296-1-1433-2-10-20211228
296-1-1433-2-10-20211228
296-1-1433-2-10-20211228
2 (2021) 20-31
Original Article
Back-stepping Control of Switched Reluctance Motor with
Artificial Neural Network based Flux Estimator
Phi Hoang Nha1,2, Pham Hung Phi2, Dao Quang Thuy3, Le Xuan Hai1,
Pham Xuan Dat2, Nguyen Ngoc Linh4,*
1
Ha Noi University of Industry, 298 Cau Dien, Bac Tu Liem, Ha Noi,Vietnam
2
Hanoi University of Science and Technology, 1 Dai Co Viet, Hai Ba Trung, Ha Noi, Vietnam
3
Ministry of Science and Technology, 133 Tran Duy Hung, Cau Giay, Hanoi, Vietnam
4
VNU University of Engineering and Technology, 144 Xuan Thuy, Cau Giay, Hanoi, Vietnam
Abstract: The paper presents a new approach to the speed control of a switched reluctance motors
(SRM) is that using a back-stepping controller combining with an artificial neuron network based
flux estimator. The nonlinear mathematical model of switched reluctance motor (SRM) is
established and the back-stepping control strategy is applied to control SRM. The ANN will be used
to estimate the flux of the motor instead of approximated model or experimental values. The ANN
flux estimator was trained off-line using backpropagation algorithm. The stability of the closed-loop
control system was analyzed and proved according to the Lyapunov stability criteria. The simulation
is carried out with both traditional back-stepping controller and the back-stepping controller
combining with ANN based flux estimator. The numerical simulation results confirmed quality of
the back-stepping controller as well as the feasibility of using ANN in the flux estimator.
Keywords: Switched Reluctance Motor (SRM), Back-stepping Control, Flux Estimator, Artificial
Neural Networks (ANN), Backpropogation Algorithm.
________
*
Corresponding author.
E-mail address: nlnguyen@vnu.edu.vn
https://doi.org/10.25073/2588-1086/vnucsce.261
20
P.H. Nha et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 37, No. 2 (2021) 20-31 21
therefore is a challenging problem and depends including armature voltage equation, electro-
much on the mathematic model of SRM. Many magnetic torque equation and mechanical
studies have tried to establish the nonlinear equation (1):
mathematical model of SRMs [1-6, 8, 13-14]. d j
Some mathematical models have been u j = R.i j +
dt
developed, however, their coefficients are
W j' (1)
difficult to determine (because depending on the T j ( , i j ) =
type of SRM, the size of SRM...[1, 8]. In [5],
d 2
author introduced a non-linear model of SRM J 2 = Te − Tl
and did a linearization after that. Besides, one of dt
the most important components in SRM’s in which j = 1, 2, …, m
model, used in controller design, is the motor’s uj is voltage of phase j
magnetic flux. Some methods have been
proposed to estimate motor’s flux of SRMs R is resistor of phase j
based on the experimental results or using an ij is current of phase j
approximately mathematical model [4, 14] ѱj is flux of phase j
However, these methods usually have Te is torque of phase
difficulties in reality and sometimes can make a
Tl is torque of load
large error. To overcome this drawback, in this
research, we will use an artificial neuron J is moment of inertia
network (ANN) to estimate the flux of SRMs. W j' is the electro-magnetic energy
Accordingly, non-linear model of the which is determined in (2):
switched reluctance motor, which including the ij
phase switch and the dynamics of the SRMs, has W j' ( , i j ) = j ( , i j )di j (2)
been established and been used to synthesize the 0
back-stepping controller. In this model, motor’s Electrical torque in SRM is a nonlinear
flux will be estimated by ANN based flux function of only current if the magnetic circuit is
estimator instead of using approximate model or linear.
measured data. This ANN is trained offline
The total torque produced is equal to the sum
based on the experimental data and can continue
of the moments in the phases:
automatically be trained in the future.. In order
m
to verify the efficiency and feasibility of Te ( , i1 , i2 ,..., im ) = T j ( , i j ) (3)
proposed method, several simulations are j =1
implemented both for back-stepping control
To control the SRMs, we need to determine
(BTP) with traditional model of SRM and back-
the magnetic flux characteristic ѱj (θ, ij) as
stepping control with the model using the ANN
accurately as possible. For convenience in the
based flux estimator (BTP - ANN flux
process of research and development of control
estimator). Conclusions and some future works
algorithms, the function of magnetic flux
are pointed out in final.
characteristic in [4] will be used:
−i j f j ( )
2. Mathematic Model of the SRMS j ( , i j ) = s (1 − e ) (4)
s f 2 ( x1 ) − x4 f 2 ( x1 ) (9)
+ N r1 − [1 + x f ( x )] e
f 2 ( x1 ) x1
2 4 2 1
1 s f3 ( x1 )
1 s f 3 ( x1 )
= + 2
N r 1 − [1 + x5 f 3 ( x1 )]e − x5 f3 ( x1 )
f c ( x) = 2
J f3 ( x1 ) x1
N r 1 − e− x5 f3 ( x1 )
J f 3 ( x1 ) x1
s f ( x ) 1 s f3 ( x1 )
+ 2 4
f 4 ( x1 ) x1
1
N r 1 − [1 + x6 f 4 ( x1 )]e − x6 f 4 ( x1 )
g c ( x) = 2
J f3 ( x1 ) x1
N r − f3 ( x1 )e− x5 f3 ( x1 )
− Bx − mgl sin( x )
2 1
1 s f 4 ( x1 )
f d ( x) = 2
J f 4 ( x1 ) x1
Nr 1 − e− x6 f4 ( x1 )
−1 (
Rx3 + s e − x3 f1 ( x1 ) )
x3 = − s e − x3 f1 ( x1 ) f1 ( x1 ) f ( x )
(
x 1 1 x
3 x1 2) (10)
g d ( x) =
1 s f 4 ( x1 )
2
Nr − f 4 ( x1 )e− x6 f4 ( x1 )
−1
J f 4 ( x1 ) x1
+ s e− x3 f1 ( x1 ) f1 ( x1 ) u1
P.H. Nha et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 37, No. 2 (2021) 20-31 23
+ f c ( x) + g c ( x) x5 + f d ( x) + g d ( x) x6 (15) s
B mgl (
6 x1 )
x f4 ( x1 ) x
2
− x2 − sin( x1 ) −1
J J qd ( x) = s e− x6 f4 ( x1 ) f 4 ( x1 )
Differentiating equation (16), we have:
Rewritten equations from (10) to (13) as
x2 = f a ( x) + g a ( x) x3 + g a ( x) x3 follows:
+ f b ( x) + g b ( x) x4 + g b ( x) x4 x3 = pa ( x ) + qa ( x )u1
x4 = pb ( x) + qb ( x)u2
+ f c ( x) + g c ( x) x5 + g c ( x) x5 (16) (17)
x5 = pc ( x) + qc ( x)u3
+ f d ( x) + g d ( x) x6 + g d ( x) x6 x6 = pd ( x) + qd ( x)u4
B mgl
− x2 − cos( x1 ) x1 Substituting (17) to (16), we have:
J J
f ( x) + g a ( x) x3 + g a ( x) pa ( x)
From (11) to (14), we set: x2 = a
Rx + g a ( x)qa ( x)u1
3 f ( x) + gb ( x) x4 + g b ( x) pb ( x)
( )
−1
pa ( x) = − s e − x3 f1 ( x1 )
f1 ( x1 ) + s e− x3 f1 ( x1 ) + b
+ gb ( x)qb ( x)u2
(
3 x1 )
x f1 ( x1 ) x
2
f ( x) + g c ( x) x5 + g c ( x) pc ( x)
−1 + c
qa ( x) = s e− x3 f1 ( x1 ) f1 ( x1 ) + g c ( x)qc ( x)u3 (18)
Rx f ( x) + g d ( x) x6 + g d ( x) pd ( x)
4 + d
pb ( x) = − s e− x4 f2 ( x1 ) f 2 ( x1 ) + ( s e − x4 f2 ( x1 ) )
−1
+ g d ( x)qd ( x)u4
B mgl
(
4 x1 )
x f2 ( x1 ) x
2
− x2 −
J J
cos( x1 ) x1
−1
qb ( x) = s e− x4 f2 ( x1 ) f 2 ( x1 ) The switched reluctance motor works with
the principle of voltage supply for each phase. If
Rx the number of phases is 4, we have u j = k j u ,
5
( )
−1
f3 ( x1 ) + s e− x5 f3 ( x1 ) with j=1, 2, 3, 4); kj is a phase transition key, so
pc ( x) = − s e − x5 f3 ( x1 )
it can only take 2 values, 0 or 1. Equation (18)
(
5 x1 )
x f3 ( x1 ) x
2
can be represented as follows:
−1
qc ( x) = s e− x5 f3 ( x1 ) f3 ( x1 )
24 P.H. Nha et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 37, No. 2 (2021) 20-31
and: e1 = z1 − zd = z2 − zd (25)
ga ( x)qa ( x)k1 + gb ( x)qb ( x)k2
G(x)= Considering e2 = z2 − 1 where 1 is virtual
+ gc ( x)qc ( x)k3 + gd ( x)qd ( x)k4 control signal for the first subsystem.
We have another expression of equation
Substitution to (25), we have:
(19) as follows:
B mgl e1 = z1 − zd = z2 − zd = e2 + 1 − zd
x2 =F(x) + G( x) − x2 − cos( x1 ) x1 (20) (26)
J J
To determine the virtual control signal to
If we set:
ensure e1 → 0 we choose Lyapunov function:
B mgl
f ( x) = F(x) − x2 − cos( x1 ) x1
J J (21) 1
V1 = e12 (27)
g ( x) = G ( x) 2
We have: Differentiating V1 over time, we have:
x2 = f ( x) + g ( x)u (22)
V1 = e1e1 = e1 ( e2 + 1 − zd ) = −c1e12 + e1e2 (28)
Setting x2 = z1 , we have the state model of
the SRMs as presented in (23): In order to have (30), the virtual control
signal has a following form:
z1 = z2
(23) 1 = −c1e1 + zd
z2 = f ( x ) + g ( x ) u
(29)
e2 = z2 − 1 (31)
V=
2
( e1 + e2 ) = V1 + e22 = V2
1 2 2 1
2
(37)
From (25) we have:
Differential V over time, we have:
e2 = z2 − 1 = f ( x ) + g ( x ) u − 1 (32)
V = −c1e12 + e1e2 + e2 f ( x ) + g ( x ) u − 1 (38)
To determine control signal u to ensure
e2 → 0 , we also choose Lyapunov function: Substituting u from equation (36) to (38),
we have:
1
V2 = V1 + e22
f ( x ) − c2 e2 − e1
(33)
2
V = −c1e12 + e1e2 + e2
Differential over time, we have: − f ( x ) − 1 − 1 (39)
V = −c1e12 − c2 e22 0
V2 = V1 + e2e2 (34)
→QED.
Substituting equations (28) and (33) to (34),
we have: 3.2. ANN Based Flux Estimator for SRM
V2 = −c1e12 + e1e2 + e2 f ( x ) + g ( x ) u − 1 (35) Based on magnetic flux equation (4) of
SRM, we will approximate function fj(θ) by an
Selecting the control signal of system from ANN The structure of ANN is illustrated in
(35): Figure 1 [10]. Inputs of the network are the
current in stator and position of rotor, and
−c2 e2 − e1 − f ( x ) − 1
u=
^
g ( x)
(36) outputs are fj(θ) and j ( ) . The network
includes 5 layers P f , R f , S f , T f , Uf with the
with c2 is positive constant. letters representing the relative number of
Theorem: The SRM has been represented in neurons in respective layers f (forward) is in
state-space model (23) controlled by a back- every variable of the network.
stepping controller defined in (36) where are i) Feedforward algorithm of the estimator
positive constants to ensure a stable Lyapunov
closed system. In every layer, x is the input and y is the
output of each neuron. Signal propagation and
Proof: We choose the Lyapunov function for activation function is illustrated.
closed loop system has a following form:
Layer P f : neuron p has input and output Layer Uf : Magnetic flux and torque of each
presented: phase are determined in this layer. After
obtaining f j ( ) and a j ( ) from the output of
x pf = and y pf x pf = x pf( ) (40)
T f , approximated magnetic flux function
with p = 0 . Activation function equals to 1
j ( ) and approximated torque function
in this layer.
Tˆj ( , i j ) are calculated in layer U f according to
f
Layer R : Each neuron r in this layer has
the formula:
input and output presented:
xuf = ytf i j and yuf = 1 − e u =
−x
( )
f
x f − c f 2 j (45)
xrf = y pf .w prf and yrf = exp − r f r (41)
r with u = 0 . In which i j is the current in stator
windings.
with p = 0 and r = 0...R f . In which crf and rf
are center and range of Gaussian Activation Electro-magnetic torque is approximated as:
function. w prf is the weight between 2 layers P f df j ( ) − i j f j ( )
Tˆj ( , i j ) = [1 − (1 + i j f j ( ))e ] (46)
and R . f
f ( )d
j
2
Layer S f : s neurons are in this layer. Inputs In above formulas, we do not need to
and outputs are calculated: consider the saturation of flux s because ANN
Rf has adaptive structure through weights and
x = yrf .w rsf
s
f
(42) activation functions. To train the ANN,
r =0
backpropagation algorithm is used.
and
ii). Backpropagation algorithm
xf −cf 2
In ANN, the main purpose of network
ysf = exp − s f s (43)
s training is to update the network’s weights
(Figure 2). The algorithm for training the ANN
with r = 0...R f and s = 0...S f . In which csf and with the forward model is error backpropagation
algorithm. Flowchart of training algorithm is
sf are center and range of Gaussian Activation presented in Figure 3.
function. w rsf is the weight between 2 layers R f
and S f .
r =0
Weights of the network are trained to Layer T f − S f : In this layer, the weights
minimize an objective function: change:
(
1 f
e ( k ) ) with k = 1,..., K E f
2
Ef =
f
(47)
2 w stf = stf − f = stf t f ysf (50)
w st
in which K f is the number of the input and
output and e f is the error between approximated in which stf is the learning coefficient of the
weight between.
value of magnetic flux j and actual value j .
Layer S f − R f : Error e f is directly fed S f
by chain rule. Therefore, error s f is determined:
ysf
s f = s f w stf (51)
xsf
E f
w rsf = rsf − f
= rsf sf yrf (52)
w rs
yrf Sf
rf =
xrf
s =0
s
f
w rsf (53)
e f = j ( k ) − j ( k ) with k = 1,..., K f (48) in which prf is the learning coefficient of the
weight between 2 layers.
Layer U f − T f : Because weights between
layers are uniform, e f is directly fed into T f in Weights w prf , w rsf and w stf are updated
chain rule. Therefore, error t f determined: through energy function E f . The change in
weights w stf , w rsf , w prf will be added to
yuf ytf
t f = −e f i j (49) weights in the ANN as in (55):
xuf xtf
28 P.H. Nha et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 37, No. 2 (2021) 20-31
R = 0.05 ( ) = 100
B = 0.2 l2 = 2500
l = 2 ( m)
Figure 4. The back-stepping controller combined Simulation results of the performance of the
with ANN based flux estimator. proposed control system are shown in Figure 5,
Figure 6, Figure 7 and Figure 8
4. Simulation Result
In Figure 5, the approximated magnetic flux
The proposed control system in the paper is from the ANN based flux estimator is compared
verified by the simulation results carried out with its values calculated by approximated
through Matlab/SIMULINK software. mathematic model (Figure 5a). The error of the
The design criteria for this problem are: two values is shown in Figure 5b. It can be seen
that, the value of the error is nearly zero. It is
o No static error. proved that the ANN work well.
o Overshoot less than 5%.
In Figure 6, the electromagnetic torque of
o Settling time less than 0.5s. SRM is presented. It is clearly that the ripple still
exists. This problem usually appears with SRM
and need to be improve in this research.
P.H. Nha et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 37, No. 2 (2021) 20-31 29
BTP-ANN
BTP
flux estimator
Static error
(rad/s)
10−4 10−4
Setting time
0.45 0.45
(s)
Overshoot (%) 0 0
Figure 6. Torque characteristic.
30 P.H. Nha et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 37, No. 2 (2021) 20-31
References
[1] A. Berdai, A. Belfqih, J. Boukherouaa, F.
Figure 8. Speed response and the error in case of Mariami, A. Hmidat, V. Vlasenko, V. Titjuk,
changing the set point
P.H. Nha et al. / VNU Journal of Science: Comp. Science & Com. Eng., Vol. 37, No. 2 (2021) 20-31 31
Similarity and Comparison of the Electrodynamics [8] L. Zeng, H. Yu, Research on a Novel Rotor
Characteristics of Switched Reluctance Motors Structure Switched Reluctance Motor, Phys,
SRM with Those of Series DC Motors, Procedia, Vol. 24, 2012, pp. 320–327.
Engineering, vol. 7, 2015 pp. 36–45. https://doi.org/10.1016/j.phpro.2012.02.048.
[2] A. Nirgude, M. Murali, N. Chaithanya, [9] M. I. Spong, R. Marino, S. M. Peresada, D. G.
S. Kulkarni, V. B. Bhole, S. R. Patel, Nonlinear Taylor, Feedback Linearizing Control of Switched
Mathematical Modeling and Simulation of Reluctance Motors, IEEE Transactions on
Switched Reluctance Motor, IEEE International Automatic Control, Vol. 32, 1987, pp. 371-379,
Conference on Power Electronics, Drives and https://doi.org/10.1109/TAC.1987.1104616.
Energy Systems (PEDES), 2016, pp. 1–6. [10] M. J. Grimble, P. Majecki, Nonlinear Industrial
[3] G. Rigatos, P. Siano, S. Ademi, Nonlinear Control Sytems: Optimal Polynomial Systems and
H-Infinity Control for Switched Reluctance State) Space Approach, Springer-Verlag London,
Machines, Nonlinear Engineering, vol. 9, no. 1, 2020.
2020, pp. 14-27. [11] V. T. Nguyen, S. F. Su, N. Wang, W. Sun,
[4] H. L. Huy, P. Brunelle, A Versatile Nonlinear Adaptive Finite-Time Neural Network Control for
Switched Reluctance Motor Model in Simulink Redundant Parallel Manipulators, Asian Journal of
Using Realistic And Analytical Magnetization Control, Vol. 22, No. 6, 2019, pp. 2534–2542,
Characteristics, 31st Annual Conference of IEEE https://doi.org/10.1002/asjc.2120
Industrial Electronics Society, 2005, [12] V. T. Nguyen, C. Y. Lin, S. F. Su, W. Sun, M. J.
pp. 1556–1561. Er, Global Finite Time Active Disturbance
[5] J. A. Makwana, P. Agarwal, S. P. Srivastava, Rejection Control for Parallel Manipulators With
Modeling and Simulation of Switched Reluctance Unknown Bounded Uncertainties, IEEE
Motor, Lect. Notes Electr. Eng, 2018, Transactions on Systems, Man, and Cybernetics:
pp. 545–558, Systems, 2020, pp. 1-12,
https://doi.org/10.1007/978-981-10-4762-652. https://doi.org/10.1109/tsmc.2020.2987056
[6] K. Deguchi, S. Sumita, Y. Enomoto, Analytical [13] O. Ustun, A Nonlinear Full Model of Switched
Method Applying a Mathematical Model for Reluctance Motor With Artificial Neural Network,
Axial-Gap-Switched Reluctance Motor, Electr. Energy Conversion and Management, Vol. 50,N.
Eng. Japan (English Transl. Denki Gakkai 9, 2009, pp. 2413–2421,
Ronbunshi), Vol. 196, 2016, pp. 30–38, https://doi.org/10.1016/j.enconman.2009.05.025
https://doi.org/10.1002/eej.22749.
[14] X. Sun, K. Diao, Z. Yang, G. Lei, Y. Guo, J. Zhu,
[7] L. Shen, J. Wu, S. Yang, X. Huang, Fast Flux Direct Torque Control Based on a Fast Modeling
Measurement for Switched Reluctance Motors Method for a Segmented-Rotor Switched Reluc-
Excluding Rotor Clamping Devices and Position tance Motor in HEV Application, IEEE Journal of
Sensors, IEEE Transactions on Instrumentation Emerging and Selected Topics in Power
and Measurement, Vol. 62,N. 1, 2013, pp. 185– Electronics, Vol. 9, No. 1, 2021, pp. 232-241.
191, https://doi.org/10.1109/TIM.2012.2212598.