Вы находитесь на странице: 1из 6

2010 8th IEEE International Conference on ThCP4.

10
Control and Automation
Xiamen, China, June 9-11, 2010

Speed Control of 3-Phase Asynchronous Motor Using Artificial


Neural Network

Pham Thuong Cat, Le Hung Linh, and Minhtuan Pham

Abstract - Speed control of an alternating current (AC) motor In this paper, we propose a motor speed control method
has been one of the difficult control problems. Many using ANN with online self learning capability to
approaches have been proposed to solve these problems. In this compensate the parameter uncertainty in the dynamics model
paper, a speed control method is proposed using artificial of the AC motor with the use of rotor speed sensors.
neural network (ANN) with an online self-learning algorithm to
Simulation on Matlab is implemented to prove the efficiency
compensate uncertain parameters in the dynamics model of AC
motor. The global asymptotic stability of the control system is of this method.
proved using Lyapunov stability method. Simulation results on
MATLAB show the reliability and accuracy of the proposed II. ASYNCHRONOUS MOTOR CONTROL MODEL
method.
The dynamics model of the induction motor is normally
I. INTRODUCTION formulated by linear control theory in reference space [5],
Three-phase asynchronous alternating current motors are [6]. The structure of the alternating motor speed control
used widely in industries and domestic appliances because of system is illustrated in Figure 1 [7].
their cheap price and robust operation in harsh environment.
It is difficult to control the speed of the AC motor due to the
highly nonlinear characteristics and many uncertain
parameters including temperature-dependent rotor resistance,
magnetic flux, application-specific friction coefficient and
variable load. Thus, controlling AC motor has been a topic
that gains attention from many researchers for recent decades
[1], [2], [3], [9]. There are two directions:
1: Sensor-based: Using sensors to measure uncertain
parameters of the AC motor and then producing suitable
control signals. It increases the motor price and the
complexity of the control connection, but the control method
is more accurate with simple control algorithm.
2: Model-based: Using mathematic models instead of
sensors to estimate motor speed. This approach focuses
using methods such as Kalman filters, nonlinear filters or
sliding-mode observers [8], [9]. However, the control
performance relies heavily on the estimation algorithm and
the accuracy of motor model.
The dynamics system of the AC motor has many uncertain
parameters that makes classical control methods, with or
without sensors, further difficult to always achieve good
control quality, especially when the load is altered. In this
case, adaptive control methods, online identification methods
[4], [5], [6], [7] and neural network supported control
Figure 1. The indirect field-oriented induction motor drive.
method shall be applied.
Manuscript received October 30, 2009.
Pham Thuong Cat, Institute of Information Technology, Viet Nam The hardware of the motor speed control system includes
Academy of Science and Technology, Ha Noi Viet Nam, e-mail: an induction motor, a current controller, an power inverter
ptcat@ioit.ac.vn controlled by PWM (Pulse-Width Modulation) with a vector
Le Hung Linh, Faculty of Information Technology, Thai Nguyen
University, Thai Nguyen Viet Nam, e-mail: linhlehung@gmail.com
control technique, a current generator and a speed controller.
Minhtuan Pham, Institute of Information Technology, Viet Nam
Assuming that a current controller and an inverter ensure
Academy of Science and Technology, Ha Noi Viet Nam, e-mail:
pmtuan@ioit.ac.vn currents through a stator coil ias , ibs , ics of an induction motor

978-1-4244-5196-8/10/$26.00 ©2010 IEEE 1391


ThCP4.10

approaching to desired currents ias* * *


, ibs , ics , the dynamics From equation (2) and (3), we have
system of the induction motor can be formulated in dω
K .u (t ) = J + Bω + TL (8)
synchronous reference coordinate (d,q) [10] as follows: dt
3P Lm
 Lm Lm  where K = is torque constant, u (t ) = φr*iqs
*
is
 Rs − Lσ p ωe Lσ p ωe 4 Lr
Lr Lr 
  control voltage [3].
 Lm Lm   iqs  vqs  Laplace conversion of equation (8) yields
 −ωe Lσ Rs − Lσ p −ωe p    
 Lr Lr   ids  = vds  Ku ( s ) = Jsω (s)+Bω (s)+TL ( s ) (9)
 L Rr  φ   0  where s is Laplace operator. Then, we have:
 − Rr m 0 +p (ωe − ω )   qr   
 Lr Lr  φdr   0  1
  ω (s)= ( Ku ( s ) − TL ( s ))
Lm Rr Js + B
 0 − Rr −(ωe − ω ) +p (10)
 Lr Lr  TL ( s )
= G ( s )u ( s ) −
Js + B
(1)
K
where G ( s ) = (11)
The torque generated by induction motor is: Js + B
Induction motor control model using the vector method is
3P Lm e e e e dω
Te = (φdr iqs − φqr ids ) = J + Bω + TL (2) illustrated in Figure 2
4 Lr dt

where vqs and vds are voltages of stator on coordinate (d,p);


iqs and ids are stator currents on coordinate (d,p), φqr and φdr
are rotor fluxes on coordinate (d,p), Rs and Rr are stator and
rotor resistance; Ls, Lr Lm are stator, rotor, mutual
inductances; Lσ=(LrLs-Lm2)/Lr is leakage inductance, s is a
differential operator, p is a quantity of pole, ωe, ω and
ωsl=(ωe – ω) are the synchronous speed, rotor speed and slip Figure 2.The motor speed control model
speed, B is friction coefficient, J is the inertia of rotor, TL is
Motor model with uncertain parameters
load torque. Let us call the desire speed ωd and the rotor speed ω, and
Field-oriented control guarantees the speed of rotor flux rewrite equation (8)
vector to reach to the synchronous speed and rotor flux to J B T
u (t ) = ω + ω + L (12)
follow vertical axis d [10], so: K K K
φqr = pφqr = 0 (3) u (t ) = J eff ω + Beff ω + TL eff (13)
where
and φdr = φr = const (4) J 
J eff = = J eff + ∆J eff (14)
K
where φr is rotor flux
B 
Beff = = Beff + ∆Beff (15)
Replace (3) and (4) into (1) and (2) K
RL * T
ωsl* = r* m iqs (5) Teff = L (16)
φr Lr K
 
J eff , Beff are definite parameters
* φr*
ids = (6) ∆J eff , ∆Beff are indefinite parameters
Lm
and Substituting (14), (15) and (16) into (12), we have:
 
4 Lr u (t ) = J eff ω + Beff ω + Teff + ∆J eff ω + ∆Beff ω (17)
*
iqs = T*
* e
(7)
3PLmφr and set:
f = Teff + ∆J eff ω + ∆Beff ω (18)
Symbol ‘*’ is specific for the control signals that must be
obtained. In equation (7), if rotor flux does not change, then Because ∆J eff , ∆Beff and Teff are limited parameters, we
magnetic Te* will linearly change with respect to a current can find out the limit of f :
*
control signal iqs . Therefore, the vector control method for f ' ≤ δ 0 ; δ 0 ≥ 0. (19)
an induction motor has the same control model as that for
DC motor control.

1392
ThCP4.10

Replacing (18) into (17) we have the dynamics system that is an output signal of the neural RBF with one node in the
describes motor speed with uncertain parameters as follows hidden layer.
 
u (t ) = J eff ω + Beff ω + f (20a) f ' = fˆ + δ (29)
In summary, the motor control problem becomes where δ is the approximation error.
determining control signal u(t) that ensures motor speed ω We choose an RBF 3-layer neural network fˆ as below
reaching to desired speed ωd while parameters J eff , Beff and
- Input layer is error signal ε
changeable load Teff are uncertain - Middle layer is the hidden layer with output being
σ in the Gaussian form:
III. CONTROL MOTOR SPEED WITH UNCERTAIN PARAMETERS  (ε − c) 
σ = exp  − (30)
A. Feedback signal and friction coefficient  λ 
We choose where
u (t ) = u0 + u1 (20b) σ is output
c, λ are Gaussian parameters that are freely chosen.
where u0 is feedback signal in terms of PD with
- Linear output layer
compensation of friction coefficient Beff
  fˆ = wσ (31)
u0 = J eff (ω d − K D (ω − ωd )) + Beff ω (21) where w is the weights, online adjustable in control
ωd , ω d are desired speeds of motor process
K D > 0 is speed feedback coefficient - Learning algorithm is chosen as follows
u1 is compensated signal for indefinite w = −ηεσ (32)
parameters f which will be estimated. where η > 0 is learning coefficient that we choose freely.
Substituting (21) and (20b) into (20a) we have C. Control method u1
   
u1 + J eff (ω d − K D (ω − ω d )) + Beff ω = J eff ω + Beff ω + f (22)
We choose a signal to compensate uncertain parameters
thus,
 by using the output of the neural network fˆ as below
u1 − f = J eff ((ω − ω d ) + K D (ω − ω d )) (23)
 ε 
or u1 = Jˆeff u ' = Jˆeff  (1 + η ) fˆ − α  (33)
u f  ε 
ε + K D ε =  1 −  (24)
J eff J eff with α > 0.
where D. The control stability
ε = ω − ωd (25)
Let us set Theorem 1:
u Given the alternating induction motor, its model with
u' =  1 (26)
J eff indefinite parameters as in (12) and its unknown function f’
approximated by a neural network as in (29)-(30), the motor
f
f′=  (27) speed ω will follow the desired speed ωd , which implies the
J eff
speed error ε = (ω - ωd ) → 0 , if the controller and the
and replace into (24) we obtain
learning algorithm of the neural network are chosen as
ε + K D ε = u ' − f ′ (28)
follows:
Consequently, the control problem becomes finding u '   
u (t ) = J eff (ω d − K D (ω − ωd )) + Beff ω + J eff u ' (34)
that makes the system asymptotically stable with unknown
 ε 
f ' . We shall use a neural network to approximate u1' = (1 + η ) fˆ − α  (35)
 ε 
function f ' .
w = −ηεσ (36)
B. Finding a neural network to approximate unknown with K D , η , α > 0 .
parameter f ' :
The structure of the control system is described in Figure
In accordance with the Stone-Weierstrass theorem [10], an 3. The control signal consists of two main parts:
RBF (Radial Basis Function) neural network can  
u0 = J eff (ω d − K D (ω − ωd )) + Beff ω is the PD and friction
approximate the unknown structured continuous function. As
compensation feedback part and u1 is a compensation part of
a result, function f ' can be approximated by function fˆ that a neural network with the online learning algorithm to
approximate indefinite parameters. This theorem is proved

1393
ThCP4.10

by Lyapunov stable method that guarantees the global system (17) is globally asymptotically stable; and ω (t ) → ωd
asymptotic stability of the control system as below. implies that the motor speed follows closely the desired
speed with speed error approaching 0. Theorem 1 is proved.

IV. SIMULATION
To prove the efficiency of algorithm, we examine the
motor control system with uncertain parameters in 2 cases -
with or without a neural network - and a varying load.
It was assumed that the desired speed signal ωd varying in
Figure 3: The motor speed control model with on-line neural network the trapezoidal shape: rise linearly from 0 rad/s to desired
speed ωd = 1 [ rad / s ] in 10s, maintain stable speed ω = ωd
Proof:
for 40s, and slow down to a complete stop in 10s (see Figure
Let us choose positive-definite function V as follows 4a)
1 2 The motor control system is modeled as in equation (7)
V=
2
( ε + w2 ) (37)
where uncertain parameters are assumed as below
B = Bˆ + ∆B
V>0 when ε , w ≠ 0 ; V=0 only ε , w = 0 , V → ∞ when
ε,w → ∞ Bˆ = 3 (Nms/rad)
∆B = 0.3
Differentiating V with accordance to time t, we have J = Jˆ + ∆J
V = εε + ww (38) Jˆ = 5 (Nms2/rad)
From (28), it is induced that ∆J = 0.5
K = Kˆ + ∆K
ε = u '- f '- K D ε (39)
Kˆ = 6
Replacing (39) into (38) produces ∆K = 0.6
The load was formulated as follows
V = − K ε 2 + ε ( u '- f ′ ) + wT w
D (40)
TL = TˆL + ∆TL
(Nm)
With the online learning algorithm (36), w = −ηεσ , we ∆TL = 1.5sin(2t ) + 0.5sin(50t )
calculate
TˆL has an amplitude changing with time as in Figure 4
ww = −ηε wσ (41)

By substituting (41) and (29) into (40), we have


V = − K ε 2 + ε ( u '− (1 + η ) wσ − δ )
D (42)

ε
Substituting (35) into (42) with u ' = (1 + η ) wσ - α ;
ε
Figure 4: a) Load changes suddenly; b) Various load TL
α > 0 yields
ε Load TˆL + ∆TL affecting on the motor varied but was not
V = − K D ε 2 + ε (−α − δ ) ≤ − K D ε 2 − α ε + ε . δ
ε (43) known to the controller. The load varied in 3 factors: sudden
2
changes, noise caused by the 50Hz frequency and noise
≤ − K D ε − α ε + ε .δ 0 caused by the vibration in low frequency 2Hz (see Figure 4b)
If we choose α = δ 0 + µ ; µ > 0 we have A. Simulation with PI control [11]

−α ε + ε .δ 0 ≤ - µ ε (44)

Giving (44) into (43), we have

V ≤ − K D ε 2 − µ ε ≤ 0 (45)
The time derivative of the Lyapunov-candidate-function
V < 0 when ε ≠ 0 and V = 0 only if ε = 0 . According to
the equilibrium Lyapunov, a speed error ε → 0 . Therefore,
Figure 5. Design of PI control for Induction motor

1394
ThCP4.10

K I = K P = 2 . The result of simulation is illustrated in


Figure 6.
As in Figure 6d shows that the load varies largely due
to the combination of 3 factors. The error between the
desired speed and the rotor speed is quite large and hard to
converge (Figure 6b). The speed of motor is not stable and
varies with a big error of about 25% (Figure 6c).
B. Simulation without using a neural network
The control signal has a form
u (t ) = u0 + u1
 
u0 = J eff (ω d − K D (ω − ωd )) + Beff ω
Figure 6. The results of simulation with PI control for Induction motor
u1 = 0 a) desire speed ωd ; b) speed error ε ;
K D = 2,5. The result of simulation is illustrated in Figure 7. c) motor speed ω ; d) control signal u(t);
The error between the desired speed and the rotor speed is
quite large and hard to converge (Figure 7b). The speed of
motor is not stable and varies with a big error of about 40%
(Figure 7c). The speed decreases when the load increases
suddenly at time 15s, 25s, 35s.
To improve, we will use a neural network with online
learning algorithm to compensate the change of load.
C. Simulation with using a neural network

The control signal u (t ) has a form


u (t ) = u0 + u1
  Figure 7. The results of simulation without a neural network
u0 = J eff (ω d − K D (ω − ωd )) + Beff ω
a) desire speed ωd ; b) speed error ε ;
 ε  c) motor speed ω ; d) control signal u(t);
u1 = Jˆeff u ' = Jˆeff (1 + η ) wσ − α 
 ε 
w = −ηεσ
where as: α = 3; c = 1; λ = 0, 3;η = 0.2 . The results of
simulation are shown in Figure 8.
Comparing the results in Figure 8 to those in Figure 6 and
Figure 7, we recognize that the control performance is much
improved with the motor speed close to the desired speed
(Figure 8b). Even when the load varies, the rotor speed can
quickly reach to the required speed.
The average error < 10−3 and at the times when the load
varies suddenly the pulse-shaped error < 0.05 (Figure 8a). It
shows the adaptive ability of the system and the efficiency of Figure 8. The results of simulation with a neural network
the motor control method using a neural network.. a) speed error ε ; b) motor speed ω and ωd
c) control signal u(t); d) various coefficient w
V. CONCLUSION
In this paper, we proposed a motor speed control method
using an artificial neural network with online learning REFERENCES
algorithm to compensate indefinite parameters and large [1] W. Leonhard: Control of Electric Drives. SpringerVerlag, 2001.
variable loads in the dynamics model of the alternating [2] P. Krause: Analysis of Electric Machinery. McGrawHill, 1986.
motor. The system stability is justified by the equilibrium [3] R. J. Wai: Robust Decoupled Control of Direct Field-Oriented
Induction Motor Drive. IEEE Transactions on Industrial, Vol. 52, No.
Lyapunov theorem. The results of the simulation on Matlab 3, June 2005.
show that the quality and efficiency of the control is quite [4] S. Rao, M. Buss, and V. Utkin: An adaptive sliding mode observer for
good. It clearly shows the highly adaptive capability of the induction machines. Proceedings of the 2008 American Control
Conference, Seattle, Washington, USA, June 2008, pp. 1947–1951.
proposed method using an online learning neural network

1395
ThCP4.10

[5] R. Marino, S. Peresada, and P. Valigi: Adaptive input output


linearizing control of induction motors. IEEE Transactions on
Automatic Control, Vol. 38, No. 2, pp. 208–221, Feb 1993.
[6] I. Utkin, J. G. Guldner, and J. Shi: Sliding Mode Control in
Electromechanical Systems. Taylor & Francis, 1999.
[7] K. Halbaoui, D. Boukhetala, and F. Boudjema: A New Robust Model
Reference Adaptive Control for Induction Motor Drives Using a
Hybrid Controller. Proceedings of the 2008 International Symposium
on Power Electronics, Electrical Drives, Automation and Motion
pp.1109-1113.
[8] Z. Yan and V. Utkin: Sliding mode observers for electric machines an
overview. Proceedings of the IECON 02, Vol. 3, No. 2, November
2002, pp. 1842 – 1847.
[9] Derdiyok, Z. Yan, M. Guven, and V. Utkin: A sliding mode speed and
rotor time constant observer for induction machines. Proceedings of the
IECON 01 , Vol. 2, December 2001, pp. 1400–1405.
[10] P. Marino, M. Milano, F. Vasca: Linear quadratic state feedback and
robust neural network estimator for field-oriented-controlled
induction motors. IEEE Trans. Ind. Electron, Vol46, pp. 150–161,
1999.
[11] Z. Liu: Hybrid Speed Control with Sliding-Mode plus Self-Tuning PI
for Induction Motor Drive. Proceedings of the 49th IEEE International
Midwest Symposium on Circuits and Systems MWSCAS '06, San
Juan, Puerto Rico, August 2006, Vol. 1, pp. 500 - 504.

1396

Вам также может понравиться