International Journal of Control, Automation and Systems 19(2) (2021) 777-787 ISSN:1598-6446 eISSN:2005-4092

http://dx.doi.org/10.1007/s12555-019-1033-1 http://www.springer.com/12555

Robust Optimal PID type ILC for Linear Batch Process

Furqan Memon and Cheng Shao* 

Abstract: The proportional-integral-derivative (PID) controller is standard technique for controlling the industrial
batch process. However, the parameter tuning and updating of PID controller has been a challenging topic for
control engineers especially for the real system having initial state error, state and output disturbances. In this
paper, a kind of robust optimal iterative learning control (ILC) scheme is suggested to update the PID gains for the
linear system with initial state error, state and output disturbances. The quadratic performance criteria is considered,
which constitute of the error and input slew rate, and the robust BIBO stability is investigated theoretically for the
proposed PID type ILC scheme. In addition, the bound of the tracking error has been calculated by using Lyapunov
composite energy function. Simulation examples are also given to demonstrate the effectiveness of the proposed
scheme in term of its ability to deal with linear as well as nonlinear system

Keywords: Discrete-time system, iterative learning control (ILC), parameter optimal ILC, PID iterative learning
control.

1. INTRODUCTION searchers have recommended many techniques for attain-

ing perfect tracking in both continuous- and discrete-time
Due to the PID controller’s better good performance, domains [6,7]. The efficacy of ILC has been demonstrated
simple structure, and robustness, it is still considered as a in literature with the applications in the field of robotics,
standard tool to solve the automatic control problem of in- mechatronics, manufacturing, building control, nuclear
dustrial processes. However, the better performance of the fusion, rehabilitation, and in network control [8–16].
PID controller depends strongly on the proper choice of ILC consists of many varieties, which have been de-
the PID parameters. Some efforts are made to update PID veloped to expedite the convergence rate of tracking er-
parameters, such as fuzzy logic [1], neural network [2], ror such as optimal ILC [9], [11] parameter optimization
adaptive [3] and etc. For batch/repetitive processes, itera- [17, 18], Model Predictive Control ILC [19, 20] and oth-
tive learning control (ILC) is known as an effective control ers. As known, the major challenge to ILC is to acquire
strategy to optimize control performance from iteration to error convergence iteratively even if there is uncertainty in
iteration. Hence, it is motivated to apply ILC schemes to the plant model. The robustness of the various ILC tech-
update PID parameters for the batch processes. ILC is an niques against model uncertainty has been studied in the
effective control technique to realize perfect tracking for literature, such as inverse model-based ILC [21], norm-
the batch process. optimal ILC [22] and two dimensional learning systems
The main principle of ILC is to use the error informa- [23, 24]. The robustness against model uncertanity in ILC
tion from the previous iteration to enhance the input or design has also been investigated in frequency domain
learning gains for the subsequent trial, such that perfect [25, 26]. There are some ILC approaches that explicitly
or bounded tracking of the reference trajectory can be at- considered model uncertainty to improve robust conver-
tained. ILC method for computing such input was first ad- gence and performance, such as author in [27] used the
dressed in 1978 by Uchiyama [4] and later mathematically convex optimization methods to calculate the LMI dealing
formulated by Arimoto et al. in 1984 [5]. Afterward, sig- with time-varying model uncertainty. In [28], the author
nificant research efforts have been devoted to the ILC de- investigates the use of time-varying filtering and [29] de-
sign to deal with actual batch process, for the improve- velops an averaging technique over uncertain models for
ment of tracking performance and convergence rate. In or- a robustly converging algorithm. Recently, a robust worst-
der to deal with linear and non-linear batch processes, re- case norm-optimal ILC technique [30] is presented, which

Manuscript received December 11, 2019; revised May 6, 2020; accepted June 2, 2020. Recommended by Associate Editor Changsun Ahn
under the direction of Editor Kyoung Kwan Ahn. This work is partly financially supported by the High-tech Research and Development
Program of China (2014AA041802). The first author is also thankful to the China Scholarship Council (CSC) for providing financial support
for his Ph.D. studies at Dalian University of Technology, China.

Furqan Memon and Cheng Shao are with the Institute of Advanced Control, Dalian University of Technology, Dalian 116024, Liaoning,
China (e-mails: [email protected], [email protected]).
* Corresponding author.

c ICROS, KIEE and Springer 2021

778 Furqan Memon and Cheng Shao

is solved by formulating as a convex optimization prob- theoretically that using proposed PID type ILC, the er-
lem. However, the error convergence is also affected by ror converges to zero in the presence of repeatable initial
the disturbance and initial state error. In most of the ex- states and repeatable external distrubances, and to a finite
isting literature, the initial state and disturbances are as- boundary in the presence of non-repeatible initial errors
sumed same and repetitive with respect to iteration re- and external disturbances. Furthermore, the bound of the
spectively. These assumptions are impractical while deal- error for the iterative-varying distrubances has been calcu-
ing with the real system, because these factors make the lated by using the lyapunov funtion for the proposed tun-
problem more complex. ing algorithm. In simulation, it can be observed that the
In order to achieve robustness against uncertainty, dis- proposed approach is effective not only in dealing with
turbances and initial state error, PID types of ILC are pre- linear systems, but also nonlinear system in the presence
ferred because of its effectiveness, simple structure, ro- of model uncertainty, non-repetitive disturbance and ini-
bustness and the merits of each control actions (propor- tial errors. The paper is organized as follows: Section 2
tional, integral and derivative) of PID in the ILC law. gives the problem fomulation and robust optimal PID iter-
Specifically, the convergent rate and robustness against ative learning control for the discrete-time linear systems
uncertainty on PD, PI, and PID types of ILC are also is discussed in Section 3. The BIBO stability analysis of
analyzed in the literature. Many researchers have recom- the proposed ILC law are presented in Section 4. Simu-
mended different online tuning methods for evaluating lation results are discussed in Section 5. Finally; conclu-
the PID gain values in the continuous-time domain, such sions are drawn in Section 6.
as least square method [31] and also intelligent tuning
method [32]. To achieve the monotonic convergence for 2. PROBLEM FORMATION
the discrete-time linear batch process, the author has pro-
posed an offline method for designing the optimal gains The plant is assumed to be described by the following
of PID [33], and extended PID [34] types of ILC. To deal sampled-time linear system with state and output distur-
with time-varying uncertainty, robustness of the PID type bances,
of ILC is analyzed using linear matrix inequality [35] in
xk (t + 1) = Axk (t) + Buk (t) + vk (t),
two-dimensional (2D) theory. To take the advantage of
the PID type ILC approach, the robustness against ini- yk (t) = xk (t) + wk (t), (1)
tial state, uncertainty and disturbance for the linear and
where xk (t)εRn , uk (t)εRm and yk (t)εRm are the state, con-
time delay systems have been discussed in [36–38]. The
trol input, and output of the system, respectively; AεRn×n ,
PD and PI types of ILC have been considered [39, 40],
BεRn×m and CεRm×n are constant matrices. Also, k repre-
which can guaranteed the monotonic convergence. More-
sents the iteration number, and and are the bounded exter-
over, other PID types of ILC can be found in the litera-
nal disturbances respectively. Regarding the plant (1) the
ture [39–42]. Recently, another effective framework uti-
following assumption is made.
lizes previous trial error information to formulate N para-
Assumption 1: The matrix pair (A, B) and (A, C) are
metric PID ILC scheme to improve the convergence rate,
assumed to be controllable and observable respectively,
as suggested in [43]. The aforementioned methods solved
and CB 6= 0.
the problem of automatic tuning parameters to a certain
Remark 1: Assumption1 means that the nominal model
extent, however, mostly existing methods for finding the
of the plant (1) is controllable and observable. In addition,
PID gains are based on offline tuning and ad hoc tuning
if CB= 0, then a regularizing procedure can be performed
methods. In batch process control, deriving the online tun-
as detailed in [9].
ing mechanism for PID controller with robustness against
Based on linear system theory, the solution to the plant
distrubances and initial states is a significant task by the
(1) can be formulated as
method of updating PID gains at each trial based on the
performance of the previous iteration. t−1
yk (t) = ∑ CAt−l−1 Buk (l) + ηk (t), (2)
Motivated by the aforementioned discussion, an opti-
l=1
mal design method for updating PID gains with applica- t−1
tion to linear batch process is proposed for the linear sys- ηk (t) = CAt xk (0) + ∑ CAt−l−1 vk (l) + wk (l). (3)
tem with initial state error, state and output disturbances. l=1
The proposed algorithm takes into account actual error in- It can be observed that initial actions of each trial and
formation of the previous iteration, and then PID gains disturbances appear in the batch process. The output (2)
are tuned iteratively. The updating mechanism for PID can be represented in super vector yk , uk and ηk for N
type ILC depends on the nominal model and error in- time samples from uk (t) , yk (t) and ηk (t) for each trial as
formation from the previous cycle, where the error in- below
formation is feedbacked to tune PID gains online and is
robust against disturbances and initial error. It is proven yk = Guk + ηk , (4)
 Robust Optimal PID type ILC for Linear Batch Process 779

 
where

1 0 ··· 0
 1 0 ··· 0
. .. 
1 ..
 −1 1 · · · 0

1 .
      
uk (0) yk (1) ηk (1) F2 =  . .  , F1 =  ..  ,

 .. .. .. . .. ..

 uk (1)
  yk (2)   ηk (2)  . 0  .. . . .
     
uk = 
 uk (2)
,

yk =   , ηk =  ηk (3)  ,
 yk (3)    0 ··· −1 1 1 1 ··· 1
..  ..   .. 
(10)
 
 .
  .   . 
uk (p − 1) yk (p) ηk (p)
  with
CB 0 ··· 0
.  KF = Kp I + Ki F1 + Kd F2 . (11)
· · · .. 

 CAB CB
G= .. .. ..  .

 .. Using (7) and (9), one can have
 . . . . 
CAN−1 B CAN−2 B · · · CB
êk = êk−1 − GKF êk−1 ,
The matrix G, a lower-triangular block matrix known as êk = (I − GKF )êk−1 , (12)
a Toeplitz matrix, can be determined from its first column,
and is full rank under Assumption1. In the absence of the where uk (t)εRN∗1 and ek εRN∗1 are the input and error at
disturbances and initial state errors, ηk = 0, the nominal iteration k, respectively. It is observed from (12) that the
system error for the reference trajectory yd , (or desired PID controller (9) for nominal system (6) can be designed
system output) can be given as to ensure Lyapunov stability with zero tracking error if and
only if the parameters of PID controller (9) are updated by
êk = yd + ŷk , (5) solving the following inequality
ŷk = Gûk , (6) k(I − GKF )k < 1. (13)
where ûk are ŷk nominal input and output of the plant (1). However, the case of initial state error, state and output
Then the nominal tracking error for two consecutive iter- disturbances will be considered in this paper, which means
ations with reference signal yd is given below the PID controller (9) will be applied to the plant (2). The
main contribution in this paper is to present a new method
êk = êk−1 − G∆ûk , (7) of the parameter updating of PID controller by means of
ILC mechanism.
with
3. ROBUST OPTIMAL PID TYPE ILC
   
ûk (0) − û0 êk (1)
 ûk (1) − û1  êk (2) 

ûk (2) − û2
   The parameter updating of PID controller will be made
∆ûk =  , êk =  êk (3)  .
   
 ..   ..  by an optimal ILC of the minimization of the following
 .   .  cost function with respect to Ks.
ûk (N − 1) − ûN−1 êk (N)
1 T
êk Qêk + ∆ûTk R∆ûk ,

min Jk (t) = (14)
In this paper, a new iterative learning online tuning Kp ,Ki ,Kd 2
method for PID gains is suggested for the batch process,
where Q and R are arbitrary symmetric positive definitive
which can take into consideration state and external distur-
weighting matrices given by the designer. Using (9) and
bance. First the proposed iterative learning PID controller
(12), the cost function (14) can be rewritten as follows:
for nominal system will be discussed which constitute of
the following scheme [33]. 1 T
Jk (t) = ê (I − GKF )T Q(I − GKF )êk−1
t+1
2 k−1
+ êTk−1 KFT RKF êk−1

ûk (t) − ûk−1 (t) =Kp êk−1 (t + 1) + Ki ∑ êk−1 ( j)
j=1 1
= êTk−1 Q − 2KFT GT Q

+ Kd [êk−1 (t + 1) − êk−1 (t)], (8) 2
+ KF (GT QG + R)KF êk−1 .

(15)
where Ks are constant with respect to time and trail axis,
and will be designed in sequal. The structure of the PID Taking partial derivative with respect to Kp , Ki and Kd
control law (8) can be rewritten by the super vectors as respectively, and solving ∂∂KJkp = 0, ∂∂ KJki = 0, ∂∂KJkd = 0 gives
follows. the following three equations:

∆ûk = (Kp I + Ki F1 + Kd F2 )êk−1 , (9) − êTk−1 QGêk−1 + êTk−1 (GT QG + R)

780 Furqan Memon and Cheng Shao

× (Kp I + Ki F1 + Kd F2 )êk−1 = 0,
− êTk−1 F1T GQêk−1 + êTk−1 F1T (GT QG + R)
× (Kp I + Ki F1 + Kd F2 )êk−1 = 0,
− êTk−1 F2T GQêk−1 + êTk−1 F2T (GT QG + R)
× (Kp I + Ki F1 + Kd F2 )êk−1 = 0. (16)

Solving above three equations gives

K̂Fk = M̂k−1 L̂k , (17) Fig. 1. Block diagram for the Robust PID Type ILC.
with
with
   
K̂p m̂1 m̂2 m̂3
K̂Fk =  K̂i  , M̂k =  m̂T2 m̂4 m̂5  ,    T 
m1 m2 m3 ek−1 QGek−1
K̂d m̂T3 m̂T5 m̂6
 T  Mk =  mT2 m4 m5  , Lk =  eTk−1 QGFk−1  ,
êk−1 QGêk−1 mT3 mT5 m6 eTk−1 QGF2 ek−1
Lk =  êTk−1 QGFk−1  ,
êTk−1 QGF2 êk−1 where

m1 = eTk−1 GT QG + R ek−1 ,


where
m2 = eTk−1 GT QG + R F1 ek−1 ,


m̂1 = êTk−1 GT QG + R êk−1 ,

m3 = eTk−1 GT QG + R F2 ek−1 ,


m̂2 = êTk−1 GT QG + R F1 êk−1 ,

m4 = eTk−1 F1T GT QG + R F1 ek−1 ,

m̂3 = êTk−1 GT QG + R F2 êk−1 ,

m5 = eTk−1 F1T GT QG + R F2 ek−1 ,

m̂4 = êTk−1 F1T GT QG + R F1 êk−1 ,

m6 = eTk−1 F2T GT QG + R F2 ek−1 .

m̂5 = êTk−1 F1T GT QG + R F2 êk−1 ,

m̂6 = êTk−1 F2T GT QG + R F2 êk−1 .

The above design parameters will be computed online
for updating gains of the PID type ILC. Specially, when
Then the parameter updating of PID controller can be
the tracking error equals to zero, then the PID paramters
made by an optimal ILC procedure as (17) and a nominal
are updated better and be kept unchangeable. So the al-
control input can be obtained iteratively for nominal plant
gorithm will work only when there are non-zero tracking
(6) as follows.
errors. The block diagram for the proposed PID type ILC
is shown in Fig. 1.The following convergence for the pro-
∆ûk =(K̂p I + K̂i F1 + K̂d F2 )ek−1 posed approach is given in sequel.

=K̂Fk êk−1 . (18)

4. CONVERGENCE ANALYSIS
This PID-ILC tuning mechanism is also optimal when
applied to (4) where ηk = 0, or the plant (1) for the The robust BIBO stability of the proposed PID type ILC
disturbance-free case. Now the nominal PID controller algorithm has been investigated and moreover, the bound
will be extended to (1) with uncertainties of initial state of the tracking error has been estimated in this section.
error, state and output disturbances. This will be made
by replacing the nominal output error in (16)- (18) with 4.1. BIBO stability anaysis
the actual tracking errors ek resulting from the plant (1). The robust BIBO (bounded-input-bounded-output) sta-
Therefore, an ILC of PID type algorithm of interest can bility as defined below together with convergence will be
be summarized as follows: discussed as follows.
Definition 1: An iterative learning control system is
∆uk = (Kp I + Ki F1 + Kd F2 )ek−1 said to be robust BIBO (bounded-input-bounded-output)
= KFk ek−1 , (19) if, and only if, for a system with any uncertain initializa-
tion and bounded disturbance, the outputs are bounded for
ek = yd − yk . (20)
each trial for arbitrary bounded control reference.
PID gains can be calculated iteratively by the following: The above defnition considers external disturbances to
the closed loop system and the initialization uncertainty
[Kp , Ki , Kd ]T = Mk−1 Lk , (21) along trial axis. The main result is given as follows:
 Robust Optimal PID type ILC for Linear Batch Process 781

Theorem 1 (Robust BIBO Stability): The application 4.2. Calculation of the error bound
of the iterative learning control algorithm of (19), (20) and To calculate the bound of error ek and to observe the
(21) to the plant (1) with uncertainties of bounded initial effect of weighting matrix of the error bound, let the fol-
state error, state and output disturbances, is robust BIBO lowing candidate of the Lyapunov function
stable if there exists a positive constant 0 < σ < 1 such
that for each trail. Vk = eTk Qek ,

k(I − GKFk )k < σ , (22) where Q is positive definite and symmetric matrices de-
fined in (14), the necessary and sufficient condition for the
k(I − KFk G)k < σ. (23)
error to converge to ∆Vk is ensured to be negative. Then it
Proof: It follows from the plant (4) that the closed loop follows from (24) and above equation, one can have
system tracking error can be written by
∆Vk =eTk Qek − eTk−1 Qek−1
ek = ek−1 − G∆uk − ∆ηk
= (ek−1 − GKF ek−1 )T Q (ek−1 − GKF ek−1 )
= ek−1 − G∆uk − ∆ηk , (24)
− 2∆ηkT Q (I − GKF ) ek−1 + ∆ηkT Q∆ηk
which gives = −2ek−1 (GKFk )T Qek−1 −2∆ηkT Q(I−GKF )ek−1
k k k
T
+ ek−1 GKFk Q GKFk ek−1 + ∆ηkT Q∆ηk .


ek = ∏(I − GKFi )e0 − ∑ ∏ (I − GKFi )∆η j . (25)
i=1 j=1 i= j+1 (30)
If the condition (22) is satisfied, then Using the proposed PID gains formula (21) and comparing
k with (16) one can derive the following:
kek k ≤ σ k ke0 k + ∑ σ k− j k∆η j k. (26)
j=1 − eTk−1 QGek−1 + eTk−1 (GT QG + R)KFk ek−1 = 0,
Therefore, for uncertainties of bounded initial state er- − eTk−1 QGF1 ek−1 +eTk−1 F1T (GT QG+ R)KFk ek−1 = 0,
ror, state and output disturbances, the closed loop system − eTk−1 QF2 ek−1 + eTk−1 F2T (GT QG + R)KFk ek−1 = 0.
tracking error is bounded. Specially, in the absence of un- (31)
certainties, i.e., ηk = 0 , zero tracking error will be derived.
In addition, it follows from control law (19) that Pre-multiplying (32) by Kpk , Kik and Kdk respectively, and
then adding them together give
uk =uk−1 + KFk ek−1
T
T
=uk−1 + KFk (yd − yk−1 ) ek−1 GKFk ek−1 = ek−1 KFk (GT QG + R)KkF ek−1 .
(32)
=(I − KFk G)uk−1 + KFk (yd − ηk−1 ), (27)
Substituting (33) into (31), one can have
which gives

T
k ∆V = − ek−1 KFk (GT QG + 2R)KFk ek−1
uk = ∏(I − KFi G)u0 − 2∆ηkT Q (I − GKF ) ek−1 + ∆ηkT Q∆ηk . (33)
i=1
k k For the Disturbance-free and no initial errors, ηk = 0,
−∑ ∏ (I − KFi G)KFi (yd − η j−1 ). (28) It follows from Lyapunov stability theorem that ∆Vk is
j=1 i= j+1
negative provided that(KFk )T (GT QG + 2R)KFk is the pos-
If the condition (23) is satisfied, then itive definite. Because the weighting matrices Q and R can
k be chosen such that (GT QG + 2R) is positive definite, so
kuk k = σ k ku0 k + ∑ σ k− j kKFi kkyd − η j−1 k j . (29) the condition on (KFk )T (GT QG + 2R)KFk being the posi-
j=1 tive definite is equivalent to the matrix is full rank. This is
The boundedness of KFk can be easily proven from its def- ideal case, as real system has varying initial state, distur-
inition and is omitted here. Thus, the proof is completed. bance and noise with respect to iteration. To deal with real
Remark 2: In [33], it has been proved that the con- system, the following bound are considered as
dition kλgc < 1k is the necessary and sufficient condi- kxk (0) − xk−1 (0)k ≤ gx ,
tion to achieve the condition (22) and (23), where λgc =
1 −CB(Kpk + Kik + Kdk ) is the eigen values of the matrices. kwk (0) − wk−1 (0)k ≤ gw ,
Remark 3: It can be observed from (25) that for the kvk (0) − vk−1 (0)k ≤ gv .
disturbance-free and no initial errors, ηk = 0, the closed
Based on the above-mentioned bounds, it can be stated
loop tracking error is zero in the presence of repeatable
that
disturbances and initial errors, whereas is bounded in the
presence of nonrepeatable disturbances and initial errors. kηk (0) − ηk−1 (0)k ≤ gη .
782 Furqan Memon and Cheng Shao

Since 5. SIMULATION RESULT

k T
(GT QG + 2R)KFk ek−1


ek−1 KF In order to illustrate the applicability and effectiveness

T
≤ λmax ( KFk (GT QG + 2R)(KFk )kek−1 k2 , of our proposed method, three simulation examples are
discussed in this section. In the first example, a compari-
∆ηkT Q∆ηk ≤ λmax g2η . (34) son of the proposed technique with respect to techniques
Also, the second term of the right side of the (33) is given mentioned in [33] is drawn. Also, the proposed method
as has the ability to deal with the nonlinear system, which is
shown in the second example. Finally, the third example is
− 2∆ηkT Q I − GKkF ek−1


given to demontrate the effectiveness of the proposed al-
≤ λmax (QT Q)g2η /ε gorithm to deal with model uncertainty with unstructured

T model.
+ ελmax I − GKkF I − GKkF kek−1 k2 .


(35) Remark 4: The error is constituted of the vector, whose
norms have been calculated by using formulae of the vec-
Using above-mention inequalities (34) and (35), (33) is
tor norms. As Matlab software has been used to simulate
bounded by
the proposed method, norms of the are calculated using

T
∆Vk ≤ − λmax ( KFk (GT QG + 2R)(KFk )kek−1 k2 the built-in commands in Matlab.

T Example 1: In this example, we consider the DC motor,
+ ελmax I − GKkF I − GKkF kek−1 k2


in which the rotational angle of the motor is controlled us-
1 ing field winding. The details of the model and desire out-
+ λmax (QT Q)g2η + λmax (Q)g2η . (36)
ε put trajectory are given in [24]. The necessary important
Suppose information of the state space model is mentioned hereby
(for details refer to [24]).
λGK = λmax (I − GKkF )T (I − GKkF ) ,
 
xk (t + 1) = AD xk (t) + BD uk (t) + vk (t),
λGQRK = λmax (KFT (GT QG + 2R)KF ),
yk (t) = CD xk (t) + wk (t),
λQ = λmax (Q), λQQ = λmax (QT Q).
t = 1, 2, 3, . . . , 1200, and k = 1, 2, 3 . . . ,
Using above notations, (36) is given as
where xk (t) = [i f (t) ω(t) θ (t)], yk (t) = θ (t), uk (t)
∆Vk ≤ −(λGQRK − ελGK )kek−1 k + gGQKη , (37) = field winding voltage, i f (t) = field winding current,
ω(t) = angular velocity, and θ (t) = angle of the DC mo-
where
  tor.
1
λQQ + λQ g2η .
 R 
gGQKη = − L ff 0 0
ε
AD =  km − f 0  ,
 
Lyapunov function can be given as J J
0 1 0
∆Vk = eTk Qek − eTk−1 Qek−1 1
Lf  
≤ −(λGQRK − ελGK )kek−1 k + gGKη , BD =  0  , CD = 0 0 1 ,
Vk −Vk−1 ≤ −((λGQRK −λGK )/λmax (Q))Vk−1 +gGKη . 0
(38)
   
vk,1 (t) 0.002 × sin(100(t + 0.2k))
Considering ρ = (1 − (λGQRK − λGK )/λmax (Q)) above vk (t) = vk,2 (t) = 0.002 × cos(100(t + 0.2k)) ,
equation is given as vk,3 (t) 0.02 × 2e(−t+0.2k)
wk (t) = 0.02×e(−t+0.2k) +0.002×sin(100(t+0.2k)).
Vk ≤ ρVk−1 + λGKη . (39) (41)
Choosing ε , R and Q such that, it follows that
Choosing the following values of the parameters, Field
Vk ≤ ρ k−1
V0 + λGKη /(1 − ρ), winding resistance R f = 20Ω, Field winding inductance
(L f ) = 1H, Motor torque ratio (km ) = 0.5 Nm/A, Mechan-
lim Vk ≤ λGKη /(1 − ρ),
x→∞ ical load inertia momentum ( j) = 2 Nm/A, Friction ratio
r
gGKη ( f ) = 2 Nm/A, sampling period (T ) = 0.01, and total time
lim kek−1 k ≤ . (40)
x→∞ (1 − ρ)λmin (Q) span [0 T ] = [0 12]. The discrete model is given as
 
ek is bounded in the presence of disturbances and initial 0.8187 0 0
errors. Also, the proper choice of the weighting matrices AD =  0.4526 0.9975 0  ,
can minimize the bound (40). 0.0023 0.0100 1
 Robust Optimal PID type ILC for Linear Batch Process 783

Fig. 2. (Example 1) Norm 2 to the error at different values

of Q when R=1I. Fig. 4. Comparison of Norm 2 of the Error with respect to
iteration.

Fig. 3. Comparison of Norm 1 of the Error with respect to

iteration. Fig. 5. Comparison of Norm 2 of the Error with respect to
iteration.
 
0  
BD = 0.0197 , CD = 0 0 1 . (42) Let the real length, mass and friction coefficient be l = 1
0.0211 m, m = 2.0 kg, and v = 1.0 kgm2 /s, respectively. Letting
The desired output trajectory, which is shown as yd (t) = h be a sampling period, one can discretize the model by
1.25 × t(t f − t), 0 ≤ t ≤ t f , t f = 12 sec. using the Euler method, as follows:
Here, the PID gains for the PID ILC scheme is calcu-
ml 2 v 2ml 2
 
lated using linear nominal model. Fig. 2 shows the con- τ(ih) = 2 θ (ih + 2h) + − 2 θ (ih + h)
h h h
vergence condition (22) is satisfied for different values of
2ml 2 v
 
weighting matrix R while Q = I at each iteration. It can + − θ (ih) + mgl cos θ (ih). (44)
be observed that for Q = I, different values of R and er- h2 h
ror vector at kth iteration and positive definite matrixes
R and Q, convergence condition (22) always less than 1. Let x1 (t) = θ (ih), x2 (t) = θ (ih + h) and x3 (t) = θ (ih + h)
Also, the appropriate choice of the weighting matrix can then above equation can be written as in state-space form
minimize the λc which can expedite the convergence rate. as
Figs. 3-5 show the Norm 1,2 and infinity of the error re-
xk (t + 1)
spectively, when are calculated using (21), in which linear  
nominal model (42) is used. Afterward, updated input is x1 (t + 1)
=
calculated using (19) and applied to the actual system (41) x2 (t + 1)
iteratively, where Q = 1 × I and R = 0.01 × I are chosen
 
0
as weighting matrices in the cost function. It can be ob- = h2 u(t) + vk (t)
ml 2
served from Figs. 3-5 that proposed algorithm performed  
x2 (t)
better than the methods proposed in [33, 43] in terms of + vh vh gh2 ,
(2− ml 2 )x2 (t)+(ml 2 −1)x1 (t)−( l cos x1 (t))
the convergence and robustness.    
Example 2: To demonstrate the ability of the proposed vk,1 (t) 0.005 sin(100(t + 0.2k))
vk (t) = = ,
method to deal with the nonlinear system, consider a vk,2 (t) 0.005e−(t+0.2k)
single-link manipulator (for detail, please refer to [43,44]) wk (t) = 0.05e(−t+0.2k) + 0.002 sin(100(t + 0.2k)),
τ(t) = ml 2 θ̈ (t) + vθ̇ (t) + mgl cos(t). (43) (45)
784 Furqan Memon and Cheng Shao

Fig. 6. (Example 1) Norm 2 to the error at different values Fig. 7. (Example 2) Norm 1 of the Error with respect to
of Q when R = 1I. iteration.

 
0.03 ∗ rand()
with x(0) . Now, suppose we
0.629 + 0.03 ∗ rand()
have a linear system model as follows:
 
x1 (i + 1)
x2 (i + 1)
 
0
= h2 u(t)
ml 2
" #
x2 (t)
+     . (46)
2− m̃ṽhl˜2 x̃2 (t)+ m̃ṽhl˜2 −1 x̃1 (t) Fig. 8. (Example 2) Norm 2 of the Error with respect to
iteration.
Here, modeled parameters are l˜ = 0.8 m, m̃ = 1.5 m,
and ṽ = 0.8 kgm2 /s. The desired output trajectory, which
is shown in yd = sin(2 × π × t) f ort = (0, 1).
Here, the PID gains for the PID ILC scheme is calcu-
lated and linear nominal model. Fig. 6 shows the conver-
gence condition (22) is satisfied for different values of
weighting matrix R while Q = 1 × I at each iteration. It
can be observed that proper choice of the weighting matrix
can expedite the convergence range.The gains of the PID
type ILC control are calculated using (21), where linear
nominal model (46) with underestimated model parame-
ters is used to calculate the PID gains, Afterward, updated
input is calculated using (19) and applied to the nonlinear
system for different cases mentioned below. Fig. 9. (Example 2) Norm infinity of the Error with respect
Case 1: The uncertainty in modeled parameters are con- to iteration.
sidered in the linear nominal i.e., l˜ = 0.8 m, m̃ = 1.5 m,
and ṽ = 0.8 kgm2 /s are taken into consideration for the real
length l˜ = 1 m, mass m̃ = 2.0 kg and friction coefficient Case 3: The uncertainty in modeled parameters to-
ṽ = 1.0 kgm2 /s, respectively. The behavior of the norm 1, gether with output disturbances appeared at 7th and 11th
2 and infinity of the error are shown using blue lines in batches as shown in Fig. 10. The behavior of the norm
Figs. 7-9 respectively, where Q = 1 × I and are chosen as 1,2 and infinity of the error are shown using green line in
weighting matrices in the cost function. Figs. 7-9, respectively, where and are chosen as weighting
Case 2: The uncertainty in modeled parameters to- matrices in the cost function.
gether with states and output disturbances and initial state Example 3: To demonstrate the ability of the proposed
error are considered as mentioned in (47). The behavior method to deal nominal plant together with uncertainty
of the norm 1, 2 and infinity of the error are shown us- weight and unstructured uncertainty models, Consider the
ing yellow in Figs. 7-9, respectively, where Q = 1 × I and following uncertain plant [30] in the frequency domain as
Q = 1 × I are chosen as weighting matrices in the cost
function. P∆ (s) = P̂(s) + ∆(s)W (s) + d(s), (47)
 Robust Optimal PID type ILC for Linear Batch Process 785

Fig. 11. (Example 3) Norm 1 of the Error with respect to

Fig. 10. (Example 2) Distrubances for the 7th and 11th it-
iteration for a=1, 5,10.
erations.

where nominal plant and disturbance model are given as

5
P̂(s) = , (48)
s+1
1
d(s) = .
3s + 1
With additive uncertainty weight transfer function and un-
structured model are given as

10s Fig. 12. (Example 3) Norm 2 of the Error with respect to

W (s) = ,
(s + 1)(s + 6) iteration for a=1, 5,10.

and
s−a
∆(s) = − ,
(s + a)

a = 1, 2, 3, . . ., 10.
In the uncertainty model, parameter a = 1, 2, 3, . . ., 10
can be considered. The desired output trajectory, which is
shown in yd = (t 3 )(4−0.3t)h for t = (0, 20) with sampling
rate h = 0.02.
Here for finding the PID gains for ILC, linear nominal
model (48) is used, but PID ILC scheme is calculated us-
ing (21). Afterward, updated input is calculated using (19) Fig. 13. (Example 3) Norm Infinity of the Error with re-
and applied to the actual system (47) iteratively. Two cases spect to iteration for a=1, 5,10.
have been considered in simulation. In first case, the un-
certainty without disturbances is simulated, where a = 1, 5
BIBO stability of the proposed PID-ILC control law is dis-
and 10 are chosen for the unstructured uncertainty and the
cussed through a mathematical proof. The performance of
Q = 1 × I and R = 1 × I are chosen as weighting matrices
proposed approach can be improved by a proper choice of
in the cost function. In second case, the disturbance to-
the weighting matrices. The effectiveness of the proposed
gether with uncertainty is also appeared with the nominal
technique has been demonstrated through simulations of
model. Figs. 11-13 show the Norm 1, 2 and infinity of the
three numerical examples.
error respectively for the different uncertainties together
with and without disturbances.
REFERENCES
6. CONCLUSION [1] M. H. Khooban, T. Niknam, and M. Sha-Sadeghi, “Speed
control of electrical vehicles: a time-varying proportional-
In this brief, a new approach for evaluating the optimal integral controller-based type-2 fuzzy logic,” IET Science,
gains of Robust PID type of ILC is proposed, which take Measurement & Technology, vol. 10, no. 3, pp. 185-192,
disturbance and initial error into consideration. The robust 2016.
786 Furqan Memon and Cheng Shao

[2] M. R. Tavana, M.-H. Khooban, and T. Niknam, “Adaptive [16] T. Ersal, M. Brudnak, A. Salvi, Y. Kim, J. B. Siegel,
PI controller to voltage regulation in power systems: STAT- and J. L. Stein, “An iterative learning control approach to
COM as a case study,” ISA Transactions, vol. 66, pp. 325- improving fidelity in internet-distributed Hardware-in-the-
334, 2017. loop simulation,” J. Dyn. Syst. Meas. Control., vol. 136, no.
6, p. 061012, 2014.
[3] M.-C. Fang, Y.-Z. Zhuo, and Z.-Y. Lee, “The application of
the self-tuning neural network PID controller on the ship [17] W. Kase, “A design of parameter optimal iterative learn-
roll reduction in random waves,” Ocean Engineering, vol. ing control for linear discrete-time systems,” ISRN Applied
37, no. 7, pp. 529-538, 2010. Mathematics, vol. 2014, Article ID 816860, Jan 2014.
[4] M. Uchiyama, “Formation of high-speed motion pattern of [18] D. H. Owens and K. Feng, “Parameter optimization in iter-
a mechanical arm by trial,” Trans. Soc. Instrum. Control ative learning control,” Int. J. Control, vol. 76, no. 11, pp.
Eng, vol. 14, no. 6, pp. 706-712, 1978. 1059-1069, 2003.

[5] S. Arimoto, S. Kawamura, and F. Miyazaki, “Bettering op- [19] S. K. Oh and J. M. Lee, “Iterative learning control inte-
eration of dynamic systems by learning: A new control the- grated with model predictive control for real-time distur-
ory for servomechanism or mechatronics systems,” Proc. bance rejection of batch processes,” J. Chem. Eng. Japan,
of 23rd IEEE Conf. Decis. Control, vol. 37, no. 7,pp. 1064- vol. 50, no. 6, pp. 415-421, 2017.
1069, 1984. [20] S. K. Oh and J. M. Lee, “Iterative learning model predictive
control for constrained multivariable control of batch pro-
[6] Y. Wang, F. Gao, and F. J. Doyle, “Survey on iterative
cesses,” Comput. Chem. Eng., vol. 93, pp. 284-292, 2016.
learning control, repetitive control, and run-to-run control,”
J. Process Control, vol. 19, no. 10, pp. 1589-1600, 2009. [21] T. J. Harte, J. Hätönen, and D. H. Owens, “Discrete-time
inverse model-based iterative learning control: Stability,
[7] H. Anwaar, Y. Y. Xin, and S. Ijaz, “A comprehensive sur-
monotonicity and robustness,” Int. J. Control., vol. 78, no.
vey on recent developments in iterative learning control al-
8, pp. 577-586, 2005.
gorithms and applications,” Proc. of 29th Chinese Control
Decis. Conf., CCDC 2017, pp. 3282-3289, 2017. [22] A. Haber, R. Fraanje, and M. Verhaegen, “Linear computa-
tional complexity robust ILC for lifted systems,” Automat-
[8] H. Sun and A. G. Alleyne, “A cross-coupled non-lifted ica, vol. 48, no. 6, pp. 1102-1110, 2012.
norm optimal iterative learning control approach with ap-
plication to a multi-axis robotic testbed,” Proc. of IFAC, [23] B. Cichy, K. Gałkowski, and E. Rogers, “2D systems based
vol. 47, no. 3, pp. 2046-2051, 2014. robust iterative learning control using noncausal finite-time
interval data,” Syst. Control Lett., vol. 64, no. 1, pp. 36-42,
[9] D. H. Owens, C. T. Freeman, and B. Chu, “Norm optimal 2014.
iterative learning control for planar tracking tasks,” Proc.
of IFAC, vol. 46, no. 11, pp. 683-688, 2013. [24] Y. Wang, H. Zhang, S. Wei, D. Zhou, and B. Huang, “Con-
trol performance assessment for ILC-controlled batch pro-
[10] J. W. Axehill, I. Dressler, S. Gunnarsson, A. Robertsson, cesses in a 2-D system framework,” IEEE Trans. Syst. Man,
and M. Norrlöf, “Estimation-based ILC applied to a paral- Cybern. Syst., vol. 48, no. 9, pp. 1493-1504, 2018.
lel kinematic robot,” Control Eng. Pract., vol. 33, pp. 1-9,
[25] H. Shokri-Ghaleh and A. Alfi, “Bilateral control of un-
2014.
certain telerobotic systems using iterative learning control:
[11] J. Bolder, J. Van Zundert, S. Koekebakker, and T. Oomen, Design and stability analysis,” Acta Astronautica, vol. 156,
“Enhancing flatbed printer accuracy and throughput: Op- pp. 58-69, 2019.
timal rational feedforward controller tuning via iterative
[26] X. Ge, J. L. Stein, and T. Ersal, “Frequency-domain analy-
learning control,” IEEE Trans. Ind. Electron., vol. 64, no.
sis of robust monotonic convergence of norm-optimal iter-
5, pp. 4207-4216, 2017.
ative learning control,” IEEE Trans. Control Syst. Technol.,
[12] D. J. Hoelzle and K. L. Barton, “On spatial iterative learn- vol. 26, no. 2, pp. 637-651, 2018.
ing control via 2-D convolution: stability analysis and com- [27] D. H. Nguyen and D. Banjerdpongchai, “A convex opti-
putational efficiency,” IEEE Trans. Control Syst. Technol., mization approach to robust iterative learning control for
vol. 24, no. 4, pp. 1504-1512, 2016. linear systems with time-varying parametric uncertainties,”
[13] C. Peng, L. Sun, W. Zhang, and M. Tomizuka, Automatica, vol. 47, no. 9, pp. 2039-2043, 2011.
“Optimization-based constrained iterative learning control [28] D. A. Bristow and A. G. Alleyne, “Monotonic convergence
with application to building temperature control systems,” of iterative learning control for uncertain systems using a
Proc. of IEEE Int. Conf. Adv. Intell. Mechatronics, pp. 709- time-varying filter,” IEEE Trans. Automat. Contr., vol. 53,
715, 2016. no. 2, pp. 582-585, 2008.
[14] C. T. Freeman, A. M. Hughes, J. H. Burridge, P. H. Chap- [29] M. Q. Phan, R. W. Longman, B. Panomruttanarug, and S.
pell, P. L. Lewin, and E. Rogers, “Iterative learning control C. Lee, “Robustification of iterative learning control and
of FES applied to the upper extremity for rehabilitation,” repetitive control by averaging,” Int. J. Control, vol. 86, no.
Control Eng. Pract., vol. 17, no. 3, pp. 368-381, 2009. 5, pp. 855-868, 2013.
[15] F. Felici and T. Oomen, “Enhancing current density pro- [30] T. D. Son, G. Pipeleers, and J. Swevers, “Robust mono-
file control in tokamak experiments using iterative learning tonic convergent iterative learning control,” IEEE Trans.
control,” Proc. of 54th IEEE CDC, pp. 5370-5377, 2015. Automat. Contr., vol. 61, no. 4, pp. 1063-1068, 2016.
 Robust Optimal PID type ILC for Linear Batch Process 787

[31] K. K. Tan, S. Zhao, and J.-X. Xu, “Online automatic tuning [43] A. Madady, H. R. Reza-Alikhani, and S. Zamiri, “Optimal
of a proportional integral derivative controller based on an N-parametric type iterative learning control,” Int. J. Con-
iterative learning control approach,” IET Control Theory & trol. Autom. Syst., vol. 16, no. 5, pp. 2187-2202, 2018.
Applications, vol. 1, no.1, pp. 90-96, 2007. [44] C. K. Chen and J. Hwang, “PD-type iterative learning con-
[32] H. Liu, Y. Li, Y. Zhang, Y. Chen, Z. Song, Z. Wang, S. trol for trajectory tracking of a pneumatic X-Y table with
Zhang, and J. Qian, “Intelligent tuning method of PID disturbances,” JSME International Journal., Series C, vol.
parameters based on iterative learning control for atomic 49, no. 2, pp. 520-526, 2006.
force microscopy,” Micron, vol. 104, no. 37, pp. 26-36, [45] C. Tianyou, “An optimal design of open and close loop
2018. PID type ILC algorithm,” Proc. of Fifth International Sym-
[33] A. Madady, “An extended PID type iterative learning con- posium on Instrumentation and Control Technology, vol.
trol,” Int. J. Control. Autom. Syst., vol. 6, no. 2, pp. 194- 5253, pp. 751-755, International Society for Optics and
203, 2008. Photonics, 2003.
[34] A. Madady, “An extended PID type iterative learning con-
trol,” Int. J. Control. Autom. Syst., vol. 11, no. 3, pp. 470-
481, 2013. Furqan Memon received his B.E. degree
in electronic engineering from Mehran
[35] T. Liu, X. Z. Wang, and J. Chen, “Robust PID based
University of Engineering and Technol-
indirect-type iterative learning control for batch processes
ogy, Pakistan, in 2007, and an master’s
with time-varying uncertainties,” J. Process Control, vol.
degree from National University of Sci-
24, no. 12, pp. 95-106, 2014.
ence and Technology, Pakistan, in 2011.
[36] Z. Baolin, T. Gongyou, and Z. Shi, “PD-Type iterative At present, he is pursuing a Ph.D. degree
learning control for nonlinear time-delay system with ex- from Institute of Advanced Control Tech-
ternal disturbance,” J. Syst. Eng. Electron., vol. 17, no. 3, nology, Dalian University of Technology,
pp. 600-605, 2006. Dalian, P. R. China. His research interests include robotics, ro-
[37] K. H. Park, Z. Bien, and D. H. Hwang, “A study on the ro- bust & optimal control and iterative learning control.
bustness of a PID-type iterative learning controller against
initial state error,” Int. J. Syst. Sci., vol. 30, no. 1, pp. 49-59, Shao Cheng is a Professor, the Head
1999. of Institute of Advanced Control Tech-
[38] G. Ji and Q. Luo, “An open-closed-loop PIDtype iterative nology, Dalian University of Technology,
learning control algorithm for uncertain time-delay sys- Dalian, P. R. China, and a Dean of Liaon-
tems,” Proc. of the International Conference on Machine ing Province Key Laboratory of Advanced
Learning and Cybernetics, pp. 1154-1159, 2005. Control System for Industrial Equipment.
He received Ph.D. degree and a master’s
[39] Y. Chen and K. L. Moore, “An optimal design of PD-type
degree in control science and engineering
iterative learning control with monotonic convergence,”
at Northeastern University, P. R. China, in
Proc. of IEEE Internatinal Symp. Intell. Control., pp. 55-
1992 and 1986, respectively, and a Bachelor’s degree in mathe-
60, 2002.
matics at Liaoning University, P. R. China, in 1982. His research
[40] Y. Chen and K. L. Moore, “PI-type iterative learning con- interests include control theory in intelligent manufacturing, ad-
trol revisited,” Proc. of 2002 Am. Control Conf, IEEE Cat. vanced process control, robust adaptive control, intelligent mea-
No.CH37301, vol. 3, 2002. surement and control.
[41] D. Wang, “On D-type and P-type ILC designs and anticipa-
tory approach,” Int. J. Control, vol. 73, no. 10, pp. 890-901, Publisher’s Note Springer Nature remains neutral with regard
2000. to jurisdictional claims in published maps and institutional affil-
iations.
[42] F. Memon and C. Shao, “An optimal approach to online
tuning method for PID type iterative learning control,” Int.
J. Control. Autom. Syst., vol. 18, pp. 1926-1935, 2020.