Professional Documents
Culture Documents
Abstract—The induction motor is a MIMO, nonlinear and high investigated. However, the decoupling and linearization of
coupling system. The reversibility of the induction motor is multivariate nonlinear system demand exact mathematics
testified. Consequently, a pseudo-linear system is completed by model of controlled object. The induction motor as a multi-
constructing a neural network inverse (NNI) system and variable, nonlinear, strong-coupled system, its rotor
combining it with the motor system. The inverse can transform parameters change very prominently with the time-varying
the MIMO nonlinear system into two SISO linear subsystems (i.e., conditions. All unmodeled dynamic effects, such as the change
rotor speed and flux subsystems). In order to approach the
of load torque, the existence of disturbance and magnetic
inversion exactly in operation of the motor, the control method
saturation make the differential geometry and inverse system
online learning based on NNI system is proposed, in which
connection value can be amended continuously on-line to make
method difficult to apply in practice.
the NN adapt to the changes of environment to strengthen its Recently, neural networks (NNs) have attracted
robustness. Experiment results have shown that NN can be considerable attention as candidates for novel tools in
adjusted in the control process. The good applicability of NN identification and control of nonlinear systems. They were also
along with the strong stability and robustness of the system can applied to the induction motor control[5-6]. In decoupling
be achieved by using the proposed method. control, if NN is used to approach the inverse model of system,
the decoupling effect will be perfect. The neural network
Keywords- -neural network; inverse system; intelligent control; inverse (NNI) system method can realize the linearization
induction motor decoupling control for generic non-linear system, and its
structure is simple enough to be applied in engineering easily [7].
I. INTRODUCTION However, the NNI is trained offline in certain operating
conditions of the controlled system with fixed parameters, so
With the well-known advantages of reliability, simple the NNI control scheme is lack of robustness and adaptability
construction, and low weight, the induction motors have been to some extent.
gradually utilized in place of DC motors with drawbacks of
spark, corrosion and necessity of maintenance. However, the
induction motors have more complexities in control II. MATHEMATIC MODEL AND REVERSIBILITY ANALYSES
characteristics than DC motors due to their coupled and For voltage controlled PWM inverter of induction system,
nonlinear time-varying dynamics. The high performance when the nonlinear and time-delay of the inverter, magnetic
control of induction motor necessarily requires that the control saturation and iron loss of the induction motor are ignored, the
of the flux and the electromagnetic torque must be entirely model of induction motor in rotating coordinates (d-q) is [8]:
decoupled. So, in the past years the decoupled control of
induction motor has been investigated largely [1-4]. di Lm Rr Lr Rs
2 2
Lm Rr Lm Rr u sd
( )isd ( n p Z r isq )isq \r
sd
® (2)
(6)
Define state variable as °̄v2 y2
x [ x1 , x2 , x3 , x4 ]
T
[isd , isq ,\ r , Zr ]
T So
System output as u M ( y1 , y1(1) , v1 , y2 , y2 (1) , v2 ) (7)
T T
y [ y1 , y2 ] [ x3 , x4 ]
III. NNI CONTROL OF THE INDUCTION MOTOR
Control variable as
T T
u [u1 , u 2 ] [u sd , u sq ] A. Offline Training of NNI
Thus equation (1) can be modified as Because it is very difficult to accurately model the
x f ( x , u) induction motor system, the inversion (7) can not be expressed
analytically. In order to implement the inverse system, a NN
Lm] 1 u1 can be trained offline to approximate the inversion (7).
° ] 3 x1 (n p x4 x x2 ) x2 ] 1] 2 x3 V L The dynamic data and static data of the induction motor
° 3 s
° Lm] 1 x2 u2 are needed to train NNI. In order to obtain the inversion, the
° ( n p x4 ) x1 ] 3 x2 n p ] 2 x3 x4 (2) given signal should cover the running domain of the induction
® x3 V Ls motor completely. In this paper, excited signal with different
° scope is used as the system input. The sampling signals
° Lm] 1 x1 ] 1 x3
° include the inputs of the induction motor which are two
Tl
°] 4 x2 x3 voltage signals and the outputs which are flux and speed. The
¯ J high accuracy seven-point numerical algorithm is adopted to
Rr Lm Lm 2 Rr Lr 2 Rs L calculate the 2-order differential coefficient of flux and speed.
where ]1 , ]2 , ]3
, ] 4 np m .
Lr V Ls Lr V Ls Lr 2 JLr In order to solve the problem of slow convergence, the
training data should be normalized into -1~+1.
In order to analyze the reversibility of the system, the 3
output formula should be differentiated until the input variable ANN of N 6,13,2 is constructed, where 3 represent layers of
is visualized firstly. According to formula (2), the following the network, 6 represent the number of the input layer neuron,
formulas can be deduced: 13 represent the number of the hidden layer neuron and 2
y1(1) ] 1 Lm x1 ] 1 x3 represent the number of the output neuron. The active function
° used for hidden neurons is tan-Sigmoid
° y ( 2) ] L [ ] x (n x Lm] 1 x ) x x
f ( x ) ( e e ) /( e e )
x x x
(8)
° 1 1 m 3 1 p 4
x3
2 2
1 m
1
0
« wu wu » » «VL Zr
A( x ) « 1 2
» » « (4)
s
Zr
« wy (2)
wy
(2)
»
] x » « 0
«
2 2
» x 4 3
V L »¼ «¬ 3
¬ wu 1
wu 2 ¼ s
] 1] 4 Lm x /(V Ls ) . When x3 \ r z 0 ,
2
Det( A( x )) 3
Figure 1. The pseudo-linear system
Det( A( x )) z 0 , and A( x , u) is nonsingular. According to the
619
620
wJ ( k )
W1 ( k 1)
T
W1 ( k ) K1
T
(11)
wW1 ( k )
T
wJ ( k )
W2 ( k 1)
T
W2 ( k ) K 2
T
(12)
wW2 ( k )
T
where
wJ ( k )
e( k ) u f (W2 X )
T
(13)
wW1 ( k )
T
W2 R
T
are the connective weight matrix of the hidden-to-
nu m 2
J (10)
2 2 compromise between the fastness for online learning and the
convergence for NN. The bigger learning rate accelerates
where X [1 y y 1
y 1
y
(1)
y y ] . According to the
1
( 2)
2 2
(1)
2
( 2)
online learning, but worsens the convergence of NN; however,
gradient descent method, the online learning algorithm of NNI the smaller learning rate results in a long learning time.
is
IV. EXPERIMENT RESULTS
The plant is an AC squirrel-cage induction motor. The
parameters of motor are as shown in table 1. Schematic
diagram of experiment system with the NNI online learning
control for induction motor is shown in Fig. 4. The current is
measured by Hall sensor, and speed of induction motor could
be obtained through photoelectrical coder. The experiments
are based on DSP TMS320LF2407 produced by TI Co. The
experiment result is shown in Fig. 5. In order to prove the
effect of the decoupling control, we change the reference of
speed and flux respectively. The speed reference steps from 20
rad/s to 60 rad/s at t=8.5s, while flux reference steps from
Figure 3. Block diagram of online learning NNI controller 0.5Wb/m2 to 0.2Wb/m2at t=4s. From Fig. 4, when speed
changes suddenly, there is almost no effect on flux. When the
flux suddenly changes there is also no effect on speed.
620
621
TABLE I. PARAMETERS OF MOTOR is beneficial to the initial stability of the motor system and the
convergence of the NN. The online learning algorithm based
Parameter Value on the idea of basis functions can reduce the computation
burden. In the end, the system is implemented on Matlab-
Pe˄kW˅ 0.5
Simulink and DSP TMS320LF2407. Simulation and
ne˄rpm˅ 1800
experiment results show that the system has good dynamic and
J(kg·m2) 0.006
static properties and excellent characteristic of speed tracking,
Ue˄V˅ 380/220
and the decoupling effect between speed and flux is achieved.
Ie˄A˅ 2.4
Ls˄H˅ 0.51
Lr˄H˅ 0.52 REFERENCES
Lm˄H˅ 0.50 [1] Shinnaka, and Shinji, “New Block Diagrams Using Vector Signals in the
General Frame of Reference for ACM Motors with Stator Core Losses,”
Electical Engineering in Japan, vol. 138, no.2, pp.59-70, 2002.
[2] X. Xia, and W. Gao, Controlling and Decoupling of Nonlinear System.
Beijing: Science Press, 1993.
[3] Z. Zhou, “A Induction Motor De-couple Control Method,” Proceedings
of CSEE, vol.23, no.2, pp.121-125, 2003.
[4] B. Mohamed, and J. Kamel, “A New Stator Resistance Estimation
Method for High Performance Stator-Flux Oriented Sensorless Induction
Motor Drives,” IEEE: Industrial Electronics Conference, pp.311-316,
2002.
[5] B. Karanayil, M. Rahman, and C. Grantham, “Rotor Resistance
Identification Using Artificial Neural Networks for an Indirect Vector
Controlled Induction Motor Drive,” IECON’01: The 27th Annual
Figure 4. Experiment system Conference of the IEEE Industrial Electronics Society, Denver, USA,
2001.
70 [6] C. Kwan, and F. Lewis, “Robust Backstepping Control of Induction
60 Motors Using Neural Networks”, IEEE Trans. on Neural Networks,
vol.11, no.5, pp.1178-1187, 2000.
50
[7] G. Liu, Y. Sun, Y. Shen, and etc. “Dynamic Decoupling Control of
40 Bearingless Switched Reluctance Motors Based on Neural Network
n/(rad/s)
(a) Speed
0.7
0.6
0.5
M/Wb/m2
0.4
0.3
0.2
0.1
0.0
-0.1
0 2 4 6 8 10 12 14
t(s)
(b) Flux
V. CONCLUSION
This paper has presented an online learning NNI
decoupling control scheme of induction motor. The online
learning NNI control is built on the NNI trained offline, which
621
622