8th IFAC Symposium on Advanced Control of Chemical Processes

The International Federation of Automatic Control

Singapore, July 10-13, 2012

Identification and Controller Tuning of Cascade Control Systems Based on

Closed-Loop Step Responses
Jyh-Cheng Jeng*, Ming-Wei Lee**

*Department of Chemical Engineering and Biotechnology, National Taipei University of Technology, Taipei 106, Taiwan
(e-mail: jcjeng@ntut.edu.tw).
**Department of New Materials Research & Development, China Steel Corporation, Kaohsiung 812, Taiwan
(e-mail: 160176@mail.csc.com.tw)

Abstract: This paper presents a new automatic tuning method for cascade control systems based on a
single closed-loop step test. The proposed method identifies the required process information with the help
of B-spline series expansions for the step responses. The two PID controllers are then tuned using an
internal model control (IMC) approach. The secondary controller is designed for enhanced disturbance
rejection, and the primary controller is designed, without requiring an additional test, based on an
identified process model that accurately accounts for the inner loop dynamics. The desired levels of system
robustness explicitly guide the selection of the IMC tuning parameters. The proposed method is robust to
measurement noises because of the filtering property of the B-splines, and can provide satisfactory control
performance. Simulation examples confirm the effectiveness of the proposed method.
Keywords: Process control; Cascade control; Closed-loop identification; B-spline series; PID tuning

controllers. Lee et al. (2002) subsequently improved the

1. INTRODUCTION
Cascade control is one of the most successful control
structures for enhancing single-loop control performance,
which has led to the extensive implementation of cascade
control in chemical process industries. The cascade control
structure is to nest one feedback loop inside another feedback
loop using two controllers, so that the fast dynamics of the
inner loop will provide faster disturbance attenuation.
Though recent studies have proposed sophisticated cascade
control schemes (e.g., Alfaro et al., 2009; Kaya et al., 2007),
the standard cascade control scheme still consists of two
nested loops with two PID controllers. The design of this
cascade control system involves the tuning of two PID
controllers, and is therefore more complicated than that of a
single-loop control system.
Previous researches have proposed relay-based auto-tuning
techniques to facilitate the design of cascade control systems.
The method proposed by Hang et al. (1994) needs a
sequential application of the conventional relay-based auto-
tuning approach, and is therefore still time consuming. The
sequential tuning procedure has been improved so that only a
single relay experiment is required for auto-tuning (Leva and
Donida, 2009; Mehta and Majhi, 2011; Tan et al., 2000).
However, an off-line or ad hoc experiment must be
performed in these methods. Veronesi and Visioli (2011)
recently proposed a simultaneous closed-loop automatic
tuning method for cascade controllers from the set-point step
response of a cascade control system.
The literature also reports tuning rules for cascade controllers.
Lee et al. (1998) proposed IMC-based PID tuning rules that
enable simultaneous tuning of primary and secondary
tuning rules using a general cascade control structure that
requires two additional set-point filters. However, these
studies fail to clarify how the procedure can be automated.
The main point of simultaneously tuning cascade controllers
(Lee et al., 1998; Lee et al., 2002; Mehta and Majhi, 2011;
Veronesi and Visioli, 2011) is to approximate the inner loop
dynamics with the inner loop design target (e.g., a desired
closed-loop transfer function for the inner loop). Such an
approximation allows obtaining a process model for the
tuning of primary controller. However, this approximation
may be inaccurate because the implemented secondary (PID)
controller cannot guarantee achieving the inner loop design
target. The proposed method addresses this problem.
This study develops an on-line automatic tuning method that
simultaneously tunes two PID controllers for cascade control
systems based on a single closed-loop step test. By
representing the step responses with B-spline series
expansions, this method first identifies the analytical
expressions of process transfer functions, which can be used
to obtain appropriate process models for controller design.
The IMC approach is then used to design secondary and
primary controllers. The cascade control design here achieves
satisfactory servo and regulatory control performance, while
the control structure is stuck to the standard cascade control
configuration (i.e., it avoids using more complicated control
schemes such as two-degree-of-freedom control). This is
accomplished by designing the inner loop for enhanced
disturbance performance, and also taking into account the
actual inner loop dynamics (without an additional experiment)
in the primary controller design. Finally, the explicit
inclusion of robustness considerations in controller tuning
results in a complete automatic tuning procedure.

978-3-902823-05-2/12/$20.00 © 2012 IFAC 414 10.3182/20120710-4-SG-2026.00024

8th IFAC Symposium on Advanced Control of Chemical Processes
Singapore, July 10-13, 2012

2. PROCESS IDENTIFICATION B-splines with a finer knot sequence will have enhanced
capability for representing a signal with sharp changes. A
Figure 1 shows the configuration of a typical cascade control suitable choice of the knot sequence depends on the “shape”
system, where G1 is the primary process and G2 is the of the signal to be represented. The B-spline series works as a
secondary process. The primary process variable y1 (with set- low-pass filter that can effectively remove the high-frequency
point r1) is used by the primary controller Gc1 to establish the noise when it is used to represent a noisy signal.
set-point (r2) for the secondary controller Gc2. The secondary
process variable y2 is transmitted to the secondary controller The Laplace transforms of Bx , m ; k (t ) , designated as Bɶ x , m ; k ( s ) ,
which adjusts the manipulated variable u. The effectiveness can be obtained by the following recurrence relations.
of a cascade control scheme is because the disturbance d2  1 m −1 ɶ 
affecting the secondary (inner) loop is effectively  1 − Bx,m −1;2− m ( s )  for k = − m + 1 (4)
 s  x1 − x2− m 
compensated before it affects the primary process variable y1. Bɶ x,m;k ( s ) = 
 1  m − 1 Bɶ m −1 ɶ 
 s x x , m −1;k ( s ) − Bx,m−1;k +1 ( s )  for k ≥ −m + 2
  k + m−1 − xk xk + m − xk +1 

The following discussion uses the 3rd-order (m=3) B-splines

to develop the closed-loop identification method, and the
knot sequence x can be inferred from the context. For
Fig. 1. Configuration of a typical cascade control system. conciseness, Bx , m ; k (t ) and Bɶ x , m ; k ( s ) are hereafter represented

The process identification is accomplished by conducting a as Bk (t ) and Bɶ k ( s ) , respectively.

simple step test in the set-point during closed-loop operation
2.2 Identification of process transfer functions
of a cascade control system. Compared with open-loop test,
the closed-loop test has the significant advantage of not From the scheme of Fig. 1, the transfer functions of the
excessively disturbing the system. The controller parameters primary and the secondary processes can be expressed as
can be chosen arbitrarily (e.g., roughly tuned). The step
y1 ( s) y2 ( s )
response data of the process variables, y1(t) and y2(t), and the G1 ( s ) = , G2 ( s ) = (5)
manipulated variable, u(t), are collected until the new steady- y2 ( s ) u ( s)
state is reached. To obtain the Laplace transforms of the measured responses,
define new variables in the deviation form as
The following subsection briefly introduces the B-splines
because the proposed identification method uses B-spline y1 (t ) = y1 (t ) − y1, s ; y2 (t ) = y2 (t ) − y2, s ; u (t ) = u (t ) − us (6)
series expansions to represent the measured step responses. where y1,s, y1,s, and us designate the new steady-state values of
y1(t), y2(t), and u(t), respectively, after the step change in the
2.1 B-spline series
set-point. Then, each of the responses in (6) is approximated
Since the step responses are functions of time for t ≥ 0 , we as a spline using the 3rd-order B-spline series expansions as
introduce the B-splines on the semi-infinite interval [0, ∞ ). Nk Nk

Let x ={xk} be a non-decreasing sequence of real numbers, y1 (t ) ≅ yˆ1 (t ) = ∑c y1 , k Bk (t ); y2 (t ) ≅ yˆ 2 (t ) = ∑c y2 , k Bk (t )

k =−2 k =−2
(7)
defined as x : 0 = x− m +1 = ⋯ = x0 < x1 < ⋯ < xk < ⋯ . For each Nk
u (t ) ≅ uˆ (t ) = ∑c Bk (t )
k ≥ − m + 1 , Bx , m ; k (t ) is called the kth B-spline of order m with k =−2
u ,k

knot sequence x. Computations with B-splines are facilitated Each response in the deviation form has a zero steady-state
by stable recurrence relations (Chui, 1997) value, which allows it to be approximated with finite terms of
t − xk x −t (1) B-splines ( k = −2, −1, 0,⋯ , N k ). After applying the Laplace
Bx ,m;k (t ) = Bx ,m −1;k (t ) + k + m Bx ,m −1;k +1 (t )
xk + m −1 − xk xk + m − xk +1 transform to (6) and (7), the approximated original responses
can be expressed in the Laplace domain as
with initial condition
Nk
y1, s Nk
y2, s
 1 if xk ≤ t < xk +1 (2) yˆ1 ( s ) = ∑c Bɶ k ( s ) + ; yˆ 2 ( s ) = ∑c Bɶ k ( s ) +
Bx,1;k (t ) =  , k = 0,1,⋯ y1 , k
s
y2 , k
s
 0 otherwise k =−2 k =−2 (8)
Nk
u
The collection of mth-order B-splines Bx , m ; k (t ) , k ≥ − m + 1 is uˆ ( s ) = ∑ cu , k Bɶ k ( s ) + s
k =−2 s
a basis of the spline space on [0, ∞ ). In this case, the B-spline
series describes a spline Sp(t) as a linear combination The B-spline coefficients c y , k , k = −2, −1, 0,⋯ , N k can be
1
∞
(3) obtained by minimizing the following cost function
S p (t ) = ∑ ck Bx ,m;k (t )
M 2
min ∑  y1 (tm ) − yˆ1 (tm ) 
k =− m +1
(9)
which can be used for representing a signal (function). c y1 ,k
m =1

The representation of a signal using B-spline series has
several advantages. The B-splines is a local basis (with finite
time duration) which has greater flexibility in signal
representation than a global basis, such as Laguerre functions.
where M is the number of measurements and tm is the time of
the mth measurement. The above optimization problem can
be easily solved by a least-squares estimation method. The
coefficients c y , k and cu , k can be obtained in a similar way.
2

415
8th IFAC Symposium on Advanced Control of Chemical Processes
Singapore, July 10-13, 2012

By choosing the knot sequence, the use of B-splines is better
able to represent the closed-loop responses. A suitable knot
sequence can be chosen based on the characteristics of the
measured responses. Because the closed-loop step responses
usually have time delay and sharp transient dynamics
(especially for the manipulated variable u), a finer knot
sequence can be chosen for the initial period of response, and
a coarser knot sequence can be chosen when the response
approaches the steady-state. This method of choosing a non-
uniform knot sequence provides a favorable trade-off
between approximation accuracy and computational effort.
From (5), the analytical expressions of process transfer
functions now can be calculated by substituting (8) as
Nk
y Nk
∑c Bɶ k ( s ) + 1, s ∑c
y1 , k
s
y2 , k
where Gm 2 − is the invertible part of the model Gm2, and F2 is
the IMC filter to deal with the robustness issue of the inner
loop. For stable processes, this IMC filter is typically chosen
as lag element without lead dynamics, which results in
sluggish response to disturbance d2 if Gm2 has slow poles.
Because the performance of cascade control in the presence
of disturbance d2 is usually the principal concern, the inner
loop must respond quickly to effectively reduce the effects of
disturbance d2. To this end, this study proposes that an IMC
filter F2 designed so that (1 − Gm 2 + F2 ) , where Gm 2 + is the non-
invertible part of the model Gm2, cancels the slow poles in
Gm2. As it is typical in the industrial context, the secondary
process is modeled as a first-order plus time delay (FOPTD)
dynamics:
y
Bɶ k ( s ) + 2, s K 2 −θ s (13)
s Gm 2 ( s ) = e 2

Gˆ1 ( s ) = kN=−2 , Gˆ 2 ( s ) = k =−N2k (10) τ 2s +1

k
y u
∑ y2 , k k
c Bɶ ( s ) +
s
2, s
∑ cu,k Bɶk (s) + ss
k =−2
In this case, the IMC filter is chosen as
k =−2
α s +1   λ2  2 −θ τ2  (14)
The process transfer functions obtained from (10) are high- F2 ( s ) = ; α = τ 2 1 − 1 − τ  e
2

( λ2 s + 1)
2
order with complex forms. These transfer functions can be   2  
directly used to calculate the process frequency responses, where the parameter λ2 is the adjustable parameter to make
Gˆ1 ( jω ) and Gˆ 2 ( jω ) , by substituting s = jω for arbitrary trade-off between performance and robustness of the inner
frequencies. The transfer functions can also be reduced to loop. The design target for the complementary sensitivity
low-order plus time delay models, which are suitable for function of the inner loop, T2, then becomes
controller design, by a simple model reduction technique α s + 1 −θ 2 s (15)
T2 t ( s ) = Gm 2 + ( s ) F2 ( s ) = e
based on the calculated frequency response data. First, the ( λ2 s + 1)
2

steady-state gain of the model, K, is determined as that of the

The lead term in (15) may cause an excessive overshoot in
approximated process transfer function by the set-point response of the inner loop. Therefore, a set-point
K = lim Gˆ ( s) (11) filter is usually used to eliminate the overshoot, resulting in a
s →0
more complicated control system (Lee at al., 2002).
The other model parameters then can be determined by Nevertheless, an additional set-point filter for the inner loop
minimizing the difference between the estimated process of cascade control system is not required because the
frequency response and the frequency response of the response of the primary process variable y1 is the major
reduced model. To avoid an iterative nonlinear optimization concern. This is because the primary process attenuates the
procedure in the parameter estimation, a two-step least- lead dynamics of the inner loop. Furthermore, the primary
squares technique is suggested. Because the amplitude ratio controller can be designed to compensate the lead dynamics
is independent of time delay, the first step estimates the of the inner loop so that the response of the primary process
model parameters, excluding time delay θ , by minimizing variable remains satisfactory.
the difference in the amplitude ratio. The estimated model
parameters are then used in the second step to estimate the The conventional feedback controller Gc2 is related to the
time delay of the model by minimizing the difference in the IMC controller Q2 as
phase angle. Both minimization problems can be solved by a
standard linear least-squares technique. The computation of
Q2 ( s ) (τ 2 s + 1)(α s + 1) (16)
Gc 2 ( s ) = =
the model parameters can be repeated for various types of 1 − Gm 2 ( s) Q2 ( s ) K 2 ( λ2 s + 1) 2 − e −θ2 s (α s + 1) 
 
low-order models, and the model that minimizes the residuals The above equivalent controller is not in the form of PID
of the least-squares problem is chosen. controller. Therefore, the Maclaurin series expansion formula
is applied to obtain a PID controller which approximates Gc 2
3. PID CONTROLLER TUNING
given in (16) as
This study proposes to use an IMC approach (Morari and 1 f 2′′( 0 ) 2   1 
Zafiriou, 1989) for the controller design. Gc 2 ( s ) =  f 2 ( 0 ) + f 2′ ( 0 ) s + s + ⋯ ≈ K c 2  1 + + τ D 2 s  (17)
s 2   τ I2 s 
3.1 Design of the secondary controller where f 2 ( s ) = s Gc 2 ( s ) . The PID parameters are obtained as
According to the method presented in the previous section, a f 2′ ( 0 ) f ′′( 0 )
K c 2 = f 2′ ( 0 ) ; τ I 2 = ; τ D2 = 2 (18)
secondary process model, Gm2, can be obtained from the f2 (0) 2 f 2′ ( 0 )
frequency response Gˆ 2 ( jω ) . The IMC controller Q2 is
Notice that a PI controller or a P controller can be
designed as
approximated using only the first two terms or using only the
Q2 ( s) = Gm 2 − ( s) −1 F2 ( s) (12) second term, respectively, in (17).

416
8th IFAC Symposium on Advanced Control of Chemical Processes
Singapore, July 10-13, 2012

3.2 Design of the primary controller 1

(23)
*
T1t ( s ) = Gm* 1+ ( s ) F1 ( s ) = e −θ1 s
λ1 s + 1
With the designed secondary controller, the primary
controller is designed based on an apparent process G1* seen For the SOPTDZ model, the conventional feedback controller
Gc1 is related to the IMC controller Q1 as
by the primary controller as
Gc 2 ( s )G2 ( s) Q1 ( s) a1* s 2 + b1* s + 1 (24)
G1* ( s) = T2 ( s ) G1 ( s) = G1 ( s) (19) Gc1 ( s) = =
1 + Gc 2 ( s )G2 ( s) (
1 − Gm* 1 ( s )Q1 ( s) K1* ( z1* s + 1) λ1s + 1 − e−θ1* s )
Most existing works for simultaneous cascade controller The Maclaurin series expansion formula is applied to obtain a
design (Lee et al., 1998; Lee et al., 2002; Mehta and Majhi, PID controller which approximates Gc1 given in (24).
2011; Veronesi and Visioli, 2011) approximate this apparent Defining f1 ( s) = s Gc1 ( s ) , the PID parameters are obtained as
process using the design target of the complementary
sensitivity function of the inner loop, i.e., G1* = T2 t G1 , for f1′( 0 ) f ′′( 0 ) (25)
K c1 = f1′( 0 ) ; τ I 1 = ; τ D1 = 1
f1 ( 0 ) 2 f1′( 0 )
primary controller design. Because the ideal secondary
controller is approximated as a classical PID-type controller
3.3 Guidelines for the selection of tuning parameters
for implementation, the actual secondary complementary
sensitivity function may deviate from its design target T2 t , In the proposed approach, λ1 and λ2 are the tuning
depending on the accuracy of PID approximation. This parameters to handle the trade-off between the speed of
deviation may be especially significant when the ideal response and system robustness. This study includes
controller is approximated as a P or PI controller, which is robustness considerations in the controller design and
usually the controller type of the secondary controller. To develops explicit guidelines for the automated selection of
account for the non-ideality of the inner loop, this study tuning parameters to achieve the desired level for the
identifies a model of G1* according to the actually robustness. This aim is accomplished by using the maximum
implemented secondary controller. After the frequency sensitivity as a robustness measure. The maximum sensitivity
responses of the secondary and primary processes have been (MS) is defined as
identified, (19) permits the calculation of the frequency M S = max S ( jω ) = max 1 − T ( jω ) (26)
ω ω
response of Gˆ1* ( jω ) in terms of frequency responses of the where S denotes the sensitivity function with S + T = 1 . As
designed secondary controller, the secondary process, and the MS decreases, the dynamic response of closed-loop system
primary process. Therefore, a model of the apparent process, becomes more robust. The recommended values for MS are
Gm* 1 , can be identified from Gˆ1 ( jω ) using the previously typically within the range 1.2 < M S < 2.0 .
*

presented least-squares method. Note that the identification

The complementary sensitivity function for the inner loop (15)
of Gm* 1 involves only simple linear least-squares computation
can be normalized by τ 2 as
and does not require conducting an additional experiment.
t α s + 1 −θ s (27)
With the model Gm* 1 , the IMC design is applied again to T2 (s ) = e 2

( λ s + 1)
2

design the primary controller Gc1. The IMC controller Q1 is 2

designed as where α = α τ 2 = 1 − (1 − λ2 )2 e −θ , λ2 = λ2 τ 2 , and θ 2 = θ 2 τ 2 .

2

Q1 ( s) = Gm* 1− ( s)−1 F1 ( s) (20)
where F1 is the IMC filter dealing with the robustness issue
of the outer loop. If the primary controller is designed for
enhanced disturbance performance, a set-point filter for the
outer loop must be incorporated to reduce the excessive
overshoot in the set-point response, which complicates the
control system. To maintain the simplicity of the control
system, the typical IMC filter design is adopted for F1, i.e.,
1 (28)
F1 ( s) = (21)
Equation (27) indicates that the maximum sensitivity of the
inner loop, MS2, depends on λ2 and θ 2 . Therefore, this study
calculates the maximum sensitivity MS2 for the combinations
of various values of λ2 and θ 2 . The resulting relations among
the normalized design parameter λ2 , the inner loop
robustness MS2, and the normalized time delay θ 2 are
correlated as the following inner loop robust design criterion:
λ2 = p1 M S 2 p + p3
2

λ1 s + 1
where
where λ1 is the adjustable parameter to make trade-off
−3.998θ 2 2 + 4.279θ 2 − 0.0397
between performance and robustness of the outer loop. p1 =
θ 23 − 3.242 θ 2 2 + 2.133θ 2 + 0.139
The apparent process for the primary controller design is 1410 θ 2 2 − 212θ 2 − 4484 (29)
p2 =
generally modeled using second-order plus time delay and θ 2 − 2228θ 2 2 + 1962 θ 2 + 759.2
3

zero (SOPTDZ) dynamics: p3 = 1.203θ 23 − 1.463θ 2 2 + 0.789θ 2 + 0.01257

Gm* 1 ( s ) =
K *
1 ( z s + 1) e
*
1 −θ1* s
, z1* ≥ 0 (22) This criterion provides the required value of λ2 , for a given
a1* s 2 + b1* s + 1
θ 2 , to achieve a desired value of MS2. Notice that (28) is
Therefore, the design target for the complementary sensitivity
function of the outer loop, T1, now becomes applicable to 1.2 ≤ M S 2 ≤ 1.8 and 0.05 ≤ θ 2 ≤ 1 .

417
8th IFAC Symposium on Advanced Control of Chemical Processes
Singapore, July 10-13, 2012

The complementary sensitivity function for the outer loop (23) G1

0

can be normalized by θ1* as -2

1 (30)
-4

T1t ( s ) = e− s -6

λ1s + 1 -8 Actual
Identified
where λ1 = λ1 θ1 . The maximum sensitivity of the outer loop,
* -10
-6 -4 -2 0 2 4 6 8 10 12

G2
MS1, depends only on λ1 . Therefore, this study calculates MS1 0.5

for various values of λ1 . The resulting relationship between -0.5

-1

the normalized design parameter λ1 and the outer loop -1.5

robustness MS1 is correlated as the following outer loop -2

-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

robust design criterion, which provides the required value of

Fig. 3. Actual and identified process frequency responses
λ1 to achieve a desired value of MS1.
(Nuquist plots) for example 1.
−0.7289 M S1 + 1.555 (31)
λ1 = 2.972 (33)
M S 1 − 1.006 Gm 2 ( s) = e −3.053 s
12.970 s + 1
Notice that (31) is applicable to 1.2 ≤ M S 1 ≤ 1.8 . Choosing MS2 = 1.55, which is the value of MS2 resulting
from the tuning of Lee et al. (1998), λ2 = 5.165 is obtained,
4. SIMULATION EXAMPLE
and the resulting PID parameters are K c 2 = 0.819 ,
Two simulation examples are provided to demonstrate the τ I 2 = 10.038 , and τ D 2 = 0.700 . The resulting inner loop has
effectiveness of the proposed automatic tuning method. MS2 = 1.524, very near to the design value. To tune the
primary controller, a SOPTD model for the apparent process
4.1 Example 1
G1* is then identified from Gˆ1* ( jω ) as
Consider the high-order process given by the following 9.979
transfer functions (Lee et al., 1998; Mehta and Majhi, 2011) Gm* 1 ( s) = e−43.77 s (34)
1922.3s 2 + 81.11s + 1
10 ( −5s + 1) 3
G1 ( s ) = e −5 s , G2 ( s ) = e −3 s (32) Choosing MS1 = 1.5, λ1 = 40.905 is obtained and the resulting
( 30 s + 1) (10s + 1) 13.3s + 1
3 2
PID parameters are K c1 = 0.109 , τ I 1 = 92.42 , and τ D1 = 30.33 .
The cascade control system with PID/PI control mode tuned Figure 4(a) shows the closed-loop response to a unit step set-
by Mehta and Majhi (2011) is adopted for closed-loop test. point change at t = 0 followed by step disturbance inputs
Gaussian random noises with standard deviations 0.02, 0.006, d 2 = 0.5 at t = 700 and d1 = 0.2 at t = 1500 , and presents
and 0.004 were added to the original unit step responses of y1,
y2, and u, respectively, to represent real measurements. the results obtained by Lee et al. (1998) for comparison.
Figure 2 shows the noisy responses and the approximated 2.5
(a)

responses by B-spline series. The B-spline series expansions 2

provide smooth approximated responses, effectively reducing 1.5

the effect of noise. Figure 3 shows the identified process

y1

frequency responses and comparisons with the actual 0.5 Proposed

Lee et al.
frequency responses for G1 and G2. An excellent agreement 0
0 500 1000 1500 2000

exists between the estimated and actual frequency responses. time

(b)
2.5

To tune the secondary controller, an FOPTD model for G2 is 2

first identified from Gˆ 2 ( jω ) as 1.5

y1

1 0.5 Proposed
0 Lee et al.
0.5
y1

0 500 1000 1500 2000

Measured response time
Approximated response
Fig. 4. Closed-loop responses for example 1: (a) nominal
0
0 100 200 300 400 500 600 700
plant (b) perturbed plant.
0.1

0.05 These results illustrate that the proposed automatic tuning

y2

0
method outperforms the method of Lee et al. (1998). To
illustrate the robustness to parameter variations, the control
-0.05
0 100 200 300 400 500 600 700

0.3 systems are evaluated on a perturbed plant where +20%

0.2 deviations in gain and time delay from the nominal values for
u

0.1 both primary and secondary processes are considered. Figure

0
0 100 200 300 400 500 600 700
4(b) shows the responses of perturbed systems. The response
time achieved with the proposed method is still better, indicating
Fig. 2. Noisy and approximated closed-loop step responses that the proposed cascade control system not only provides a
for example 1. fast response, but also maintains a reasonable robustness.

418
8th IFAC Symposium on Advanced Control of Chemical Processes
Singapore, July 10-13, 2012

4.2 Example 2 1.6

1.4
As mentioned previously, most existing works design the
primary controller based on an apparent process model that is 1.2

obtained using the design target of the inner loop. To 1

illustrate the drawback in previous works, the cascade control 0.8

y1
system with the PID/P control mode and the following 0.6

processes is considered. 0.4

1 1 −s (35)
G (s) =
1 e −2 s , G ( s ) = e
2
5s + 1 s +1
0.2
Proposed
Lee et al.
This study compares the proposed tuning method with that of 0
0 20 40 60 80 100 120 140 160 180

Lee et al. (1998). Perfect models are assumed for controller time

Fig. 5. Closed-loop responses for example 2.

design. By choosing MS2 = 1.5, the proposed method suggests
the secondary P controller with K c 2 = 0.565 . The tuning 0.2

parameter in the method of Lee et al. (1998) is chosen as 0.1

λ2 = 1.18 so that the same Kc2 is resulted.

0

-0.1

To tune the primary controller, the proposed method -0.2

identifies a SOPTDZ model for the apparent process as -0.3

0.361( 0.634s + 1) −3.578 s

-0.4 Actual

Gm* 1 ( s) = e (36) -0.5

Proposed
Lee et al.
1.698s 2 + 5.34s + 1 -0.6

Choosing MS1 = 1.7, the resulting PID parameters are -0.7

obtained as K c1 = 3.157 , τ I 1 = 5.936 , and τ D1 = 0.766 . On the -0.8

-0.4 -0.2 0 0.2 0.4 0.6 0.8 1

other hand, the method of Lee et al. (1998) uses the following
approximated model for the primary controller design.
 1  1
G* (s) =  e −θ s G ( s ) =
2
 e −3 s
m1
 λ2 s + 1 
1
(1.18s + 1)( 5s + 1)
The tuning parameter λ1 in Lee et al. (1998) is selected so
that the resulting system has the same MS1 with the proposed
system, leading to K c1 = 2.371 , τ I 1 = 7.586 , and τ D1 = 1.999 .
Figure 5 compares the closed-loop responses to a unit step
set-point change at t = 0 followed by unit step disturbance
inputs d2 and d1 at t = 60 and t = 120 , respectively. These
results illustrate the superior performance of the proposed
method. Figure 6 shows the frequency responses of the actual
apparent process and the models considered in the proposed
method and in Lee et al. (1998). The model considered in
Lee et al. (1998) (i.e., (37)) significantly deviates from the
actual apparent process because the secondary controller is a
P controller. This figure reveals that the superior performance
by the proposed method is the result of accurately modeling
the apparent process in the primary controller design.
5. CONCLUSIONS
A novel automatic tuning technique for cascade control
systems has been developed to tune the two PID controllers
simultaneously. Based on a single closed-loop step test, the
process information is identified with the help of B-spline
series representation for the step responses. Distinctive
features of the presented tuning method include: (1) the
secondary controller is designed for enhanced disturbance
performance; (2) the primary controller is designed based on
a process model that accurately accounts for the inner loop
dynamics; (3) the desired robustness levels explicitly guide
the selection of the IMC tuning parameters. The effectiveness
of the proposed method is apparent from the satisfactory
performance shown in the simulation examples.
Fig. 6. Comparison of frequency responses (Nyquist plots) of
the apparent process model for example 2.
(37)
REFERENCES
Alfaro, V. M., R. Vilanova and O. Arrieta (2009). Robust
tuning of two-degree-of-freedom (2-DoF) PI/PID based
cascade control systems, J. Process Control, 19, 1658-
1670.
Chui, C. K. (1997). Wavelets: A Mathematical Tool for
Signal Processing, SIAM, Philadelphia, USA.
Hang, C. C., A. P. Loh and V. U. Vasnani (1994). Relay
feedback auto-tuning of cascade controllers. IEEE Trans.
Control Syst. Technol., 2, 42-45.
Kaya, I., N. Tan and D. P. Atherton (2007). Improved
cascade control structure for enhanced performance, J.
Process Control, 17, 3-16.
Lee, Y., S. Park and M. Lee (1998). PID controller tuning to
obtain desired closed loop responses for cascade control
systems, Ind. Eng. Chem. Res., 37, 1859-1865.
Lee, Y., S. Oh and S. Park (2002). Enhanced control with a
general cascade control structure, Ind. Eng. Chem. Res.,
41, 2679-2688.
Leva, A. and F. Donida (2009). Autotuning in cascaded
systems based on a single relay experiment, J. Process
Control, 19, 896-905.
Mehta, U. and S. Majhi (2011). On-line identification of
cascade control systems based on half limit cycle data,
ISA Trans., 50, 473-478.
Morari, M. and E. Zafiriou (1989). Robust Process Control,
Prentice-Hall, New York, USA.
Tan, K. K., T. H. Lee and R. Ferdous (2000). Simultaneous
online automatic tuning of cascade control for open loop
stable processes, ISA Trans., 39, 233-242.
Veronesi, M. and A. Visioli (2011). Simultaneous closed-
loop automatic tuning method for cascade controllers,
IET Control Theory Appl., 5, 263-270.

419