ω =
( )
( ) ( )
( )
( )
( )
( )( )
s ss s ss s ss s ss
s s s s s s s s s s s s
L i L i L i L i
L i L i L i L i
α α β β β β α α
α α α α β β β β
ψ ψ σ ψ ψ σ
ψ ψ σ ψ ψ σ
− − − − −
− − + − −
� � � �
� � � �
(3)
where
2
σ = 1− L
m
L
s
L
r
i
s
α
,i
s
β
—— The stator current components
ψ
s
α
,
ψ
s
β
—— The stator flux components
L
s
, L
r
, L
m
—— The stator inductance, rotor inductance
and mutual inductance
It can be seen that, in the DTC system, NN-based flux observer can compute out an accurate stator
flux value without ω, because the speed information is embedded in relative stator contents. Fig.2
is the training scheme of the NN-based stator flux observer. Where f, e(k) are the addition
distortion and error, respectively. It should be noted that a random distortion plus on the time-
delayed ψ
s
, so that the control range is widen and stability is strengthened.
Fig.3 is the structure of the BP-based flux observer.
Fig.2 Schematic diagram of NN-based flux observer
Fig.3 Structure of the BP-based flux observer
2.1.2 The modified BP algorithm
According to analysis above, the BP-based supervised network is used to implement the stator-
flux observer. The input/output pairs for training are produced by the combined model, input
samples: u
s
α
, u
s
β
, i
s
α
, i
s
β
, u
s
α
(k
−1), u
s
β
(k
−1), i
s
α
(k
−1) ,i
s
β
(k
−1) ψ
s
α
(k
−1) and ψ
s
β
(k
−1); output
samples: ψ
s
α
and ψ
s
β
. It is well known that slower convergence and longer training times are the
disadvantages of the conventional BP algorithm compared with other competing techniques. In
addition, in the conventional BP algorithm, the learning rate is fixed. This paper applies a
modification of the original BP algorithm which is a combination of weight update and learning
rate update. Corresponding to the input vector X
p
in the training set, we consider the Eq. (4)
measuring the error of the function approximation network.