DTC induction motor are presented in section Ⅲ. A conclusion is presented in the final section,
section Ⅳ.
2 Neural network-based DTC
DTC scheme includes usually 3/2 transformations of current and voltage, flux estimation, torque
calculation, flux location, flux magnitude computation, flux and torque hystereses comparators,
and optimal switching table. Fig.1 shows a DTC system, which consists of a DTC controller, an
inverter, and an induction motor. Based on DTC principle, the NN-based DTC controller can be
designed as following.
Fig.1 Construction of the DTC system
2.1 Stator-flux observer subnet
2.1.1 Structure of stator-flux observer
Recently, proposed approaches in the literatures to observe stator flux are voltage model, current
model. Although voltage model is easy and precision theoretically, the effectiveness of the
observer depends upon the accuracy of measuring the stator resistance, especially in the low speed
range its value could not be ensured. Current model includes a lot of motor parameters, and these
parameters deeply depend on operating condition, especially in the high-speed range. A novel
stator flux model (named combined model) can be achieved if combines the current model into
voltage model. In high-speed range let the current model be out of function through a low pass
filter, while in low one let the voltage model be out of function through a high pass filter. The
combined model equation is
s s
(
voltage model
)
s
(
current model
)
1
1 1
sT
sT sT
ψ =
ψ + ψ
+ +
(1)
So the stator flux based on Eq.(1) can be
expressed as
( )
ψ
s
= f i
s
, u
s
,
ω (2)
where i
s
——Stator current
u
s
——Stator voltage
ω——Mechanical angle speed
It is not easy to apply Eq.(1) practically because this model is somewhat complex though the
accuracy of observing flux can be obtained. An appropriate neural network can fulfill the
combined nonlinear model. In addition, since there are five inputs, i.e., i
s
α
, i
s
β
, u
s
α
, u
s
β
and
ω to
the stator flux model, and the output flux is the input of computing torque, especially, which are
observed real-timely for selecting voltage vector in the look-table, the stator flux observer is a
dynamic, recurrent and nonlinear system exactly. Actually, a feed-forward neural network with
back-propagation training algorithm is a static system and it needs rebuilding if it is applied in a
dynamic system. The method of rebuilding a multilayer feed-forward neural network with
dynamic characteristic is the output of one layer neurons (hidden layer or output layer) input to
another layer neurons through time-delay unit
[8]
. According to Ref.[9] analysis