
ADAPTIVE INTERNAL MODEL CONTROLLER DESIGN 991
the score matrices; P and Q are the loading matrices of T and U respectively. t
a
and u
a
are the ath orthogonal vectors of T and U, p
a
, q
a
are the ath loading vectors of P and Q.
E* and F* are residual matrices of X and Y, respectively.
In the inner model, an algebraic relationship between the input and output latent
variable is obtained by least squares (LS) method,
u
a
= b
a
t
a
⇒ b
a
= u
T
a
t
a
/t
T
a
t
a
(3)
Finally, the PLS model can be expressed as,
Y − F
∗
=
A
∑
a=1
b
a
t
a
q
T
a
= T BQ
T
(4)
where B = diag(b
1
, b
2
, . . . , b
A
) is a diagonal matrix containing the inner model regression
coefficients. Actually, PLS decouples the multivariate regression problem into a series of
univariate regression problems. The PLS modeling is generally implemented by means of
numerical approach, called NIPALS (non-linear iterative partial least squares) algorithm
which is proposed by H¨oskuldsson [18]. However, with the algebraic structure, the tradi-
tional PLS model can only deal with the static relationship. Therefore, it is necessary to
incorporate dynamics into the PLS model.
2.2. Dynamic partial least squares. In recent years, dynamic PLS modeling that de-
veloped to obtain a better representation of dynamic process is presented in literature
[7-9,12,13,19-21]. Among them, modifying traditional PLS by incorporating the previ-
ous data into each observation vector [19,21] is an intuitive method. However, these
approaches require a substantial increase in the dimension of the X matrix. Kaspar and
Ray proposed another method by including filters before applying the standard PLS al-
gorithm [7,8]. However, the dynamic filter is designed either based on a-prior knowledge
or by minimizing the sum of squares of the output residuals. Lakshminarayanan et al.
proposed a dynamic extension of the PLS algorithm that incorporating the dynamic re-
gression relationship, like ARX, into the PLS inner model [9]. This method modified the
standard PLS algorithm. Chen et al. proposed a dynamic PLS method by combining the
standard PLS model and ARX model [12,13]. The standard PLS model can decouple the
MIMO system into several SISO subsystems and pair the loops automatically. While the
dynamic characteristics of the system can be inferred from the analysis of an ARX model
fitted to the observations of the system [13]. The model structure is shown in Figure 1.
The input x(t) is scaled and centered by matrix W x to be the input vector of standard
PLS. Using PLS model, one can obtain the output vector of PLS, y
P LS
(t). G
p
represents
the process plant. The output y(t) is scaled and centered by matrix W y, then passes the
delay operators, q
−1
, q
−2
, . . . , q
−Na
, to be the input vector of ARX part. The prediction
output ˆy(t + 1) is expressed as a weighted sum of the past outputs plus the outputs from
the PLS model.
This dynamic PLS structure can be formulated as two parts,
y
P LS
(t) = P LS(x(t)) (5)
ˆy(t + 1) =
Na
∑
j=1
A
j
y(t + 1 − j) + y
P LS
(t − d) (6)
where N
a
is the number of lagged, A
j
is the ARX model parameters and d is the pure
delay.
The procedure of the dynamic PLS modeling is presented as follows.