This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS
a few reported studies focusing on performance assessment
of ILC. Chen and Kong [21] used MVC as the benchmark to
assess the control performance of ILC, which appeared to be
the first reported study on CPA of ILC. Under the assump-
tion that each ILC controller influences either stochastic or
deterministic performance, their method estimated a mini-
mum variance bound to assess the performance. However, the
above-mentioned assumption neglects the interactions between
stochastic and deterministic performance, which exist com-
monly in practice. Also based on the MVC benchmark,
Farasat and Huang [2] proposed a new method to assess ILC’s
performance, which introduces a tradeoff between determinis-
tic and stochastic performance described by a tradeoff curve.
Because MVC does not consider the control moves and has
poor robustness, the MVC benchmark seems to be undesirable
and unachievable in many practical applications. In addi-
tion, the systems considered in both of the above-mentioned
studies are described by transfer function, which is gener-
ally inapposite to deal with multi-input multi-output (MIMO)
systems.
In this paper, the authors suggested a new performance
assessment method for ILC based on a 2-D Roesser model that
is transformed from the ILC-controlled batch process model.
Moreover, this paper proposed a novel 2-D LQG benchmark
for the transformed 2-D system and first introduced a con-
cept named as “performance assessment tradeoff surface.” The
proposed method can be implemented for both single-input
single-output (SISO) and MIMO systems. To design the 2-D
LQG benchmark controller, the system model is required.
When the model is unknown, system identification should
be done before assessing control performance. A subspace
system identification algorithm was recently developed in [22]
for ILC, specifically on a gantry robot. Their identification
algorithm performed effectively, but there is a very rigorous
assumption that the initial state of the following trial equals
to the final state of the preceding trial. However, this assump-
tion generally does not hold for batch processes. Because the
ILC-controlled batch process is transformed to a 2-D system,
a novel subspace identification method is proposed based on
2-D system in a unified framework in this paper.
In recent decades, the identification of 2-D systems has
attracted increasingly interests, and there are considerable
studies in the emerging field of 2-D control systems [23], [24],
2-D signal processing [25], and 2-D recursive filters [26].
Because 2-D systems have two coupled state variables: 1) hor-
izontal state and 2) vertical state, the analysis and synthesis of
2-D systems are more complex and challenging than those of
1-D systems [27]. Consequently, it is difficult to extend reg-
ular system identification methods from 1-D systems to 2-D
systems. To overcome the 2-D system identification problem
in state space, a special type of 2-D system structure—causal,
recursive, and separable-in-denominator (CRSD) system—
in the Roesser form, was considered in the past decade.
Fortunately, the converted 2-D Roesser model for the ILC-
controlled batch process is CRSD. To identify MIMO systems,
subspace method is a very powerful tool with good numerical
reliability and modest computational complexity. Its calcu-
lation is relatively simple without complicated iterations.
In addition, its calculated results are in the state space form,
which is very convenient for estimation, filtering, prediction,
and control. Therefore, 2-D subspace method is applied to
identify the converted 2-D CRSD model.
Based on traditional 1-D Hankel approximation methods,
Lashgari et al. [28] introduced a method to approximate a gen-
eral 2-D filter with a CRSD model. Hinamoto et al. [29]
suggested a method to realize a CRSD model directly from
2-D impulse response data. Nam et al. [30] used genetic algo-
rithms to identify a CRSD model in the frequency domain by
simultaneously approximating both magnitude and phase char-
acteristics. Ramos [31] first applied the subspace method to
identify 2-D CRSD model using only input–output data. Since
then, four standard subspace-based 2-D CRSD model identifi-
cation algorithms have been developed by Ramos et al. [32].
However, to the authors’ best knowledge, all existing stud-
ies are concerned with subspace identification of open-loop
2-D CRSD systems. In the standard open-loop subspace-based
algorithms (e.g., N4SID, MOESP, and CVA), the fundamental
assumption that there is no correlation between input signals
and unknown noises, which does not hold in the closed-loop
system case [33]–[35]. In other words, the existing subspace
identification methods cannot be directly applied to closed-
loop 2-D CRSD system. As an increasing number of industrial
processes are in closed-loop, the problem of closed-loop 2-D
CRSD system identification needs to be addressed [36].
This paper tackled the problem of identifying unknown
closed-loop 2-D CRSD systems in the converted Roesser form.
Using the standard subspace identification algorithm and prin-
cipal component analysis [37], consistent estimations of the
2-D CRSD state space model can be achieved. Based on the
identified model, the performance assessment for ILC can be
continued.
This paper can be considered an integration and improve-
ment of our two conference papers [38], [39]. In [38], it is
assumed that the system model is known, which is generally
unsubstantial in practice. In [39], the model-unknown situ-
ation has been studied. Furthermore, compared with these
two conference papers, this paper has the following improve-
ments: first, the new benchmark is a real 2-D controller,
but the previous benchmarks in conference papers [38], [39]
are in fact 1-D controllers; second, more details about the
proposed closed-loop subspace identification method have
been presented; in addition, more simulation tests have
been done.
In conclusion, this paper has the following contributions.
First, this paper proposed a novel framework to assess the
control performance of ILC by using 2-D system theory
and correspondingly introduced a novel concept—performance
assessment tradeoff surface. The proposed framework is
a powerful tool to deal with MIMO batch processes. Second,
a novel 2-D LQG benchmark control is designed for the trans-
formed 2-D system. Unlike the corresponding 1-D problem,
it is generally very difficult to express the 2-D linear-
quadratic (LQ) optimal control in terms of a state feedback.
Ntogramatzidis and Cantoni [40] first proposed an LQ opti-
mal control via semistate feedback for Roesser models over
a 2-D signal-index set of finite extent. We extended this