Article

pubs.acs.org/IECR

Design of Proportional Integral Controllers with Decouplers for

Unstable Two Input Two Output Systems
Sukanya Hazarika and M. Chidambaram*
Department of Chemical Engineering, Indian Institute Of Technology Madras, Chennai 600 036, India

ABSTRACT: A method is proposed based on the equivalent transfer function (ETF) model to design multivariable propor-
tional integral (PI) controllers for unstable multivariable systems with delay. The simplified decoupler decomposes the unstable
multiloop systems into independent loops with ETFs as the resultant decoupled process model having unstable poles. PI
controllers are designed for the diagonal elements of ETFs by the synthesis method meant for unstable first order plus time delay
(FOPTD) systems. Since the overshoot is higher for the unstable systems, a double-loop control structure (inner-loop diagonal
proportional (P) controllers with the decoupler and outer loop only with diagonal PI controllers) is proposed to get a reduced
overshoot. Two examples of two input two output (TITO) unstable systems are considered to demonstrate the simplicity and
effectiveness of the proposed method.

1. INTRODUCTION PI controllers stabilize such systems. Govindhakannan and

Unstable single input single output (SISO) systems with time Chidambaram14 have applied the two stage P−PI controllers
delays are more difficult to control than that of the stable systems. for the unstable systems based on the Tanttu and Lieslehto
The performance specifications such as overshoot and settling method. The method gives significant interactions.
time for such systems are larger for unstable systems than There are three types of decoupling methods reported for
those of the stable systems.1 Several methods for the design of stable systems:15 ideal, simplified, and inverted decoupling. The
PI/PID controllers for SISO unstable systems are available.1 A ideal decoupling method needs to calculate the inverse of the
review of the control of unstable systems is recently given by Rao process transfer function matrix. It results in complicated de-
and Chidambaram.2 To reduce the overshoot, Jacob and coupling elements, and the ideal decoupling is sensitive to
Chidambaram3 have proposed a two stage design of controllers modeling errors. The simplified decoupler has a simple
for unstable SISO systems. First, the unstable system is stabilized decoupler form, but the controller cannot be designed directly
by a simple proportional controller. For the stabilized system, a from the decoupled process model without any model reduction
PI controller is designed. It is shown that an improved per- method. The inverted decoupling method is also sensitive to
formance is obtained. modeling errors. Jevtovic and Matausek16 have proposed an
The method of designing the multivariable PI controllers is optimization method under the constraints on robustness and
complicated due to the interactions among the loops. A review sensitivity to measurement noise based on the ideal decoupler
on the design of decentralized PI controllers for a stable MIMO method.17 Recently several works18−20 appeared for stable
system is given by Luyben and Luyben4 and Maciejowski.5 systems, that introduced the concepts of equivalent transfer
Tanttu and Lieslehto6 and Katebi7 have given simple methods functions/effective open-loop transfer functions (ETFs/EOTFs)
of tuning centralized PI controllers for stable systems. These to take into account the loop interactions in the design of
methods include Davison’s method8 (requires only the steady multiloop control systems.
state gain matrix, SSGM) and the Tanttu and Lieslehto method6 Rajapandiyan and Chidambaram21 recently proposed a
(requires the transfer function matrix). More rigorous methods method of designing controllers for MIMO stable systems
of designing multivariable PI controllers are reviewed by Wang by combining the simplified decoupler approach with the ETF
et al.9 and Wang and Nie.10 model approximation.This method is applicable when the
Only a few methods are reported for the design of controllers decoupler is a stable one. Both the formulation EOTFs and
for unstable multivariable processes. Georgiou et al.11 have
ETFs (decomposed into individual loops) are based on
suggested an optimization method. However the system con-
perfect control approximations and the assumption is fully
sidered does not have a significant time delay. Agamenoni et al.12
have proposed a method of designing controllers based on an validated only by the decouplers. This method gives less
optimization method. In the above methods, the systems con- interactions and better performances when compared to the
sidered have unstable components only for one of the input ideal and inverted decoupling methods. In the present work,
variables. Govindhakannan and Chidambaram13 have applied the applicability of this method to unstable TITO systems is
the method of Tanttu and Lieslehto to unstable MIMO pro- evaluated.
cesses.The interactions are found to be significant. Decentralized
PI controllers are also designed by the detuning method pro- Received: November 9, 2013
posed by Luyben and Luyben.4 The decentralized PI controllers Revised: January 16, 2014
do not stabilize the system if the unstable pole is present in Accepted: February 21, 2014
each of the transfer functions of the system. Only centralized Published: February 21, 2014

© 2014 American Chemical Society 6467 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476
Industrial & Engineering Chemistry Research Article

2. DESIGN BASED ON ETF MODEL ⎡ϕ ϕ ⎤

ϕ = K N ⊗ K N −T = = ⎢ ⎥
11 12
In the two input two output (TITO) system, if the second ⎢⎣ ϕ21 ϕ22 ⎥⎦
feedback controller is in the automatic mode, with yr2 = 0, then (8)
the overall closed-loop transfer function between y1 and u1 is
where ⊗ denotes Hadamard multiplication. RARTA, which is
given by
defined as the ratio of loop yi − uj average residence times, when
y1 (g p,12g p,21gc,2) other loops are closed and when other loops are open, is given by
= g p,11 −
u1 (1 + gc,2g p,22) ⎡ γ11 γ12 ⎤
(1) Γ=ϕ⊙Λ=⎢ ⎥
This can be further written as ⎣ γ21 γ22 ⎦ (9)

y1 g p,12g p,21(gc,2g p,22) Hence for an unstable system, the ETF can be expressed as
= g p,11 − ̂
u1 g p,22(1 + gc,2g p,22) (2) e−θijs
g p,̂ ij (s) = k p,̂ ij
Similarly, for the second loop τiĵ s − 1 (10)

g p,21g p,12(gc,1g p,11) ̂ = kp,ij/Λij, τ̂ij = γijτij, and θ̂ij = γijθij.

where kp,ij
y2
= g p,22 − It should be noted that the method is applicable when EOTF =
u2 g p,11(1 + gc,1g p,11) (3) ETF.
Equations 2 and 3, being complicated, can be simplified by two 3. PAIRING CRITERIA
assumptions:22 the first assumption is the perfect controller
approximation for the other loop (the output attains steady state The problem of loop interaction can be minimized by a proper
with no transient) which is used to simplify both equations; i.e., choice of input output pairings. The degree of interaction is
quantified using the RGA analysis which helps in choosing
gc, ig p, ii
=1 i = 1, 2 the manipulated−controlled variable pairings that best suits
(1 + gc, ig p, ii) the control problem. For stable plants, the pairing is selected
(4)
corresponding to the positive values of NI (Niederlinski index)
The second is assuming ETFs have the same structure of the and RGA, where
corresponding open-loop model. By using the perfect controller
approximation, eqs 2 and 3 can be approximated as det[K ]
NI =
g p,12g p,21 ∏ K ii (11)
eff
y
g p,11 = 1 = g p,11 −
u1 g p,22 (5) ⎡ λ11 λ12 ⎤
RGA(Λ) = K ⊗ K −T = ⎢ ⎥
y2 g p,12g p,21 ⎢⎣ λ 21 λ 22 ⎥⎦ (12)
eff
g p,22 = = g p,22 −
u2 g p,11 (6)
23
The pairing criteria for unstable systems will differ when the
number of open-loop unstable poles of Gp(s) is different from
Here, geff eff
p,11 and gp,22 are the effective open-loop transfer functions G̅p(s) = diag[gp,ii(s)] . Hence pairing is carried out by the follow-
(EOTF). These EOTFs are complicated transfer function
models, and it is difficult to directly use them for the controller ing ways: If for an (n × n) plant (1) with one unstable pole which
design. For the purpose of controller design, resulting EOTFs are appears in all elements of Gp(s), pairing should be such that NI is
reduced to FOPTD models using the Maclaurin series.22 This positive if n is odd and negative if n is even; or (2) with P unstable
method leads to complications in higher dimension systems, in poles which appears in all elements of Gp(s), pairing should be
the formulation of EOTFs, and in the model reduction step. By such that NI is positive if (n − 1)P is even and negative if (n − 1)P
using relative gain array (RGA), relative normalized gain array is odd.
(RNGA), and relative average residence time array (RARTA)
concepts, the expression for ETF can be derived easily for higher 4. DESIGN OF CONTROLLERS
dimension systems also.21 Figure 1 shows the block diagram of the process, simplified
The controllers are designed on the diagonal elements of the decoupler, and the controllers for a TITO process.
ETF, and its closed-loop responses have to be matched with the For the preceding TITO system, we have
closed-loop response of geff eff
p,11 and gp,22. The result of getting
the same closed-loop response would signify that both ETF Y (s) = Gp(s) D(s) U (s) (13)
and EOTF are the same. The basic steps required to obtain ETF
for a TITO system are as follows: The normalized gain KN,ij is ⎡ g p,11 g p,12 ⎤⎡1 d12 ⎤ ⎡ g p,11 * 0 ⎤
⎢ ⎥
defined as Gp(s) D(s) = ⎢ ⎥⎢ ⎥=
⎢⎣ g p,21 g p,22 ⎥⎦⎢⎣ d 21 1 ⎥⎦ ⎢⎢ 0 *
⎥
g p,11⎥⎦
⎡ KN,11 KN,12 ⎤ ⎣
K N = K ⊙ Tar = ⎢ ⎥ (14)
⎢⎣ KN,21 KN,22 ⎥⎦ (7)
where the simplified decoupler is designed as
where ⊙ denotes Hadamard division, K is the steady state gain,
and Tar = τij + θij is defined as the average residence time which g p,12(s) g p,21(s)
signifies the response speed of the controlled variable yi to d12(s) = − ; d 21(s) = −
g p,11(s) g p,22(s) (15)
manipulated variable uj. The RNGA can be obtained as
6468 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476
Industrial & Engineering Chemistry Research Article

Figure 1. Simplified decoupled control system of a TITO process. (d11 = 1;d22 = 1).

For these systems, having time delay may lead to unwanted ⎡ 3.6 4 ⎤ ⎡ 0.444 0.1500 ⎤
situations (nonrealizable cases) by eq 15. Hence, an extra time Tar = ⎢ ⎥; KN = ⎢ ⎥;
⎣ 4.5 3.2 ⎦ ⎣ 0.1556 0.5313 ⎦
delay (θ) is to be incorporated into the decoupler matrix which is
further added to the corresponding ETF.21 For any uncertainty ⎡1.1097 −0.1097 ⎤ ⎡ 0.9383 0.6005⎤
ϕ=⎢ ⎥; Γ = ⎢ ⎥
in time delay, time constant, or process gain, the robustness of ⎣−0.1097 1.1097 ⎦ ⎣ 0.6005 0.9383⎦
the closed-loop system is to be evaluated by simulation. In the (21)
presence of the decoupler, the TITO system behaves like two
independent loops for which the controllers can be designed By using the preceding concepts, the ETF matrix is obtained as
independently. In the present work, diagonal PI controllers are ⎡ 1.3529e−0.9383s −3.2857e−0.9008s ⎤
designed by the synthesis method1 based on the unstable ETFs: ⎢ ⎥
⎢ ( −2.4396s + 1) (1.5013s + 1) ⎥
⎛ 1 ⎞ Ĝp(s) = ⎢ ⎥
gc, ii(s) = kc, ii⎜⎜1 + ⎟⎟ ⎢ −3.8333e
−0.9008s
1.4375e−0.9383s ⎥
⎝ τI, ii(s) ⎠ (16) ⎢⎣ (1.8016s + 1) ( −2.0643s + 1) ⎥⎦ (22)
kc, iik p, ii = 0.8668εii−0.8288 for 0.1 ≤ εii ≤ 0.7 (17) According to eq 15, the decoupler is given by
τI, ii ⎡ (7.8s − 3)e−0.5s ⎤
= 0.1523e7.9425εii for 0 ≤ εii ≤ 0.7 ⎢1 ⎥
τii (18) ⎢ (20s + 8) ⎥
D(s) = ⎢ ⎥
where εij = θii/τii. ⎢ (15.4s − 7)e
−0.5s
⎥
⎢⎣ (51s + 17) 1 ⎥⎦
(23)
5. SIMULATION EXAMPLES
To demonstrate the ETF method, two simulation examples of The PI controllers for the diagonal elements of ETF are calcu-
unstable systems are considered. lated by eqs 16−18 as
5.1. Example 1. Consider an example given by Flesch et al.,24 ⎡ ⎤
⎛ 1 ⎞
which has diagonal elements of unstable and unequal poles: ⎢−1.3797⎜1 + ⎟ 0 ⎥
⎢ ⎝ 8.4s ⎠ ⎥
⎡ 1.6e−s 0.6e−1.5s ⎤ Gc(s) = ⎢
⎢ ⎥ ⎢0
⎛ 1 ⎟ ⎥⎥
⎞
⎢ ( −2.6s + 1) (2.5s + 1) ⎥ −0.98⎜1 +
Gp(s) = ⎢ ⎣ ⎝ 11.62s ⎠ ⎦
−1.5s ⎥
⎢ 0.7e 1.7e−s ⎥ (24)
⎢⎣ (3s + 1) ( −2.2s + 1) ⎥⎦ (19) The actual EOTF models for Gp(s)D(s) can be derived as
According to section 3, the pairing criteria for this system will eff
y1 1.6e−s 0.6 ⎡ 15.4s − 7 ⎤ −2s
g p,11 = = + ⎢ ⎥e
be the same as that for a stable one since the number of open- u1 ( −2.6s + 1) (2.5s + 1) ⎣ 51s + 17 ⎦
loop unstable poles is the same for both Gp(s) and G̅p(s) = (25)
diag[gp,ii(s)]. To determine the pairing for this system, K and
RGA are calculated as eff
y2 1.7e−s 0.7 ⎡ 7.8s − 3 ⎤ −2s
g p,22 = = + ⎢ ⎥e
u2 ( −2.2s + 1) (3s + 1) ⎣ 20s + 8 ⎦
⎡1.6 0.6 ⎤ ⎡1.1826 − 0.1826 ⎤
K=⎢ ⎥; RGA(Λ) = ⎢ ⎥ (26)
⎣ 0.7 1.7 ⎦ ⎣−0.1826 1.1826 ⎦
As stated earlier, the PI controllers are designed based on ETF.
(20) The corresponding closed-loop responses are obtained by simu-
Since NI = 0.6969 is positive as calculated by eq 11, the pairing lation . The same controller settings are applied on EOTF, and
can be kept as it is. The dynamic elements such as normalized the closed-loop responses are found to be same. Hence the
gain matrix (KN), RNGA (ϕ), average residence time (Tar), and reduced model (ETF) depicts the EOTF. Figure 2 shows the
RARTA (Γ) to obtain ETF matrix are calculated by using eqs 7,8, performance of the controlled system for a step change in set
9, and 10: point y1 and y2 separately. The overshoot is found to be large
6469 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476
Industrial & Engineering Chemistry Research Article

Figure 2. Performance of the PI controllers with decouplers for a step input in y1 and y2 separately.

Figure 3. Responses of the manipulated variable for a step input in y1 and y2 separately.
′
which is the typical performance of any unstable system with k p,′ ii e−θiis
delay. However, no interactions among the loops are observed. Gp,′ ii =
Figure 3 shows the manipulated variable versus time behavior τii′s + 1 (28)
corresponding to Figure 2. In order to reduce the overshoot, a Here k′p,ii is the steady state value reached, θ′ii is the initial value
double-loop control structure is applied. As shown in the block noted, and τ′ii = t − θ′ii where t is the time taken to (0.63k′p,ii) from
diagram of Figure 4, first the decouplers are to be used. The Figure 5. The fitted parameters are kp,11 ′ = 2.628, τ11 ′ = 7, kp,22
′ =
3.225, τ22
′ = 7, and θ11 ′ = θ22
′ = 1 . For the stabilized FOPTD
system, diagonal PI controllers by the synthesis method are
designed and placed in the outer loop.

1 ⎛ τii′ ⎞ 0.5
kc, ii = ⎜⎜ + θii⎟⎟ ≈ ; τI, ii = τii′
k p,′ ii ⎝ τc, ii ⎠ k p,′ ii (29)
Hence the obtained diagonal PI controller matrix for the outer
Figure 4. Block diagram for the two stage control scheme: Gp, loop is given by
multivariable system; GD, decouplers; Gc,I, inner-loop stabiliz-
ing decentralized proportional controllers; Gc,O, outer-loop decentral- ⎡ ⎛ 1⎞ ⎤
ized PI controllers; Y, output variables; Yr, set point values; V, load ⎢ 0.1902⎜1 + ⎟ 0 ⎥
⎢ ⎝ 7s ⎠ ⎥
variables. Gc(s) = ⎢
⎢0
⎛ 1 ⎞⎥
0.155⎜1 + ⎟ ⎥
⎣ ⎝ 7s ⎠ ⎦ (30)
resulting system is to be then stabilized by diagonal proportional
controllers which are designed on the diagonal elements of ETF Figure 6 shows that the responses of the double-loop controller
by using this formula25 are much superior to that of a single-loop PI controller. The
1 interactions are found to be negligible (not shown in Figure 6).
k p, iikc, ii = 0.5 Figures 7 and 8 show the responses of single- and double-loop
εii (27)
controllers when the load variables v1 and v2 enter the system
The proportional controller tuning parameters are kc,11 = −1.19 along with the inputs u1 and u2. Stable responses with high
and kc,22 = −1.007. Further for the stabilized inner-loop system, oscillations and large interactions are observed for the single-
diagonal PI controllers are designed based on the identified loop systems, whereas the double-loop method reduces the
simple FOPTD model. Figure 5 shows the step responses of the overshoots and eliminates the interactions.The robustness of the
single loop alone. Based on the above step response simple double-loop system is evaluated by perturbing each process gain
FOPTD models are fitted: by +10% of the normal value in the process.The same controller
6470 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476
Industrial & Engineering Chemistry Research Article

Figure 5. Responses of single-loop P controllers with decouplers for a step input in y1 and y2 separately.

Figure 6. Comparison of servo responses of double-loop and single-loop multivariable PI controllers with the decoupler. (No interactions for double-
loop system.

Figure 7. Responses and interactions of the single-loop PI controllers with decoupler for regulatory problems: (left) v1 = 1, v2 = 0; (right) v1 = 0, v2 = 1.

Figure 8. Comparison of responses of double-loop and single-loop multivariable PI controllers with the decoupler for regulatory problems
(no interactions for the double-loop system): (left) v1 = 1, v2 = 0; (right) v1 = 0, v2 = 1).

6471 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476

Industrial & Engineering Chemistry Research Article

Figure 9. Robustness of double-loop controllers for perturbation in each kp (10%) in the process: (top panels) servo problem; (bottom panels)
regulatory problem.

Table 1. Comparison of IAE and ISE Values for Perturbations in the Model Parameters for the Double-Loop Control System
(Example 1)
servo (y1) servo (y2) regulatory (y1) regulatory (y2)
perturbation IAE ISE IAE ISE IAE ISE IAE ISE
kp 13.98 6.378 16.77 7.664 36.4 29.7 53.08 63.2
1.1kp 16 7.071 14.11 6.7 36.2 27.9 52.6 58.6
0.9kp 11.86 5.62 11.92 5.69 36.7 32.5 54.28 71.49
θ 13.98 6.38 14.11 6.7 36.47 29.74 53.08 63.27
1.1θ 13.98 6.43 14.35 6.8 36.46 31.97 53.6 73.62
0.9θ 13.99 6.37 14.13 6.69 36.4 28.74 53.1 60.37
τ 13.98 6.37 14.11 6.7 36.45 31.16 53.08 63.27
1.1τ 14.02 6.5 14.2 6.8 36.49 29.12 53.14 61.04
0.9τ 13.98 6.2 14.2 6.71 36.47 29.74 53.29 71.43

setting and decouplers are used. Similar studies are also carried The NI for this 2 × 2 system is 0.5833, and hence the columns are
out for −10% perturbations. The robust performance is shown in interchanged to get the correct pairing. Therefore the newly paired
Figure 9. Similar studies are carried out for perturbations in time
delay (θ) and time constant (τ) separately, and the results are system is given by
given in Table 1 in terms of IAE/ISE values.
5.2. Example 2. Consider the example given by ⎡ −1e−s −1.6667e−s ⎤
⎢ ⎥
Govindhakannan and Chidambaram,13,14 containing all elements ⎢ ( −1.6667s + 1) ( −1.6667s + 1) ⎥
unstable with equal poles: Gp(s) = ⎢
⎢ −1.6667e
−s
−0.8333e−s ⎥⎥
⎡ −1.6667e−s −1e−s ⎤ ⎢⎣ ( −1.6667s + 1) ( −1.6667s + 1) ⎥⎦ (33)
⎢ ⎥
⎢ ( −1.6667s + 1) ( −1.6667s + 1) ⎥
Gp(s) = ⎢
−1.6667e−s ⎥⎥
−s The normalized gain matrix (KN), RGA (Λ), RNGA (ϕ), average
⎢ −0.8333e
⎢⎣ ( −1.6667s + 1) ( −1.6667s + 1) ⎥⎦ residence time (Tar) and RARTA (Γ) are calculated as
(31)
⎡− 0.4283 1.4283 ⎤ ⎡ 2.6667 2.6667 ⎤
Since the number of open-loop unstable poles of Gp(s) is Λ=⎢ ⎥; Tar = ⎢ ⎥;
⎣1.4283 − 0.4283⎦ ⎣ 2.6667 2.6667 ⎦
different from G̅p(s) = diag[gp,ii(s)], the pairing criteria for this
⎡− 0.375 − 0.625 ⎤ ⎡− 0.4283 1.4283 ⎤
system will differ from that of a stable system. Hence, the pairing KN = ⎢ ⎥; ϕ = ⎢ ⎥;
is carried out according to point 1 of section 3: ⎣− 0.625 − 0.3124 ⎦ ⎣1.4283 − 0.4283⎦
⎡1 1⎤
⎡−1.6667 − 1 ⎤ ⎡1.4283 − 0.4283⎤ Γ=⎢
⎣1 1⎦⎥ (34)
K=⎢ ⎥; Λ=⎢
⎣−0.8333 − 1.6667 ⎦ ⎣−0.4283 1.4283 ⎥⎦
(32) The ETF model matrix is given by
6472 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476
Industrial & Engineering Chemistry Research Article

Figure 10. Performance of the PI controllers with decouplers for a step input in y1 and y2 separately (example 2).

Figure 11. Responses of a single loop with the set point filter (example 2).

Figure 12. Responses of single-loop P controllers with decouplers for a step input in y1 and y2 separately.

⎡ 2.3348e−s −1.16711e−s ⎤ The obtained diagonal PI controller matrix by the synthesis

⎢ ⎥ method is given by
̂ ⎢ ( −1.6667s + 1) ( −1.6667s + 1) ⎥
Gp(s) = ⎢
⎢ −1.16711e
−s
1.9456e−s ⎥⎥ ⎡ ⎛
⎢−0.5681⎜1 +
1 ⎞
⎟ 0
⎤
⎥
⎢⎣ ( −1.6667s + 1) ( −1.6667s + 1) ⎥⎦ ⎢ ⎝ 24.3798s ⎠ ⎥
(35) Gc(s) = ⎢
⎛ ⎞ ⎥
⎢0 1
−0.68⎜1 + ⎟⎥
The simplified decoupler matrix is given by ⎢⎣ ⎝ 24.3798s ⎠ ⎥⎦ (38)
⎡1 −1.6667 ⎤
D(s) = ⎢ ⎥ Figure 10 shows the responses of the single-loop PI controller
⎣−2.001 1 ⎦ (36) with the decoupler. The overshoots are large but with negligible
The combined model Gp(s)D(s) is the same as that obtained by interactions. To reduce the overshoot, we use a set point filter
the ETF method. with τI,ii as the time constant of the filter. The designed filter is
given as follows: 1/(1 + τI,iis) where τI,ii is the integral time
eff * = 2.3348e−s constant of the diagonal PI controller. The performances of
g p,11 = g p,11 ;
( −1.6667s + 1) the single-loop system with the set point filters are shown in
Figure 11. The overshoot is reduced. However, the set point filter
eff * = 1.9456e−s
g p,22 = g p,22 can reduce overshoot only for a servo problem. In order to re-
( −1.6667s + 1) (37) duce overshoot for a regulatory problem, the double-loop control
6473 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476
Industrial & Engineering Chemistry Research Article

Figure 13. Comparison of servo responses of double-loop and single-loop multivariable PI controllers with the decoupler (no interactions for the
double-loop system.).

Figure 14. Responses and interactions of the single-loop PI controllers with decoupler for regulatory problems: (left) v1 = 1, v2 = 0; (right) v1 = 0, v2 = 1.

Figure 15. Comparison of double-loop and single-loop multivariable PI controllers with the decoupler for regulatory problems (no interactions for the
double-loop system): (left) v1 = 1, v2 = 0; (right) v1 = 0, v2 = 1).

Table 2. Comparison of IAE and ISE Values for Perturbations in the Model Parameters for the Double-Loop Control System
(Example 2)

servo (y1) servo (y2) regulatory (y1) regulatory (y2)

perturbation IAE ISE IAE ISE IAE ISE IAE ISE
kp 8 3.78 7.99 3.78 27.48 41.41 22.95 28.8
1.1kp 10.49 4.65 10.5 4.66 27.46 35.08 22.93 24.4
0.9kp 4.94 3.12 4.93 3.12 27.48 57.35 22.95 39.93
θ 8 3.78 7.99 3.78 27.48 41.41 22.95 28.8
1.1θ 8 3.89 7.99 3.89 27.62 50.3 23.04 34.98
0.9θ 8 3.80 8 3.8 27.48 37.95 22.95 26.42
τ 8 3.78 7.99 3.78 27.48 41.41 22.95 28.82
1.1τ 8 3.934 8 3.93 27.48 38.98 22.95 27.14
0.9τ 8 3.75 7.99 3.756 27.48 49.19 22.94 34.12

6474 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476

Industrial & Engineering Chemistry Research Article

Figure 16. Robustness of double-loop controllers for perturbation in each kp (10%) in the process: (top panels) servo problem; (bottom panels)
regulatory problem.

structure is required. The design of a double-loop controller is

carried out as stated in Example 1. Also given are the responses of
■ AUTHOR INFORMATION
Corresponding Author
single-loop P controllers with decouplers in Figure 12. *E-mail:chidam@iitm.ac.in.
The proportional controller tuning parameters are kc,11 =
Notes
−0.55 and kc,22 = −0.66. The fitted parameters of the simplified
The authors declare no competing financial interest.

■
FOPTD model for the inner loop are obtained as follows: k′p,11 =
4.435, τ11 ′ = 4.444, τ22
′ = 4, kp,22 ′ = 4, and θ11
′ = θ22
′ = 1. The diagonal NOMENCLATURE
PI controller matrix by the synthesis method is calculated as
dij = decoupler elements
⎡ ⎛ 1⎞ ⎤ EOTF = effective open-loop transfer function
⎢ 0.112⎜1 + ⎟ 0 ⎥
⎢ ⎝ 4s ⎠ ⎥
ETF = equivalent transfer function
Gc(s) = ⎢ ⎥ FOPTD = first order plus time delay
⎛ 1⎞ Gp, Gc = process and controller transfer function matrixes
⎢0 0.112⎜1 + ⎟ ⎥
⎣ ⎝ 4s ⎠ ⎦ (39) gp,ij, gc,ij, ĝij = process, controller, and equivalent transfer
function models
Figures 13 and 14 show the responses of the single- and the geff * = effective open-loop transfer function and decoupled
p,ij, gp,ii
double-loop systems for step changes in the load variables v1 and process model
v2 entering along with u1 and u2. The double-loop system gives an IAE = integral of absolute error
improved performance. The robustness of the double-loop con- ISE = integral of the squared error
trol system is studied by perturbing each gain, time delay, and K = process steady state gain
time constant in the process as explained earlier in the previous KN = normalized gain matrix
simulation example. Figure 15 shows the response when the KNij = normalized gain
uncertainty is given in the process gain. Table 2 shows the IAE/ kc,ii = diagonal proportional gain
ISE values for the above robustness studies (Figure 16) for the ̂ = process and effective steady state gain
kp,ij, kp,ij
servo and regulatory problems. NI = Niederlinski index
P = proportional
6. CONCLUSION PI = proportional integral
Based on the equivalent transfer function (ETF) model, multi- RARTA = relative average residence time array
variable PI controllers are designed for unstable multivariable RGA = relative gain array
systems with time delay. The method uses simplified decouplers RNGA = relative normalized gain array
which decompose the unstable multiloop systems into inde- s = Laplace domain
pendent loops with ETFs as the resulting decoupled process SSGM = steady state gain matrix
model having unstable poles. This step is followed by the design t = time
of PI controllers for the diagonal elements by the synthesis TITO = two input−two output
method. Since the overshoot is found to be larger, a double-loop uj = manipulated variable
control structure is proposed to reduce the overshoot. To vk = load variable
demonstrate the simplicity and effectiveness of the proposed Y = closed-loop process response vector
method, two examples of two input two output TITO unstable yi = closed-loop process response
systems, one with mild interactions and another with large Γ = relative average residence time array (RARTA)
interactions, are given. θij, θ̂ij = process and effective time delay
6475 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476
Industrial & Engineering Chemistry Research Article

Λ = relative gain array (RGA) (23) Hovd, M.; Skogestad, S. Pairing criteria and decentralized control
τij, τ̂ij = process and effective time constant of unstable plant. Ind. Eng. Chem. Res. 1994, 33, 2134.
τI,ii = diagonal integral time constant (24) Flesch, R. C. C.; Torrico, B. C.; Normen-Rico, J. E.; Cavalcante,
τIi = integral time M. U. Unified aapproach for minimal output dead time compensation in
MIMO processes. J. Process Control 2011, 21, 1081.
ϕ = relative normalized gain array (RNGA) (25) DePaor, A. M.; OMalley, M. Controllers of Ziegler-Nichols type
Γij = relative average residence time for unstable process with time delay. Int. J. Control 1989, 49, 1273.
ϕij = element of relative normalized gain
Λij = relative gain array element
Subscripts
i, j= loop representation

■ REFERENCES
(1) Padmasree, R.; Chidambaram, M. Control of Unstable Systems, 1st
ed.; Narosa Publishing House: New Delhi, India, 2006.
(2) Rao, A. S.; Chidambaram, M. PI/PID Controllers Design for
Integrating and Unstable Systems. In PID control in the third Milleneium;
Vilanova, R., Visioli, A., Eds.; Springer Verlag: London, 2012; p 75 .
(3) Jacob, E. F.; Chidambaram, M. Control of unstable delay plus first
order systems. Comput. Chem. Eng. 1996, 20, 579.
(4) Luyben, W. L.; Luyben, M. L. Essentials of Process Control; McGraw-
Hill: Singapore, 1997.
(5) Maciejowski, J. M. Multivariable Feedback design; Addison-Wesley:
New York, 1989.
(6) Tanttu, J. T.; Lieslehto, J. A. Comparative study of some
multivariable PI controller tuning methods. In Intelligent tuning and
adaptive control methods; Devanathan, R., Ed.; Pergamon: Oxford, U.K.,
1991; p 357.
(7) Katebi, R. Robust multivariable tuning methods. In PID Control in
the Third Millennium, Vilanova, R., Visioli, A., Eds.; Springer Verlag:
London, 2012; p 255.
(8) Davison, E. J. Multivariable tuning regulators: The feed-forward
and robust control of general servo-mechanism problem. IEEE Trans
Autom. Control 1976, AC-21, 35.
(9) Wang, Q.-G.; Ye, Z.; Cai, W.-J.; Hang, C.-C. PID Control for
Multivariable Processes; Springer-Verlag: Berlin, Heidelberg, 2008.
(10) Wang, Q.-G.; Nie, Z.-Y. PID Control for MIMO Processes. In
PID Control in the Third Milleneium; Vilanova, R., Visioli, A., Eds.;
Springer-Verlag: London, 2012; p 177.
(11) Georgiou, A.; Georgakis, C.; Luyben, W. L. Control of a
multivariable open loop unstable process. Ind. Eng. Chem. Res. 1989, 28,
1481.
(12) Agamenoni, O. E.; Deagas, A. C.; Romagnoli, J. A. A multivariable
delay compensation scheme. Chem. Eng. Sci. 1992, 47, 1173.
(13) Govindhakannan, J.; Chidambaram, M. Multivariable PI control
of unstable systems. Process Control Qual. 1997, 10, 319.
(14) Govindhakannan, J.; Chidambaram, M. Two stage multivariable
PID controllers for unstable plus time delay systems. Indian Chem. Eng.
2000, 42, 34.
(15) Gagnon, E.; Pommerleau, A.; Desbiens, A. Simplified, Ideal or
Inverted Decoupling? ISA Trans. 1998, 37, 265.
(16) Jevtovic, B. T.; Matausek, M. R. PID controller design of TITO
system based on ideal decoupler. J. Process Control 2010, 20, 869.
(17) Nordfeldt, P.; Hagglund, T. Decoupler and PID controller design
of TITO systems. J. Process Control 2006, 16, 923.
(18) Huang, H. P.; Jeng, J. C.; Chiang, C. H.; Pan, W. A. Direct method
for multi loop P/PID controller design. J. Process Control 2003, 13, 769.
(19) Xiong, Q.; Cai, W. J. Effective transfer function matrix method for
decentralized control system design of multi-input multi-output
processes. J. Process Control 2006, 16, 773.
(20) Shen, Y.; Cai, W. J.; Li, S. Multivariable Process Control:
Decentralized, Decoupling or Sparse? Ind. Eng. Chem. Res. 2010, 49, 761.
(21) Rajapandiyan, C.; Chidambaram, M. Controller design for
MIMO processes based on simple decoupled equivalent transfer
functions and simplified decoupler. Ind. Eng. Chem. Res. 2012, 51, 12398.
(22) Vu, T. N. L.; Lee, M. Independent design of multi loop PI/PID
controllers for interacting multivariable processes. J. Process Control
2010, 922.

6476 dx.doi.org/10.1021/ie403791q | Ind. Eng. Chem. Res. 2014, 53, 6467−6476