You are on page 1of 5

Robust Sensorless Control for Induction Motor Drives

Fed by a Matrix Converter with Variable Structure

Model Reference Adaptive

* Chonbuk National University ** Korea Advanced Institute *** Aalborg University

Division of Electronics and of Science and Technology Institute of Energy Technology
Information Engineering 335 Gwahangno, Yuseong-gu, Pontoppidanstraede 101, DK-9220,
664-14, Deokjin-dong, Jeonju, Daejeon 305-701, KOREA Aalborg East, DENAMRK
561-756, KOREA Email:

Abstract - This paper presents a new robust sensorless control In [6, 7], the authors proposed adaptive methods to tune
system for high performance induction motor drives fed by a the control gain by using neural networks (NN). However,
matrix converter with variable structure. The lumped some parameters of the NN are updated based on a
disturbances such as parameter variation and load disturbance back-propagation strategy, which normally leads to a local
of the system are estimated by a variable structure approach minima problem. Uncertain parameters in VSC can be
based on model reference adaptive scheme. A Reduced Order
estimated by adopting proper adaptive laws. Recently, VSC
Extended Luenberger Observer (ROELO) is also employed to
bring better responses at the low speed operation. Experimental combining model reference adaptive control (MRAC) has
results are shown to illustrate the performance of the proposed been intensively researched in relation to complex nonlinear
system. systems and has shown improved transient and steady state
responses under a bounded disturbance [1-5, 8]. In [2],
I. INTRODUCTION unknown parameters in a 2-degree of freedom robot
manipulator are adaptively updated, resulting in improved
Variable structure control (VSC) employs a discontinuous
transient response. However, only exponential stability is
manifold called a sliding surface. VSC can make a controlled
guaranteed, and there is no estimation strategy for unknown
system robust to a certain degree with the help of switching
boundary values. In [4], actuator nonlinearities such as
control actions [4, 11]. However, VSC entails two well
dead-zones and back-lash are considered, and inertia and
known problems, a chattering phenomenon and unknown
joint flexibility terms are approximated using a NN based on
control gain. To date, numerous attempts to address these
back stepping and a VSC approach. However, the norm of
issues have been reported [1-8, 11, 12]. For the chattering
the unknown term is assumed to be bounded by a known
problem, practically efficient approaches have been reported
constant, and all variables are uniformly bounded rather than
including making use of a saturation function [15], and a soft
asymptotically stable. In [8], for a perturbed large scale
computing approach [12]. Control gain in VSC is an effective
system with a time-varying delay interconnection, an
element that influences control performance [1-8].
Power grid
v A , vB , vC

Z r + Proposed ROBUST
iqse + q-axis current
vqee va PWM
Controller (II) - control 2/3 vb Signal
- Indirect
Zˆ r +
vdse &
ej vc SVM Gate Matrix Converter
e d-axis current Drives
ds - control

iq Z sl+ Zˆ e Tˆ
³ dt
Wid +
iqse idse
+ ia , ib , ic
Zˆ r Reduced order 3/2
extended + + &
- -
Luenberger e-j
observer [13] 3 kW Induction

iqcomp idcomp v A , vB , vC
ia , ib , ic
Non-linearity compensation
using PQR transformation [9]

Fig. 1: The proposed robust sensorless vector control for matrix converter drives.

1-4244-0655-2/07/$20.00©2007 IEEE 2423

adaptive algorithm for estimating an uncertain upper bound
based on a VSC frame was proposed. e1n  k1n e1n  k0 n e1n 0 (4)
In this paper, to achieve the speed sensorless process, or
speed estimation is completed by a conventional preferred
d ª e1n º ª 0 1 º ª e1n º
method proposed by the authors [13]. Adaptive sliding-mode « » «k » « » Aen e n (5)
control with a NN observer is presented to control the dt ¬e1n ¼ ¬ 0 n  k1n ¼ ¬ e1n ¼
Induction motor drive, which includes parametric and Which means that the dynamic system (2) can be stable,
un-modeled uncertainties. Under the framework of MRAC,
lim ^e1n , e 2 n ` ^0, 0` with an appropriate design constant k n .
the actual dynamics of the induction motor drive are t of
As easily guess, the nominal controls are not perfect for the
controlled so as to follow the reference dynamics, which is a
practical model (1), and additional control is required.
nominal model.
Time varying uncertain parameters are approximated by In the next section, to cope with the composite
the NN in an on-line manner, and the sliding control is uncertainty U , NN observer and adaptive SMC based on
designed so as to compensate for the reconstruction error. MRAS are presented, and the stability issue will be detailed
Finally, the actual system robustly follows the nominal model described.
dynamics even under hard nonlinearity. NN weight vectors
B. Radial basis function network (RBFN)
and sliding mode control gains including unknown variation
bounds are adaptively updated with stabilizing adaptive laws The radial-basis function network (RBFN) is a kind of NN,
based on Lyapunov criterions. The proposed control scheme and is simply constructed with input, output, and hidden
of sensorless vector control for induction motor drives fed by layers of normalized Gaussian activation functions.
a matrix converter is shown in Fig. 1. Experimental results
are presented to show the effectiveness of the proposed
system. o1 " " oj om


(VS-MRAC) w1i
z1 zi
A. System modeling using RBFN (Radial Basis Function
" " zl

Concerning parametric uncertainty and un-modeled

dynamics, mechanical formulation of induction machines can
be described by state variable form with the set of as
follows: xrb1 " " xrbi xrbn

x1 x2 Fig. 2: Schematic diagram of the RBFN.

x2 f  f c x2  g  g c u  hTL  H (1)
f x2  g u  U A schematic diagram of the RBFN is shown in Fig. 2. In
general, each of the hidden nodes has a normalized Gaussian
where f  B J , g  KT J and U f cx2  g cu  hTL  H activation function as follows [9]:
Without considering uncertainty U, nominal dynamics can 2
exp ª  x rb  m i / 2 Pi2 º
be rewritten with the nominal vector x n > x1n x2 n @T as follows: -i (x rb ) ¬ ¼ (6)
zi T i ( x) ' l l
¦-k (xrb ) ¦ exp ª  x rb  m k / 2 Pk2 º
x1n x2 n k 1 k 1
¬ ¼
x2 n fx2 n  gun .
where z i is the ith RBFN of Z z1 , z2 , ! , zi , ! zl  R l ,
Stabilizing computed torque controls for the nominal x rb xrb1 , xrb 2 , ! , xrbn  R n is the input vector, mi  R and

modes (1) can be easily derived as follows: n

Pi  R denote the center and width of the ith Gaussian
function, respectively.
un g 1 x2d  fx2 n  k Tn e n (3) Therefore, the hidden node zi gives a large response as the
input vector approaches m i . The output of the RBFN is
where kn > k0 n k1n @ ,
>e1n e1n @ >e1n e2 n @ , and
e1n x1d  x1n , simply the weighted sum of the hidden layer output, and is
e2 n x  x2 n .
d given in the following linear combination form:

Applying (3) to (2), nominal error dynamics is computed

as follows:

with the notation of 
W W  W* as follows:
Rj ¦w
k 1
k ˜ zk
[ w j1 w j 2 ... w jl ][ z1 z2 ... zl ]T w Tj z U wT z  U wT z  wT * z  G wT z  G (11)
where w jk , (k 1, 2,......, l ) , is the weight between the kth
hidden node and the RBFN output , and wTj  Rl is the vector where G means a reconstruction error of the neural network.
of w jk ’s. Using (11), original lumped uncertainty can be re-written
Finally, the output vector set is o o1 , o2 , ! , om  R m . as
Some theorems reflect that the 3-layered RBFN is a universal
approximator like a fuzzy and multi-layered NN, if the U wT z  wT z  G (12)
activation functions used in the hidden layer are infinitely
differentiable and in polynomial forms [9, 10]. The RBFN To describe the theorem, the following assumption is
offers a viable alternative to the three-layered NNs in many required.
applications such as signal processing, pattern recognition, Assumption. The following inequality holds
control and functional approximation. To update the RBFN
parameters such as the weight vectors, and the center and G d] (13)
width of the Gaussian functions, some simple update laws of
delta-learning rules can be derived [10]. where ] is a positive definite constant.
Now, the main theorem will be described.
C.Variable Structure-Model Reference Adaptive Control
System (VS-MRAC) Theorem. Let the overall control input u un  uc , and
As was shown in the previous section, nominal state vector
uc g 1 k1e1  uc1
based on known parameters can be controlled at will. In this
paper, nominal dynamics is concerned as a reference model §1 T O · (14)
¨ c A e e   w z  [ ¸ sgn S
for the actual dynamic model. © c2 c2 ¹
From (1) and (2) with additional controls uc
where u un  uc , error dynamics of e1 x1n  x1 can be where un is determined as (3), O is a strictly positive constant,
derived as follows: and [ is a control parameter which is to be updated by its
adaptive law.
x1 f x1n  x1  guc  U
(8) By using RBFN, lumped uncertainty is to be approximated
k1e1  f e1  uc1  U as follows:

where uc g 1 k1e1  uc1 Uˆ wT z (15)

Error dynamics can be summarized as follows:
The update laws for the weight vector W and control gain
e Ae e  B eMe (9) [ with [ * | ] are determined as

ªe1 º ª 0 1º ª0 º w J w c2 Sz
where e «  » , Ae « k f » , B «1 » , and Me uc1  U (16)
¬e1 ¼ ¬ 1 ¼ ¬ ¼ [ J ks c2 S
In (8), f is a negative constant, and so, states of the
dynamic system (1) can follow those of the nominal models
Then, desired tracking performance S as well as all
(2) with additional control uc and appropriate design
parameters is to be asymptotically stable.
constant k1 .
Now, with the error vector defined in (10), sliding surface For deriving the adaptive laws, we define a Lyapunov
and its derivative can be defined as follows: function as the following form.

S cT ˜ e 1 S 2  1 w  w* T w  w*  1 [  [ * 2
(10) Ve (t )
2 2J w
2J (17)
c ˜ e c ˜ Ae e  B eMe

where T The derivative of (18) is to be as follows:

c >c1 c2 @
In this paper, system uncertainty U is to be modeled by
neural network, and its modeling error U can be depicted

1 T 1  has a robust characteristic against parameter variation. This
Ve SS  w w
Jw Jk superiority can also be seen in Fig. 5(a) and (b). There is an
1 1  abrupt load disturbance at 5 s. For the speed control with
S cT Ae e  c2 uc1  c2 U  wT w
 [[ RBFN control, it takes about 0.5 s for the tracking error to be
Jw J ks
zero, but the speed control without the RBFN observer uses 2
§1 O · s. From these experimental results, the best speed control
d S cT Ae e  c2 S ¨ cT Ae e   wT z  [ ¸ sgn S
© 2 c2 ¹ characteristics in the sensorless vector control system are
observable when the lumped disturbances exist. Especially,
1 T 1 
 c2 S U  w  w  [[ the proposed control scheme has robust characteristics
Jw Jk
against parameter variation even though there is a negligible
O S  c2 S w z  c2[ S  c2 S w T z  w
 Tz G external disturbance effect.
1 T 1 
w   [[
Jw Jw
d O S  c2 S w T z  c2[ S  c2 S w T z  c2 Sw
1 1 
 c2 S [   Tw
w  [[
Jw Jk
1 1  
O S 
 T w
w   J w c2 Sz 

[ [  J k c2 S
O S .
From (17) and (18), it is guaranteed that all signals e , W  ,
and [ are bounded, and from (7), e is also bounded.
With integration of (18), 2³ f Ve (t ) d  ³ f 1 S , the following
0 0 2
equation is true:
(a) speed and phase current
³ S d 2 Ve 0  Ve f (19)

Because Ve 0 is bounded Ve f , and e is non-increasing,

³0 S is also bounded. By using Babalat’s lemma, S o 0 for


To confirm the validity of the proposed control algorithm,
experiments are carried out. The low speed operation is
especially in focus during the experiments. The hardware (b) speed and phase current, A zoom of the zero crossing of speed from (a)
Fig. 3: Forward and reverse operation.
consists of a 3-phase, 380 V, 50 Hz, 4-pole, 3 kW induction
motor and power circuit with a matrix converter. The
modulation used a double-sided SVM algorithm with the
minimized number of commutations and the commutation of
the bi-directional switches was performed in a four-step way
depending on the output current sign [13].
Fig. 3(a) and (b) show the speed and phase current
responses of the proposed sensorless vector control system in
the forward and reverse operation. Fig. 3(b) shows a zoom of
the zero-crossing in the speed response. Fig. 3(a) and (b)
show the speed response for a ramp-speed reference with a
slope 200 rpm/s. It is noted here that a smooth and stable zero
crossing of speed is obtained. Fig. 4(a) and (b) show the
experimental results when the inertia value variation has
occurred abruptly. For speed control without RBFN
disturbance observer, 1 s is taken to overcome the affection
for the tracking error to be zero. (a) with IP controller
However, speed control with RBFN disturbance observer

approximated by the radial basis function network. In
addition, control laws for stabilizing a controlled system and
adaptive laws for updating both the weights of RBFN and
unknown control parameters have been established.
Experimental results show that the proposed control
scheme can provide good performance at the low speed
region and robust and stable characteristics against parameter
variations and load disturbances.

[1] C. W. Park and Y. W. Cho, “Adaptive tracking control of flexible joint
manipulator based on fuzzy model reference approach,” IEE Proc.
Control Theory Appl., 2004, vol. 150, no. 2, pp. 198-204, 2004.
[2] S. K. Tso, X. Xu, and H. Y. Shum, “Variable Structure Model
Reference Adaptive Control of Robot Manipulators,” Proc. of IEEE Int.
(b) with proposed controller
Conf. on Robot and Auto., pp.2148-2153, April 1991.
Fig. 4: Parameter variation at 500 rpm.
[3] L. Hsu, “Variable Structure Model-Reference Adaptive Control
(VS-MRAC) Using Only Input and Output Measurements: Part II,”
Proc. of IEEE 27th Conf. on Decision and Control., pp.2396-2400, Dec.
[4] R. J. Wai, ‘Total sliding-mode controller for PM synchronous servo
motor drive using recurrent fuzzy neural network’, IEEE Trans.
Industrial Electronics, vol. 48, no. 5, pp.926-944, 2001.
[5] M. A. Demetriou, I. G. Rosen, “Variable structure model reference
adaptive control of parabolic distributed parameter systems,” Proc. of
the American Control Conference, pp. 4371-4376, May 2002.
[6] C. H. Tsai, H. Y. Chung, and F. M. Yu, “Neuro-sliding mode control
with its applications to seesaw systems,” IEEE Trans. Neural networks,
vol. 15, no. 1, pp. 124-134, 2004.
[7] R. J. Wai, “Total sliding-mode controller for PM synchronous servo
motor drive using recurrent fuzzy neural network,” IEEE Trans.
Industrial Electronics, vol. 48, no. 5, pp. 926-944, 2001.
[8] C. H. Chou and C. C. Cheng, “A decentralized model reference
adaptive variable structure controller for large-scale time-varying delay
systems,” IEEE Trans. on Automatic Control, vol. 48, no. 7, pp.
1213-1217, 2003.
(a) with IP controller [9] K. B. Lee and F. Blaabjerg, “Improved Sensorless Vector Control for
Induction Motor Drives Fed by Matrix Converter Using Nonlinear
Modeling and Disturbance Observer,” IEEE Trans. Energy Conversion,
vol. 21, no. 1, pp. 52-59, Mar. 2006.
[10] C. T. Lin and C. S. Lee, ‘Neural fuzzy systems, Prentice Hall’ (Upper
Saddle River, NJ, 1996, pp.328-330)
[11] J. E. Slotine and W. Li, ‘Applied nonlinear control’ (Prentice Hall,
[12] H. X. Li, H. B. Gatland, and A. W. Green, “Fuzzy Variable Structure
Control,” IEEE Trans. System, Man, And Cyber.-Part B: Cyber., vol.
27, no. 2, pp.306-312, 1997.
[13] K. B. Lee and F. Blaabjerg, “Reduced Order Extended Luenberger
Observer Based Sensorless Vector Control Driven by Matrix Converter
With Non-Linearity Compensation,” IEEE Trans. Industrial
Electronics, vol. 53, no. 1, pp. 66-75, Feb. 2006.

(b) with proposed controller

Fig. 5: Load disturbance at 500 rpm.

In order to realize high performance control of induction
motor drives fed by matrix converter, a robust speed
controller using a VS-MRAC has been proposed in this
To achieve a robust control characteristic against
un-modeled disturbances such as parameter variations and
load disturbances, the un-modeled disturbances have been