Published: October 12, 2011
r
2011 American Chemical Society
13427 dx.doi.org/10.1021/ie200021t
|
Ind. Eng. Chem. Res. 2011, 50, 13427–13434
Optimal Structure of Learning-Type Set-Point in Various Set-Point-
Related Indirect ILC Algorithms
Youqing Wang,
†,
* Jianyong Tuo,
†
Zhong Zhao,
†
and Furong Gao
‡,†,
*
†
College of Information Science and Technology, Beijing University of Chemical Technology, Beijing, China
‡
Department of Chemical and Biomolecular Engineering, Hong Kong University of Science and Technol ogy, Clear Water Bay,
Kowloon, Hong Kong
ABSTRACT: According to the literature statistics, only less than 10% of reported iterative learning control (ILC) methods have
been devoted to the indirect approach. Motivated by the full potential of research opportunities in this field, a number of studies on
indirect ILC were proposed recently, where ILC-based P-type control and learning-type model predictive control (L-MPC) are two
successful stories. All indirect ILC algorithms consist of two loops: an ILC in the oute r loop and a local controller in the inner loop.
The local controllers are, respectively, a P-type controller in the ILC-based P-type control and a model predictive control (MPC) in
the L-M PC. Logically, this leads to the question of what type of ILC should be chosen respectively for the two above-mentioned
indirect ILC methods. In this study, P-type ILC and anticipatory P-type (A-P-type) ILC are studied and compared, because they are
typical and widely implemented. Based on mathe matical analysis and simulation test, it has been proved that the A-P-type ILC
should be used in the ILC-based P-type control and while the P-type ILC should be used in the L-MPC. Furthermore, an improved
L-MPC with batch-varying learning gain was proposed to handle the trade-off between convergence rate and robustness
performance. The simulation results on injection molding process and a nonlinear batch process validated the feasibility and
effectiveness of the proposed algorithm.
1. INTRODUCTION
Intelligent machines including industrial computers can be led
from repeated training to reach superior performance of a speci-
fied task. Scholars and engineers have there fore developed
iterative learning control (ILC) methods to formulate the learn-
ing procedure systematically. ILC was first presented in Japanese
in 1978
1
and then was introduced in Englis h in 1984.
2
These
contributions are widely considered to be the origins of ILC. Up
to the present, ILC has become an ad hoc research topic and
numbers of publications were achieved every year.
39
After three
decades’ development, ILC has been successfully implemented
in industrial manipulators,
4,10,11
chemical batch processes,
1214
bio-
medical processes,
7,15,16
and o ther processes.
5,17,18
There are mainly two application modes for ILC. First, ILC is
used to determine the control signal directly, and this kind of ILC
is named direct ILC. Second, there is a local feedback controller
in each batch and ILC is used to update some parameter settings
of the local controller, so this kind is named indirect I LC.
Compared with direct form, indirect ILC has some advantages.
First, the existing p rocess structure need no change; if there
already exists a controller, only an I LC module is added in the
outer loop to update some parameters of the e xisting controller
and this ILC module could be easily moved at any time. Second,
in most cases, indirect ILC has better robustness than the direct
one; this is because direct ILC must have a feedforward term,
which is sensitive to variations in batch direction, but a
feedforward term is not necessary for the local controller of
the indirect ILC. In addition, the idea of the indirect method is
consistent with the d eveloping trend of control engineering:
stability and robustness are not the only requirements f or
control design, and an optimization scheme should be uti lized
to improve the closed-loop control performance. According to
the literature statistics in a recent survey,
19
however, only less
than 10% of the reported ILC methods were implemented in
the indirect mode.
The following two issues are essential for an indirect ILC: what
algorithm is used to design the local controller, and which param-
eters of the local controller are adjusted by the ILC. Generally
speaking, ILC could be used to adjust set-point,
20
control gain,
21
weight,
22
and other parameters
23,24
for the local controller. Par-
ticularly, an indirect ILC that updates the set-point for the local
controller is termed as set-point-related (SPR) indirect ILC. Less
than 20% of reported indirect ILC algorithms are SPR indirect
ILC. For clarity, the block diagram of SPR indirect ILC was
shown in Figure 1.
Motivated by the full potential of research opportunities in this
field, several SPR indirect ILC methods were proposed recently.
7,25
In the literature,
7
a novel combination of ILC and model pre-
dictive control (MPC), termed L-MPC, was proposed, where the
local controller is MPC. It is valuable to point out that ILC and
MPC have long been used together, in combinations such as
BMPC,
26
2D-GPILC,
27
and MPILC.
28
However, in each of these,
MPC was used to design the updating law of ILC; therefore,
these combinations belong to the direct ILC category. To our
best knowledge, ref 7 is the first reported wo rk on ILC-based
MPC, or L-MPC. In the L-MPC framework, the set-point y
r
(t,k)
for MPC could be different in various batches, and it is updated
Received: September 8, 2010
Accepted: October 12, 2011
Revised: October 6, 2011