Professional Documents
Culture Documents
Sung-Hoe Huh,ofKyo-Beum
Control, Automation, and Systems,
Lee, Dong-Won vol. 3, and
Kim, Ick Choy, no. 4, pp. 612-619,
Gwi-Tae Park December 2005
Abstract: A robust adaptive speed sensorless induction motor direct torque control (DTC)
using a neural network (NN) is presented in this paper. The inherent lumped uncertainties of the
induction motor DTC system such as parametric uncertainty, external load disturbance and
unmodeled dynamics are approximated by the NN. An additional robust control term is
introduced to compensate for the reconstruction error. A control law and adaptive laws for the
weights in the NN, as well as the bounding constant of the lumped uncertainties are established
so that the whole closed-loop system is stable in the sense of Lyapunov. The effect of the speed
estimation error is analyzed, and the stability proof of the control system is also proved.
Experimental results as well as computer simulations are presented to show the validity and
efficiency of the proposed system.
Keywords: Neural network (NN), uncertainty observer, robust adaptive speed sensorless
control, speed estimation error.
The universal approximators (UAs) such as NN, FL, where x is the input vector, wi 's, i = 1," , l are the
and hybrid system of FL and NN have been weights between the i th node and the output, and
successfully applied to many nonlinear control W ∈ R n is the vector of wi ’s.
problems. The design objective of NN or FL is aimed
at approximating some nonlinear mappings into
3. ROBUST STABLE SENSORLESS SPEED
idealistic approaches. This means that an arbitrary
CONTROLLER IN DTC SYSTEM
function f : R n → R is to be approximated by NN or USING A NEURAL NETWORK
FL. It is well known that any continuous function
3.1. Integration and Proportion (IP) control approach
Input Layer Hidden Layer Output Layer In the field of speed control systems, IP speed
v11 z1 control is a generally used scheme because it has
X1 v12 several advantages such as negligible overshoot in its
w1
. z2
step tracking response, good regulation characteristics
. compared to proportion and integration (PI) control
. w2
.
scheme and zero steady state error [6]. The IP
.
. y controller is to be designed to stabilize the speed
X n− 1 . control loop, and its parameters Ki, Kp are derived to
. wl show the desired response. A block diagram of the
Xn zl basic induction motor DTC system including the IP
V W speed controller, torque and flux comparators fed
switching logic generator, adaptive observer and 3-
Fig. 1. Structure of the neural network. level inverter system is presented in Fig. 2. Now, if
Torque comparator
Te* + 3-level inverter
ω r _ ref + eω
IP
eT Switching
− − Logic
+ eλ
λ̂ s _ ref Generator Gating
− signal
ω̂ r λ̂ s T̂ e Flux comparator
Adaptive 3
Observer 2
Fig. 2. Block diagram of the basic DTC for 3-level inverter system.
614 Sung-Hoe Huh, Kyo-Beum Lee, Dong-Won Kim, Ick Choy, and Gwi-Tae Park
-
variable form:
ω r* + + U
Te 1 ωr
K /s KT
I Js+B
ωr
_
-
+
x n = An x n + B n u m + CnTL , (7)
where x n = ω r , An = − B J , Bn = K T J , u m = Te*
Kp
and Cn = − 1 J .
The above equation (7) is expressed by nominal
Fig. 3. Simplified DTC system. values, but in most practical cases, the inherent
uncertainties in the induction motor system should
some reasonable assumptions are adopted into exist. Besides, for the speed sensorless control, the
nominal model dynamics such as rotor speed, then estimated rotor speed instead of the sensed value is to
flux and torque estimations and regulations are stably be fed back for generating a control signal based on
worked, and the DTC system can be further simplified the assumption that the estimated rotor speed is
as indicated in Fig. 3. perfectly identical to the real one. However, in any
For the speed control loop on Fig. 3, the transfer real system, estimation error is inevitable even though
it is negligible, and thus, the influence of the error
function GIP ( s) for the reference signal ω r* is to be
must be analyzed on feedback control systems.
derived as follows: Now, considering the inherent uncertainties with
ωr ( s) KT KI the estimated speed, equation (7) is further expressed
GIP ( s ) = = . (2) as in the following state variable form:
ω r* ( s) 2
Js + (B + K T K P ) s + K T K I
x q = A n x q + Bn u o + ε , (8)
For the unit step reference signal, speed response can
be derived as follows:
where ε = (Δ A n x q + ΔBn u o + Cn TL +ρ) , x q = ω r ,
ωr ( s) =
KT KI J 1
⋅ . (3) A n = − B J , B n = k T J , Cn = −1 J , u o is control
s + ( B + KT KP
2
J ) s + ( KT KI J ) s input, ρ is the unmodeled uncertainties, and Δ A n ,
Δ Bn are modeling errors of A n , Bn .
And the following equation is true.
Let the two variables, ê x and e x , be defined as
ω n2
ωr ( s) = , (4)
(
s s 2 + 2ς s + ω n2 ) ê x = ω r* − ω r and e x = ω r* − ω r , (9)
σ1 ⎛ z ϕ ⎞ σ 2 ⎛ ϕ ⎞
2 2 (20)
Then, the tracking error as well as other signals =− ⎜W− ⎟ − ⎜ξ − ⎟
involved in the closed-loop system are unified 2 ⎝ σ1 ⎠ 2 ⎝ σ2 ⎠
ultimately bounded (UUB).
− Ke 2x + λ ,
Proof: For deriving the adaptive laws for unknown
where the constants z and λ are defined as
constant limit, ξ , and compensating control input, up,
we define Lyapunov function as in the following form. z = sup Z ,
z
Ve (t ) = 1 ( W − W* )T ( W − W* )
2γ W ( z ϕ )2 ϕ2 σ 2 σ 2
(15) λ= + + A n ϕ 2 + ϕζ + 1 W* + 2 ξ * .
2σ 1 2σ 2 2 2
+ 1 (ξ − ξ * ) 2 + 1 e2x ,
2γ ξ 2
From (20) and the Lyapunov’s direct method, it is
* ,
easily observed that the upper bounds for ê x , W
where W is an optimal value of W . In the
equation (15), because e x is an unknown value, the and ξ are
following equality should be required.
λ ≤ z ϕ + 2λ , ξ ≤ ϕ + 2λ ,
ex ≤ , W (21)
( )
e x =x d -x q =x d - x q + ϕ = e x − ϕ . (16) K σ1 σ1 σ2 σ2
The time derivative of (16) is which shows the UUB of the signals.
An update law for the weight vector V can be easily
e x = x d - x q = x d - ( A n x q + Bn ( u r + u p ) + ε ) derived by using the gradient decent rule with back
(17) propagation algorithm. Let the activation (sigmoid)
( )
T
= − A nϕ − K e x − Bn u p + W − W* Z +ζ.
( )
−1
function be Τ( I ) = 1 + e−α I , and its derivative is to
Let W = W - W* , ξ ξ − ξ * and take the time be ∂Τ ∂I = α (1 − Τ ) Τ . Now, using the gradient decent
derivative of the Lyapunov equation, then rule, we can find the following equation.
Ve (t ) = − Kex2 + Bn u P (ϕ − ex ) − A n exϕ ∂e 2 l
Δv1k = −ηv = ηv x1 ∑ δ1k , (22)
∂v1k
+ A nϕ 2 + Kexϕ − ϕζ + W T Z ( e − ϕ ) (18) m =1
x
1 T 1 T ∂y ∂z ∂z
+ γ W W + γ ξ ξ + exζ where ηv > 0 , δ1k = 2e wk1 k = δ k1wk1 k ,
W ξ ∂I z ∂I xk ∂I xk
616 Sung-Hoe Huh, Kyo-Beum Lee, Dong-Won Kim, Ick Choy, and Gwi-Tae Park
400
feedback control response.
200
However, the NN observer approximates the
uncertainties to minimize the rotor speed error, and
0
0 1 2 3 4 5 6 7 8 9 10
the compensated control input is applied for
reconstruction error by the feedforward manner. Only
15 under 0.15s is taken for the speed error to be zero, and
as a result, the tracking performance can be drastically
control input[A]
10
improved despite abrupt uncertainties. The lowest part
5
in Fig. 5(b) shows its identification ability.
0
400
10
is still unstable. The sampling time of the control
5
cycle is set at 200 µ s for the torque ripple reduction
0
and the proposed algorithms. Because the experimental
-5
0 1 2 3 4 5 6 7 8 9 10
setup is for a high power induction motor DTC system,
time[sec] the switching frequency remains in the region of 500
(b) Proposed control (CASE 1) Hz – 1.0 kHz. Fig. 7 shows the comparatively
experimental results of CASE 1 at 500rpm of the
Fig. 4. Simulation results of CASE 1. desired rotor speed. In this case, both controllers have
Sensorless Speed Control System Using a Neural Network 617
600 600
400 400
speed[rpm]
speed[rpm]
200 200
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
15 15
10 10
control input[A]
control input[A]
5 5
0 0
-5 -5
-10 -10
-15 -15
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time[sec] time[sec]
600 600
speederr[rpm]
speederr[rpm]
400 400
200 200
0
0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
observed uncertainty biased uncertainty
0 0
-5 -5
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
5 5
0 0
-5
-5
0 1 2 3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10
time[sec]
time[sec]
3φ
Gate Drives
500[rpm] 500[rp m ]
ω r*
ω r* ωˆ r ωr
(a) Conventional IP control (CASE 1). (a) Conventional IP control (CASE 2).
500[rpm]
ω r*
ωˆ r
(b) Proposed control (CASE 1). (b) Proposed control (CASE 2).
Fig. 7. Experimental results of CASE 1. Fig. 8. Experimental results of CASE 2.
effective tracking performances as gathered from the can be drastically improved. A simple NN structure
simulated results. IP gains are derived by (4) and (5) quickly approximates the uncertainties, and the
as in the simulation processes. Even though a additional compensating control input engaged to
modeling error from the simplified mechanical model control the input. As a result, tracking performance can
used for experimental test exists, the IP control shows be drastically improved under serious disturbances.
robust response by the integrating operation. The With these experimental results, the feasibility and
overall DTC system including 3-level inverter is robustly stable characteristics are established. The
approximated by a simple mechanical model as shown proposed control scheme shows relatively small
in Fig. 3, and the approximation error and unmodeled tracking error compared to conventional IP control
dynamics are not serious enough to deteriorate the schemes even when a definite inherent uncertainty is
speed regulation. However, for CASE 2, concerning abruptly engaged.
the keener disturbances in which the parameter
variations and abrupt external disturbance are engaged 5. CONCLUSION
at 2.5s, the distinctions between the two approaches
are obviously displayed. As shown in Fig. 8, the In this research, a robustly stable speed controller
proposed control scheme shows advanced robust for the induction motor system using the NN observer
performance comparing the IP scheme. When the test is presented. To cope with the inherent uncertainties
condition is engaged, for the proposed control (Fig. such as parametric uncertainty, external disturbance
8(b)), it takes about 0.1s for tracking error to be zero, and unmodeled dynamics, the NN is used as an
but IP control (Fig. 8(a)) needs 1.5s for the recovery uncertainty observer approximates the inherent
process, with a greater degree of tracking error and uncertainties. Moreover, the stability analysis of the
fluctuation. IP control is based on the nominal model whole control system considering the effect of the
and is not effective for the serious disturbances of speed estimation error is presented. A control law for
CASE 2. It takes a great deal of time to melt the stabilizing the system and adaptive laws for updating
effects of unmodeled uncertainties that deteriorate both the weights in the NN and a bounding constant
feedback response. However, using the NN observer are established so that the whole closed-loop system is
and the compensated control input, speed response stable in the sense of Lyapunov. The proposed control
Sensorless Speed Control System Using a Neural Network 619
algorithm is relatively simple and requires no Kyo-Beum Lee received the B.S. and
restrictive conditions on the design constants for the M.S. degrees in Electrical and Electronic
stability. Moreover, various kinds of control or motor Engineering from Ajou University,
systems such as vector control and PM synchronous Suwon, Korea, in 1997 and 1999,
respectively. He received the Ph.D.
motors can be applied with this approach. In this
degree in Electrical Engineering from
research, as one of the application examples, a high Korea University, Seoul, in 2003.
power 3-level fed induction motor direct torque Currently, he is an Assistant Professor
control system is presented. To achieve the speed with the Institute of Energy Technology,
sensorless process, speed estimation is completed by a Aalborg University, Aalborg, Denmark. His research
conventionally preferred method. The effectiveness interests include electric machine drives and power
and validity of the proposed system are shown in electronics.
simulation and experimental results.
Dong-won Kim is currently a Ph.D.
candidate in Electrical Engineering at
REFERENCES Korea University, Seoul, Korea. His
[1] R. J. Wai, “Hybrid control for speed sensorless research interests are Humanoid and
induction motor drive,” IEEE Trans. on Fuzzy mobile robots, soft computing such as
Systems, vol. 9, no. 1, pp. 116-138, 2001. fuzzy systems, neural networks, genetic
[2] B. Heber and L. Xu, “Fuzzy logic enhanced algorithms, GMDH-type algorithms,
speed control of an indirect field-oriented and their applications to complex
induction machine drive,” IEEE Trans. on Power systems.
Electronics, vol. 12, no. 5, pp. 772-778,
September, 1997. Ick Choy received the B.S., M.S. and
Ph.D. degrees in Electrical Engineering
[3] P. Vas, Sensorless Vector Control and Direct
from Seoul National University, Korea,
Torque Control, Oxford Univ. Press, 1998. in 1979, 1981 and 1990, respectively.
[4] Z. Uykan, C. Guzelis, M. Celebi, and H. Koivo, From 1981 to 2003, he was with the
“Analysis of input-output clustering for determin- Intelligent System Control Research
ing centers of RBFN,” IEEE Trans. on Neural Center, Korea Institute of Science and
Networks, vol. 11, no. 4, pp. 851-858, 2000. Technology, Korea. Since 2003 he has
[5] W. Rudin, Principle of Mathematical Analysis, been with Kwangwoon University,
McGraw-Hill, New York, 1976. where he is currently an Associate Professor in the
[6] S. H. Kim, T. S. Park, J. Y. Yoo, and G. T. Park, department of information and control engineering. His
“Speed-sensorless vector control of an induction main research interests include microprocessor application,
high-performance drives and emerging technologies.
motor using neural network speed estimation,”
IEEE Trans. on Power Electronics, vol. 48, no. 3, Gwi-Tae Park received the B.S., M.S.
pp. 609-614, September 2001. and Ph.D. degrees in Electrical
[7] H. Kubota, K. Matsuse, and T. Nakano, “DSP- Engineering from Korea University in
based speed adaptive flux observer of induction 1975, 1977 and 1981, respectively. He
motor,” IEEE Trans. on Industry Applications, was a Technical Staff Member in the
vol. 29, no. 2, pp. 344-348, 1993. Korea Nuclear Power Laboratory and
[8] K.-B. Lee, J.-H. Song, I. Choy, and J.-Y. Yoo, an Electrical Engineering Faculty
“Torque ripple reduction in DTC of induction Member at Kwangwoon University, in
1975 and 1978, respectively. He joined
motor driven by three-level inverter with low
Korea University in 1981 where he is currently a Professor
switching frequency,” IEEE Trans. on Power in Electrical Engineering. He was a Visiting Professor at the
Electronics, vol. 17, no. 2, pp. 255-264, March University of Illinois, UC and the University of Maryland,
2002. in 1984 and 1996, respectively. Dr. Park is presently serving
as the President of the Intelligent Building System (IBS)-
Sung-Hoe Huh received the Ph.D. Korea. His research interests include soft computing
degree in Electrical Engineering from techniques, adaptive signal processing, computer & control
Korea University 2004. He was a networks and their applications to robots, home automation,
student research worker at Intelligent security systems, smart car, and IBS.
Robotics Research Center in Korea
Institute Science & Technology, Seoul,
Korea from 1995 to 1998. He has been
with the University of Wisconsin since
July, 2005, where he is currently a post
doctor in Wisconsin Electric Machines & Power Electronics
Consortium. His major interests include motor drives,
mobile robots, adaptive and intelligent controls.