
sensorless saturated robust optimal control law and the worst-
case disturbance law be approximated by NNs (57) and (58), e−0.1ξ̂{p,e} − 1), the update rates λ{p,e} = 0.01.
where the NN parameters are tuned online by (64). Then, For the control scheme, the matrices of the value func-
the tracking errors and the approximation error are UUB, tion (48) are chosen as Qc ∈ I 3×3 , Rc ∈ I 4×4 . NN weights
the parameters of the value function, the control and the Ŵc ∈ R6 are set to be zeros. The activation functions are
disturbance laws synchronously converge to the near-optimal chosen as φc (ê) = [ê2ω , êω êid , êω êiq , ê2id , êid êiq , ê2iq ] , λc1 =
values. 0.1, λc2 = 0.001. The size of the past data array P = 6, ta = Ts .
Proof: The proof follows from [38] and is omitted herein due The disturbance attenuation level γ = 10. The saturation ampli-
to space limitations. tudes of the control inputs ū = 0.75.

A. Simulation Results
IV. SIMULATION AND EXPERIMENT
The evolution of the NN weights with SSROC is shown in
In this section, the effectiveness of the NN-observer scheme Fig. 4. It can be seen that the weights quickly converge to the
and the SSROC scheme is validated by simulations and experi- approximated optimal values.
ments. Fig. 5 shows the observed and actual rotor speeds, where
In the simulation and experiment, the PMSM model ESTUN the observed speed ω̂ tracks the reference speed ωd with the
EMJ-04APB22 including the parameters in Table I is used. near-zero error during the control process while the actual speed
The reference speed ωd is made smooth by a first-low-pass slightly fluctuates at the early stage but tracks the reference speed
filter 0.001z/(z − 0.999). The sampling period Ts = 100 μs. with the small error when the NN weights Wc converge.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 TAN et al.: NEURAL NETWORK OBSERVERS AND SENSORLESS ROBUST OPTIMAL CONTROL 12053

Fig. 7. Observed, actual, and reference rotor speeds when Rs is changed.

Fig. 4. Evolution of NN weights for SSROC scheme.

Fig. 8. Observed, actual, and reference rotor speeds when B is changed.

Fig. 5. Observed, actual, and reference rotor speeds.

Fig. 9. Observed, actual, and reference rotor speeds when TL is changed.

B. Comparative Experiment
Fig. 6. Rotor positions from NN-observers and actual measurement.
In this section, PMSM is experimented in both clockwise and
counterclockwise. The tracking control performances in dynam-
ical environment between two control schemes using different
The observed and actual mechanical rotor positions from observers are compared. The former includes the observer of the
Fig. 6 show that the phase delay of the rotor position computed estimated position and speed extracted from the conventional
by the output of the back-EMF NN-observer is large while the SMO with PLL (SMO-PLL) governed by PI controllers [31].
phase delay from the output of the tracking error NN-observer The latter includes the back-EMF-direct observer (EMF-DO) [8]
is small. and tracking error NN-observer defined in (27), served as a
The robust ability with external and internal disturbances back-EMF-compensated observer (EMF-CO), regulated by the
of the NN-observers and the SSROC scheme without using proposed SSROC in this article. Note that the former observer
the knowledge of TL , B, and Rs is verified under changes of requires the PMSM parameters but the later observers do not.
the stator resistance, the viscous friction, and the load torque. The overall performances are also determined by control laws of
Figs. 7 –9 show the performance of the observed and actual the conventional PI controller and the robust optimal controller
rotor speeds when Rs is added to 100% of the nominal value, SSROC.
the viscous friction B/J increases by 5000% of the nominal Since SSROC learns online, the poor control performance at
value, and the load torque is added TL =1.0 N · m at 1 s. It can the early stage may cause hardware damage. Therefore, after the
be seen that the performance of the observed speed is almost approximate NN weights have been obtained from the simula-
unchanged, and the tracking performance of the actual speed is tion, we continue to apply them for SSROC in the real system,
robust with the disturbances. and thus, the weights are further refined.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 12054 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 10, OCTOBER 2021

Fig. 10. Experimental prototype of PMSM.

Fig. 11. Estimated back-EMF of NN-observer using for tracking error NN-
The experiment is conducted on a PMSM prototype model observer.
of an application of load drive equipped a brake. The platform,
presented in Fig. 10, includes a PMSM ESTUN EMJ-04APB22,
of which the nominal parameters are shown in Table I. The DSP
board TMS320F28379D of TI, a dc voltage source of 78 V, and a
board of IGBT three-phase inverter are employed for controlling.
To get the feedback currents, a circuit board with Hall effect
current sensors HAX25 is made, and the currents are converted to
digital signals via analog-to-digital converter (ADC) functions.
The firmware inside the DSP computes dq-axis currents gener-
ates the inverter phase voltages via the SSROC algorithm and
provides the sine pulsewidth modulation signals for the IGBT
driver with switching frequency of 10 kHz via enhanced pulse Fig. 12. Comparison of rotor speeds by EMF-DO, EMF-CO, and PLL when
width modulator (ePWM) functions of DSP. Load torque can be nominal moment of inertia changes four times.
observed by an extended Kalman filter.
The experimental data are uploaded online to files via the se- where np ωLq iq and np ω(Ld id + ψf ) are FF terms used to
rial communications interface (SCI) between DSP and Simulink compensate the feedback control signals. The control gains are
of MATLAB. Because SCI takes a while (about 50 s in our sys- tuned based on [39], i.e., KPd = Ld wci , KPq = Lq wci , KId =
tem) to establish a connection before transmitting data, SSROC KIq = Rs wci , where wci is the cut-off frequencies, which is
is executed after the connection. a design parameter. Thus, the tuned PI gains guarantee that
The conventional SMO uses the sigmoid function F (x) = closed-loop transfer functions from i∗q and i∗d to iq and id become
2/(1 + e−x ) − 1. In the conventional PLL, the normalized po-
sition error, which a PI regulator dealing with, is defined as ωci ωci
Gid (s) = , Giq (s) = (73)
s + ωci s + ωci
êα cos θ̂ + êβ sin θ̂
f = −  . (69) where s is the Laplace variable. wci is high to yield Gid (s) ≈ 1 :
ê2α + ê2β id → i∗d and Gid (s) ≈ 1 : iq → i∗q . Given the nominal parame-
ters of the motor ESTUN EMJ-04APB22 in this article, the con-
Then, the parameters of the PI can be chosen as [p. 79, eqs. trol gains are tuned as KPω = 0.1, KIω = 0.2, KPd = KPq = 60,
(3.70)–(3.74)] [31] KId = KIq = 6000 [39].
 First, the sudden change of the torque disturbance is absent
np Td
Kp = 2ζ, Ki = ζ 2 , ζ = (70) from the experiment, but the moment of inertia and the load
Jeθ̂ max torque are changed when the PMSM is coupled with an in-
where eθ̂ max = 0.1 rad is the desired value of the estimation ductance motor. The experimental moment of inertia is four
error, and J is the nominal moment of inertia in Table I. times higher than the nominal one in Table I. Fig. 11 shows
The control voltages from classical PI controllers for the the estimated back-EMF of the back-EMF NN-observer, which
SMO-PLL scheme are given by [39] provides the extracted speed and position to the tracking error
⎧ NN-observer. Fig. 12 presents the actual and estimated rotor
 t
⎪
⎪ speed of three methods. When using the back-EMF directly,
⎨ ud = −KPd id − KId id dτ − np ωLq iq
0 
the chattering phenomenon occurs within EMF-DO, while it
t
⎪
⎪ is removed completely by the EMF-CO. The SMO-PLL also
⎩ uq=KPq (i∗q − iq ) + KIq (i∗q − iq )dτ + np ω(Ld id + ψf )
0
removes the chattering, but its tracking error is large and over-
shoot when speed reversal happens to the PMSM. Since the PI
(71) gains defined in (70)–(72) remain unchanged and, thus, are not
where adaptive with the variation of the moment of inertia. The rotor
 t
positions of the EMF-DO and the SMO-PLL in Fig. 13 are worse
i∗q = KPω (ωd − ω) + KIω (ωd − ω)dτ (72) than that of the EMF-CO at the moment of direction change of
0 the rotor.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 TAN et al.: NEURAL NETWORK OBSERVERS AND SENSORLESS ROBUST OPTIMAL CONTROL 12055

V. DISCUSSION
Despite the computational complexity, the two NNs are
needed to estimate and compensate the estimation error of rotor
position and speed. The proposed weight-tuning laws are effec-
tive in the case of unknown current dynamics and disturbances.
By using the data-driven method, NNs do not require offline
training using a prior sample set. However, how to choose and
balance the parameters in Lyapunov equations (8) and (29)
determine the convergence of NNs and the quality of initial
observers. In order to reduce trial and error times, one should
choose the parameters close to the dynamics of each particular
PMSM.
Note that the SSROC, based on a reinforcement learning
Fig. 13. Comparison of rotor positions by EMF-DO, EMF-CO, PLL, and technique, plays a very important role in online observer and
actual position. control. Robust optimal control signals with constraint bring the
reasonable inputs for training the NN-observer in reducing esti-
mation error, otherwise despite being UUB, the error trajectory
may not approach closer to the origin.

VI. CONCLUSION
In this article, the NN-based observers for the back-EMF
and tracking error are provided. The requirement of known
PMSM dynamics to design the observers is not needed. For
closed system control, the saturated robust optimal tracking
controller with partially unknown dynamics, saturation voltages,
and disturbance is proposed, where the feedforward control is
Fig. 14. Estimated back-EMF of NN-observer using for tracking error NN-
designed so that the problem of the speed and current tracking
observer when TL changes. control is converted to the equivalent optimal control problem.
The sensorless saturated robust optimal control law and the
worst disturbance law are approximated based on IRL technique
without identifying the drift dynamics. It is guaranteed that all of
closed state and approximation errors are uniformly ultimately
bounded and the value function converges to the near-optimal
value. The results from the simulation and experiment show the
effectiveness of the proposed control scheme. The future work
will concentrate on distributed cooperative optimal control of
multiple PMSMs.

REFERENCES
[1] F. Aghili, “Optimal feedback linearization control of interior synchronous
Fig. 15. Comparison of rotor speeds of EMF-DO, EMF-CO, and PLL when motors subject to time-varying operation conditions minimizing power
TL changes. loss,” IEEE Trans. Ind. Electron., vol. 65, no. 7, pp. 5414–5421, Jul. 2018.
[2] V. Smidl, S. Janous, L. Adam, and Z. Peroutka, “Direct speed control of
a PMSM drive using SDRE and convex constrained optimization,” IEEE
Trans. Ind. Electron., vol. 65, no. 1, pp. 532–542, Jan. 2018.
Next, a sudden load torque disturbance change with the [3] Y. Sun, M. Preindl, S. Sirouspour, and A. Emadi, “Unified wide-speed
sensorless scheme using nonlinear optimization for IPMSM drives,” IEEE
engaged brake is introduced to the experiment. Fig. 14 shows Trans. Power Electron., vol. 32, no. 8, pp. 6308–6322, Oct. 2017.
the estimated back-EMF of the back-EMF NN-observer, which [4] F. Toso, D. Da Ru, P. Alotto, and S. Bolognani, “A moving horizon
provides the extracted speed and position to the tracking error estimator for the speed and rotor position of a sensorless PMSM drive,”
IEEE Trans. Power Electron., vol. 34, no. 1, pp. 580–587, Apr. 2019.
NN-observer. Fig. 15 presents actual and estimated rotor speed [5] H. Yao, Y. Yan, T. Shi, G. Zhang, Z. Wang, and C. Xia, “A novel SVPWM
corresponding to the change of the torque disturbance. It can be scheme for field oriented vector controlled PMSM drive system fed by
seen that when using the back-EMF directly, the speed of the cascaded H-bridge inverter,” IEEE Trans. Power Electron., to be published,
doi: 10.1109/TPEL.2021.3054642.
EMF-DO fluctuates strongly at the time the torque disturbance [6] T. Chan, W. Wang, P. Borsje, Y. Wong, and S. Ho, “Sensorless permanent-
changes. Meanwhile, the tracking error of the compensated magnet synchronous motor drive using a reduced-order rotor flux ob-
speed of the EMF-CO is small. Since the moment inertia is server,” IET Electron. Power Appl., vol. 2, pp. 88–98, 2008.
[7] S. Bolognani, R. Oboe, and M. Zigliotto, “Sensorless full-digital PMSM
unknown against the calculated inertia, the estimated error of drive with EKF estimation of speed and rotor position,” IEEE Trans. Ind.
the SMO-PLL occurs even in steady state. Electron., vol. 46, no. 1, pp. 184–191, Feb. 1999.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.
 12056 IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 36, NO. 10, OCTOBER 2021

[8] H. Kim, J. Son, and J. Lee, “A high-speed sliding-mode observer for the [31] G. Wang, G. Zhang, and D. Xu, Position Sensorless Control Techniques
sensorless speed control of a PMSM,” IEEE Trans. Ind. Electron., vol. 58, for Permanent Magnet Synchronous Machine Drives. Singapore: Springer,
no. 9, pp. 4069–4077, Dec. 2011. 2020.
[9] Z. Qiao, T. Shi, Y. Wang, Y. Yan, C. Xia, and X. He, “New sliding-mode [32] S. Li, M. Zhou, and X. Yu, “Design and implementation of terminal sliding
observer for position sensorless control of permanent-magnet synchronous mode control method for PMSM speed regulation system,” IEEE Trans.
motor,” IEEE Trans. Ind. Electron., vol. 60, no. 2, pp. 710–719, Feb. 2013. Ind. Informat., vol. 9, no. 4, pp. 1879–1891, Nov. 2013.
[10] W. Xu, S. Qu, L. Zhao, and H. Zhang, “An improved adaptive sliding mode [33] H. Zargarzadeh, T. Dierks, and S. Jagannathan, “Optimal control of nonlin-
observer for middle-and high-speed rotor tracking,” IEEE Trans. Power ear continuous-time systems in strict-feedback form,” IEEE Trans. Neural
Electron., vol. 36, no. 1, pp. 1043–1053, Jan. 2021. Netw. Learn. Syst., vol. 26, no. 10, pp. 2535–2549, Oct. 2015.
[11] T. Orlowska-Kowalska and M. Dybkowski, “Stator-current-based MRAS [34] A. J.Van der shaft, “L2 -gain analysis of nonlinear systems and nonlinear
estimator for a wide range speed-sensorless induction-motor drive,” IEEE state feedback H∞ control,” IEEE Trans. Autom. Control, vol. 37, no. 6,
Trans. Ind. Electron., vol. 57, no. 4, pp. 1296–1308, Sep. 2010. pp. 770–784, Jun. 1992.
[12] D. Abramovitch, “Phase-locked loops: a control centric tutorial,” in Proc. [35] C. Wen, J. Zhou, Z. Liu, and H. Su, “Robust adaptive control of uncer-
Amer. Control Conf. (ACC), May 2002, vol. 1, pp. 1–15. tain nonlinear systems in the presence of input saturation and external
[13] L. Qu, W. Qiao, and L. Qu, “An enhanced linear active disturbance rejec- disturbance,” IEEE Trans. Autom. Control, vol. 56, no. 7, pp. 1672–1678,
tion rotor position sensorless control for permanent magnet synchronous Jul. 2011.
motors,” IEEE Trans. Power Electron., vol. 35, no. 6, pp. 6175–6184, [36] T. Basar and P. Bernhard, H Optimal Control and Related Minimax
Nov. 2020. Design Problems: A Dynamic Game Approach, 2nd ed. Boston, MA, USA:
[14] S. Lin and W. Zhang, “An adaptive sliding-mode observer with a tangent Birkhauser, 1995.
function-based PLL structure for position sensorless PMSM drives,” Int. [37] G. Chowdhary and E. Johnson, “Concurrent learning for convergence in
J. Electron. Power Energy Syst., vol. 88, pp. 63–74, 2017. adaptive control without persistency of excitation,” in Proc. 49th Conf.
[15] S. Ye and X. Yao, “A modified flux sliding-mode observer for the sen- Decis. Control, 2010, pp. 3674–3679.
sorless control of PMSMs with online stator resistance and inductance [38] L. N. Tan, “Distributed H∞ optimal tracking control for strict-feedback
estimation,” IEEE Trans. Power Electron., vol. 35, no. 8, pp. 8652–8662, nonlinear large-scale systems with disturbances and saturating actuators,”
Aug. 2020. IEEE Trans. Syst., Man, Cybern., Syst., vol. 50, no. 11, pp. 4719–4731,
[16] C. Xia, C. Guo, and T. Shi, “A neural-network-identifier and fuzzy- Aug. 2018.
controller-based algorithm for dynamic decoupling control of permanent- [39] F. Mendoza-Mondragon, V. M. Hernandez-Guzman, and J. Rodriguez-
magnet spherical motor,” IEEE Trans. Ind. Electron., vol. 57, no. 8, Resendiz, “Robust speed control of permanent magnet synchronous mo-
pp. 2868–2878, Aug. 2010. tors using two-degrees-of-freedom control,” IEEE Trans. Ind. Electron.,
[17] S. Tong, K. Sun, and S. Sui, “Observer-based adaptive fuzzy decentralized vol. 65, no. 8, pp. 6099–6108, Aug. 2018.
optimal control design for strict-feedback nonlinear large-scale systems,”
IEEE Trans. Fuzzy Syst., vol. 26, no. 2, pp. 569–584, Apr. 2018. Luy Nguyen Tan received the B.S. and M.Sc. degrees
[18] D. Liu, Y. Huang, D. Wang, and Q. Wei, “Neural-network-observer-based in computer science and automation and control en-
optimal control for unknown nonlinear systems using adaptive dynamic gineering and the Ph.D. degree in automation from
programming,” Int. J. Control, vol. 86, no. 9, pp. 1554–1566, 2013. the Ho Chi Minh City University of Technology, Ho
[19] D. Vrabie and F. Lewis, “Neural network approach to continuous-time Chi Minh City, Vietnam, in 1996, 2006, and 2015,
direct adaptive optimal control for partially unknown nonlinear systems,” respectively.
Neural Netw., vol. 22, no. 3, pp. 237–246, 2009. He is currently with the Faculty of Electronics
[20] H. Modares, F. L. Lewis, and Z. Jiang, “H∞ tracking control of completely Technology, Industrial University of Ho Chi Minh
unknown continuous-time systems via off-policy reinforcement learning,” City. His current research interests include adap-
IEEE Trans Neural Netw. Learn. Syst., vol. 26, no. 10, pp. 2550–2562, tive dynamic programming, distributed control, large-
Oct. 2015. scale systems, and deep learning.
[21] L. N. Tan, “Omnidirectional-vision-based distributed optimal tracking Dr. Tan was a reviewer for many journals, including IEEE TRANSACTIONS
control for mobile multirobot systems with kinematic and dynamic distur- ON NEURAL NETWORK AND LEARNING SYSTEMS, IEEE TRANSACTIONS ON IN-
bance rejection,” IEEE Trans. Ind. Electron., vol. 65, no. 7, pp. 5693–5703, DUSTRIAL ELECTRONICS, IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS,
Jul. 2018. IEEE TRANSACTIONS ON CYBERNETICS, IEEE TRANSACTIONS ON SYSTEMS,
[22] K. G. Vamvoudakis, M. F. Miranda, and J. P. Hespanha, “Asymptotically MAN CYBERNETICS: SYSTEMS, IET Control Theory and Applications, Neuro-
stable adaptive-optimal control algorithm with saturating actuators and computing, and Applied Soft Computing (Elsevier Journal).
relaxed persistence of excitation,” IEEE Trans. Neural Netw. Learn. Syst.,
vol. 27, no. 11, pp. 2386–2398, Nov. 2016. Thanh Pham Cong received the B.S. and M.Sc.
[23] C. Ben Regaya, W. Chakchouk, A. Zaafouri, A. Sallami, and A. Chaari, degrees in automation and control engineering from
“Adaptive discrete-time backstepping control of irrigation station,” in the Ho Chi Minh City University of Technology, Ho
Proc. Int. Conf. Adv. Syst. Elect. Technol., 2017, pp. 158–163. Chi Minh City, Vietnam, 2002 and 2006, respectively,
[24] C. Ben Regaya, A. Zaafouri, and A. Chaari, “Speed sensorless indirect and the Ph.D. degree from the Department of Control
field-oriented of induction motor using two type of adaptive observer,” in Science and Engineering, Huazhong University of
Proc. Int. Conf. Elect. Eng. Softw. Appl., 2013, pp. 1–5. Science and Technology, Wuhan, China, 2014.
[25] C. Ben Regaya, F. Farhani, A. Zaafouri, and A. Chaari, “A novel adaptive He is currently with the Faculty of Aeronautical
control method for induction motor based on backstepping approach using Electronics-Telecommunications, Vietnam Aviation
dspace DS 1104 control board,” Mech. Syst. Signal Process., vol. 100, Academy, Ho Chi Minh City, Vietnam. His current
pp. 466–481, 2018. research interests include ac motor, power electronics,
[26] C. B. Regaya, W. Chakchouk, A. Zaafouri, A. Sallami, and A. Chaari, adaptive dynamic programming.
“High-performance control of IM using MRAS-fuzzy logic observer,” Int.
J. Tomography Simul., vol. 80, no. 2, pp. 44–51, 2017.
[27] C. Ben Regaya, A. Zaafouri, and A. Chaari, “An adaptive sliding-mode
speed observer for induction motor under backstepping control,” Innov. Duy Pham Cong received the B.Eng. and M.Eng.
Comput., Inf. Control (ICIC Exp. Lett.), vol. 11, no. 4, pp. 763–771, 2017. degrees in electrical and electronics engineering from
[28] M. Abu-Khalaf and F. L. Lewis, “Nearly optimal control laws for nonlin- the Ho Chi Minh City University of Technology, Ho
ear systems with saturating actuators using a neural network approach,” Chi Minh, Vietnam, in 2001 and 2007, respectively,
Automatica, vol. 41, no. 5, pp. 779–791, 2005. and the Ph.D. degree from the Hunan University,
[29] B. A. Finlayson, The Method of Weighted Residuals and Variational Changsha, China, in 2012.
Principles. New York, NY, USA: Academic, 1990. He is currently with the Faculty of Electrical Engi-
[30] F. L. Lewis, S. Jagannathan, and A. Yesildirek, Neural Network Control neering Technology, Industrial University of Ho Chi
of Robot Manipulators and Nonlinear Systems. Philadelphia, PA, USA: Minh City. His key research interests include power
Taylor and Francis, 1999. conversion, electrical machine, and advanced control
techniques for electrical drives.

Authorized licensed use limited to: Karunya Institute of Technology and Sciences. Downloaded on March 01,2025 at 07:55:48 UTC from IEEE Xplore. Restrictions apply.