This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY 1

Decentralized PI Controller Design Based on Phase Margin Specications

Petr Huek
Abstract An effective method for design and tuning of decentralized PI controllers for stable multi-input, multi-output systems is presented in this brief. The direct Nyquist array is applied to shape the Gershgorin bands for each loop separately such that they pass through a specied point corresponding to phase margin specication that is used as a single tuning parameter. The procedure is applied on the control of a laboratory model of quadruple-tank system. Index Terms Decentralized control, multivariate systems, Nyquist stability analysis, phase margin, quadruple-tank process.

I. I NTRODUCTION

ESPITE the signicant progress in applicability of advanced control techniques resulting in a growing number of industrial implementations, the proportionalintegral-differential (PID) controllers remain the most popular controllers used in process control. The reason for that consists in their simple implementation, high reliability and robustness, and sufcient performance in the cases when the closed-loop requirements are not too strong[1]. Unfortunately, even though consisting of three parameters only their tuning is not trivial even for single-input, single-output (SISO) systems [2], [3] and consequently plenty of them are tuned poorly. However, most of the plants encountered in industry contain more closed loops. The multivariate systems can be successfully treated by classical control approaches. The resulted state space or full transfer function matrix controllers are nevertheless of a high order and thus difcult for nal tuning. For severe interactions, the decoupling control schemes are preferable [4][7]. When the interactions among different loops are moderate, decentralized PID control achieves satisfactory behavior. Moreover, even if advanced strategies, such as increasingly applied model predictive control are used, PID controllers are still implemented at the lowest level loops. Even though the SISO design procedures cannot be applied directly due to process interactions analysis and synthesis become much easier than in the case of full dimensional controller. A comprehensive survey of design methods can be found in [8]. For multiloop PID control design, several methods have been developed based on frequency-domain analysis. A common practice consists in using SISO techniques for diagonal elements disregarding the other ones followed by a detuning decreasing the coefcients by some ad hoc factor such that

Manuscript received January 28, 2012; accepted February 4, 2013. Manuscript received in nal form February 16, 2013. Recommended by Associate Editor C. Lagoa. The author is with the Department of Control Engineering, Faculty of Electrical Engineering, Czech Technical University in Prague, Prague 166 27, Czech Republic (e-mail: husek@fel.cvut.cz). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TCST.2013.2248060

the inuence of the other loops will not affect the main loops too much [9]. The most well-known method using detuning control scheme is the biggest log modulus tuning [10]. If a model of the controlled MIMO plant is not available, methods adopting relay feedback tests resulting in modications of the ZieglerNichols (ZN) tuning rules can be applied [11][14]. Chen and Seborg [15] modied the ZN rules using localization of the stability region in the controller parameter space. A detailed analysis of generalization of ultimate quantities to MIMO systems was elaborated by Campestrini et al. [16]. Another approach relies on extension of Nyquist stability criterion to MIMO systems [17], [18]. The direct Nyquist array ensures closed-loop stability and performance by an appropriate shaping of the Gershgorin bands, the MIMO version of the Nyquist curve. For a second-order plus deadtime model, Ho et al. [19], [20] derived analytic formulas to guarantee gain and phase margin specications for Gershgorin bands. An iterative tuning using the structural decomposition of n n MIMO system to n SISO ones is used in [21] and [22]. A similar idea allowing independent design of diagonal controllers called equivalent subsystems method is presented in [23]. Garcia et al. [24] minimized a frequency criterion for each independent loop weighting the deviations of the innity norms of sensitivity and complementary sensitivity functions as well as crossover frequency from the specied values. This brief presents a multiloop PI control tuning method that shapes the Gershgorin bands such that they do not contain the point ej M where the angle M is a tuning parameter. The motivation comes from SISO systems controller design where such angle corresponds to phase margin that often serves as a closed-loop step response tuning parameter since it is related to its damping. All the PI controller candidates for each loop are plotted as a curve in the kP kI plane and the ones with the highest integral part are chosen since they minimize the sum of integral errors. The presented method is implemented on control of a laboratory model of quadruple-tank process [25], [26]. The main advantage of the presented procedure consists in its simplicityit contains only one tuning parameter and the computation is one-step and simple. On the other hand, the method does not allow ne tuning and is not suitable for systems with strong interactions. II. D IRECT N YQUIST A RRAY D ESIGN Consider an n n system G (s ) = [gi j (s )]nn controlled by a decentralized controller C (s ) = diag{c1 (s ), c2 (s ), . . . , cn (s )}, see Fig. 1.

10636536/$31.00 2013 IEEE

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

1.5

0.5

Fig. 1.

Decentralized control scheme.

0 M 0.5 Im 1

The individual control loops are paired according to relative gain analysis (RGA) [18]. The following denote: L (s ) = G (s )C (s ) H (s ) = ( I + G (s )C (s ))1 G (s )C (s ) (1) (2)

the open-loop and closed-loop transfer matrix, respectively. Consider the Nyquist plot of gkk (j)ck (j) with a superimposed circle of the radius
n

1 lmm(j ) 1.5 1.5 1 0.5 0 Re 0.5 1 1.5

|gkm (j)cm (j)|.

k =1,k =m

(3)

This circle is referred to as the Gershgorin circle. The whole band composed of the circles for all > 0 is called the Gershgorin band. Stability of the closed loop can be tested by the following theorem. Theorem 1 (Direct Nyquist Array (DNA) [17], [18]): Let the Gershgorin bands be centered on the diagonal elements lmm (j) of L (j), and m = 1, . . . , n exclude the point (1 + j0) (the transfer matrix L (s ) is called column diagonally dominant). Let the i -th Gershgorin band encircles the point (1 + j0) Ni times anticlockwise. Then, the closed-loop transfer matrix H (s ) is stable if and only if
n

Fig. 2.

Phase margin for Gershgorin band.

Ni = p0
i =1

(4)

where p0 is the number of unstable poles of L (s ). Since most practical processes are open-loop stable, p0 = 0 is assumed in this brief. In that case the closedloop transfer matrix H (s ) with column diagonally dominant open-loop transfer matrix L (s ) is stable if and only if Nyquist plots of lmm (j) do not encircle the point (1 + j0) for all m = 1, . . . , n . Remark 1: The stability analysis by direct Nyquist array represents sufcient condition only, i.e., if some of the Gershgorin bands contains the critical point nothing can be said about stability or instability of the closed loop. Remark 2: To diminish the interactions as much as possible, a constant decoupling compensator can be inserted between the controller and the plant. The frequency on which the compensator will act is usually chosen zero or equal to the bandwidth of the closed loop. III. P HASE M ARGIN S PECIFICATION R EGION FOR D ECENTRALIZED PI C ONTROL In compliance with the phase margin denition for SISO systems, let us try to shape the Gershgorin bands such that they pass through the point ejM in the complex plane

(point A in Fig. 2) where the value of M is usually chosen between 0 and /2. Let us remind that the extension of phase margin to MIMO systems is not straightforward because of complexity of matrix perturbations that result in many different denitions. The definition employed here [27] corresponding to perturbations of characteristic loci is simple for computation but it requires the diagonal dominance of the system. The most general denition formulated by Ba-ron and Jonckheere [28] considers arbitrary unitary matrix perturbations in the feedback path. However, from practical reasons, it is not used for a controller design because it is tough to imagine its relation to phase changes in individual loops. The most useful denition, proposed in [29], takes into account only diagonal phase perturbations but even for this simplication, the phase margin computation requires solving a constraint optimization task. Although a closed-loop system satisfying phase margin specication used in this brief may not be stable, it is widely accepted by the practitioners since such a case does is unlikely to occur in practical applications. The condition can be expressed as |ejM + gmm (j)cm (j)|
n

|gkm (j)cm (j)| , m (5)

k =1,k =m

with the equality sign holding for just one = gm for each m = 1, . . . , n corresponding to the touching point. Since, according to the assumption, the diagonal elements are all stable, one can always nd a sufciently small proportional controller such that inequality 5 holds for every . Since 0 < M /2 for a stable gmm (s ) of at least second order, one can nd a frequency = gm such that gmm (jgm ) = (ej M ) = + M and thus by an appropriate proportional controller, one can make the left-hand side of (5) arbitrarily small (even zero) to force the

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
HUEK: DECENTRALIZED PI CONTROLLER DESIGN
3

4 3.5 3 2.5 2 1.5 1 0.5 k

I

x 10

=0
M M M M

=15 =25 =40

0 0

0.1

0.2

0.3 k
P

0.4

0.5

0.6

Fig. 4.

PI controllers with phase margin specications.

x 10 8
3

M=0
M M M

Fig. 3.

Principal scheme of quadruple-tank system.

7 6

=15 =25

TABLE I
k

=40

N ONMINIMUM P HASE S YSTEM PARAMETER VALUES A1 , A3 [cm 2 ] A2 , A4 [cm 2 ] a1 , a3 g [cm 2 ] [cm /s2 ] a2 , a4 [cm 2 ]
0 h0 1 , h 2 [cm ] 0 h3, h0 4 [cm ] 0 u0 , u 1 2 [V] k1 , k2 [cm 3 /Vs]

5
I

25 30 0.084 0.065 981 (11.4 12) (4.99 4.72) (4.15 4.41) (2.32 2.58) (0.35 0.27) Fig. 5.

4 3 2 1 0 0 0.2 0.4 0.6 k

P

0.8

PI controllers with phase margin specications. TABLE II

1 , 2 []

equality. For a rst order gmm (s ), a purely integral controller will do the same job. Thus, at least one PI controller solving (5) exists for every 0 < M /2 and each m = 1, . . . , n . Let us write each element of the plant transfer matrix gkm (j) (possibly with time delay) as gkm (j) = akm () + jbkm () = rkm ()ejkm () and use a decentralized PI controller C (s ) {c1 (s ), . . . , cn (s )} c m (s ) = k P m + to control the plant. With
n

D ECENTRALIZED PI C ONTROLLERS C ONSTANTS M [] 0 15 25 40 CS [15] kP1 0.5342 0.3944 0.3301 0.2651 0.4958 kI1 0.0082 0.0050 0.0037 0.0023 0.0040 kP2 0.2754 0.2351 0.2119 0.1786 0.2667 kI2 0.0032 0.0023 0.0017 0.0012 0.0016

(6) diag (5) becomes

2 2 2 2 kP m rmm () Rm () + k Im 2 () R 2 () rmm m 2 + 2kPm (amm () cos M + bmm () sin M ) amm () sin M bmm () cos M + 1 0. (9) + 2k I m

k Im , s

m = 1, . . . , n

(7)

Rm () =
k =1,k =m

|gkm (j)|,

m = 1, . . . , n

(8)

and after some algebraic manipulations, the inequality

All the PI controllers with kPm > 0, kIm > 0 of the m th loop satisfying (9) for 0 < form a region in kP kI plane for which the corresponding Gershgorin band does not contain the point ejM . The controllers lying on the boundary of that region guarantee that the Gershgorin band passes through that point.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4
12.3 12.2 12.1 12

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

12.3 12.2 12.1 12 11.9 y [cm]

y [cm] 1

11.9 11.8 11.7 11.6 11.5 11.4 0 = 0

M M M M

11.8 11.7 = 0
M M M M

= 15

11.6 11.5 11.4

4000

= 15 = 25

= 25 = 40

= 40 500 1000 1500 2000 t[s] 2500 3000 3500 4000

500

1000

1500

2000 t[s]

2500

3000

3500

11.3 0

Fig. 6.

Model responses of level in tank 1 for step setpoint changes.

Fig. 8.

Real system responses of level in tank 1 for step setpoint changes.

12.8 12.7 12.6 12.5

y [cm]

12.8 12.7 12.6 12.5 12.4 y [cm] M = 0

M 2

12.4
2

12.3 12.2 12.1 12 11.9 11.8 0 500 1000 1500 2000 t[s] 2500 3000

12.3 12.2 12.1 12 11.9

=0
M M M M

= 15 M = 25 = 40
M

= 15 = 25

= 40

3500

4000

11.8 0

500

1000

1500

2000 t[s]

2500

3000

3500

4000

Fig. 7.

Model responses of level in tank 2 for step setpoint changes.

Fig. 9.

Real system responses of level in tank 2 for step setpoint changes.

According to the DNA theorem, in order to achieve closedloop stability, the frequency plots of the open-loop diagonal elements lmm (j) should not encircle the critical point (1 + j0). Solving the equation lmm (j) = gmm (j)cm (j) k Im = k Pm + (amm () + jbmm ()) j = 1

(10)

for real and imaginary part separately, one obtains the stability region in kP kI plane, which is delimited by k Pm = amm () bmm () , k I m = 2 2 () rmm rmm () (11)

plotted for 0 < . The PI controllers guaranteeing the phase margin M are those lying on the boundary of the region characterized by (9) and inside the stability region (11). The region (9) can be determined in a graphical way as the intersection of the ellipses plotted for each 0 < . Numerically, a range

for kPm should be specied at rst and then the endpoints of the admissible range of kIm are chosen from solutions of quadratic equations for each value of kPm and 0 < . Remark 3: As it was mentioned above, similarly to SISO case, the controllers guaranteeing the phase margin specication may not guarantee overall system stability since the Gershgorin band passing through the point ejM still may contain the critical point (1 + j0). Although such cases are very rare, the inequality (9) with M = 0 can be added to avoid the Gershgorin band to contain the critical point. From the whole set of controllers guaranteeing the phase margin, we will choose those with maximum integral part for each loop since they minimize the sum of integral errors and with reasonable damping they generally produce the best values for the speed of response and performance, see [30]. IV. C ONTROL OF N ON M INIMUM P HASE Q UADRUPLE -TANK P ROCESS The decentralized PI controller design procedure derived above will be implemented on a laboratory model of

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
HUEK: DECENTRALIZED PI CONTROLLER DESIGN 5

quadruple-tank process in nonminimum phase conguration. A schematic diagram of the process is depicted in Fig. 3. The process inputs u 1 and u 2 [V] are the voltages to the pumps and the outputs y1 and y2 [cm] are the levels in the lower tanks. After linearization around an operating point, we obtain a fourth-order state-space model x =A x+B u y=C x+D u (12)
y [cm]
1

12.1 12 11.9 11.8 11.7 11.6 11.5 11.4 0 Chen

M = 25

y2[cm]

with u i = u i u 0 yi = yi yi0 , i = 1, 2, x i = h i h 0 i, i, i = 1, . . . , 4 where h i [cm] denote the levels in the tanks and A3 1 0 0 T1 A1 T3 1 A4 0 0 T2 A2 T4 A= 1 0 0 0 T3 1 0 0 0 T4 1 k 1 0 A2 2 k 2 0 A 2 (13) B = (1 2 )k 2 0 A3 (1 1 )k 1 0 A4 C = D= 1 0 0 0 000 100 0 0

500

1000

1500

2000 t[s]

2500

3000

3500

4000

Fig. 10. Comparison of PI controller proposed and Chen and Seeborg [15].
12.5 12.45 12.4 12.35 12.3 12.25 12.2 12.15 12.1 12.05 12 0 500 1000 1500 2000 t[s] 2500 3000 3500 4000 M = 25 Chen

where the time constants are given as Ti = Ai ai 2h 0 i , i = 1, . . . , 4 g (14)

Fig. 11. Comparison of PI controller proposed and Chen and Seeborg [15].

with Ai [cm2 ] and ai [cm2 ], i = 1, . . . , 4, denoting the cross sections of the tanks and their outlets, respectively, ki [cm3 V/s], i = 1, 2 being the pump constants and i [], i = 1, 2, i [0, 1] the relative valves constants dividing the pump ows into the upper and lower tanks. g [cm/s2 ] denotes the gravitational constant. The transfer matrix of the system is 1 c1 (1 2 )c 1 1 + sT1 (1 + sT3 )(1 + sT1 ) G (s ) = (15) 2 c2 (1 1 )c 2 (1 + sT4 )(1 + sT2 ) 1 + sT2 with ci = (Ti ki / Ai ), i = 1, 2. An interesting feature of the system is that by setting the valve constants i , one can change the minimum phase system (if 1 + 2 > 1) to the nonminimum phase one (if 1 + 2 < 1), see [25]. For the controller design, we will apply the nonminimum phase conguration with the parameter values given in Table I. It follows from the RGA analysis that if taking into account interactions in closed loop, the more

suitable pairing is u 1 y2 and u 2 y1 , hence we interchange the columns in the plant transfer matrix (15). The corresponding transfer matrix yields 3.77 0.81 1 + 47.7s (1 + 37.7s )(1 + 80.1s ) G (s ) = . 0.69 3.59 (1 + 30s )(1 + 47.7s ) 1 + 80.1s Using the proposed procedure, the sets of all PI controllers satisfying phase margin specications for M = 0, 15, 25, 40 for both loops are depicted in Figs. 5 and 4. The values of controller parameters with maximum integral part on each curve, which will be considered for control, are summarized in Table II. The courses of the levels in tanks 1 and 2 for step reference signals for the model and the real laboratory system are plotted in Figs. 69, respectively. One can see that the phase margin M represents a suitable parameter to tune the decentralized control ranging from an underdamped behavior for too low values to an overdamped response if its value is too high.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

The comparison of the responses of decentralized PI controller obtained using the proposed method for M = 25 and the procedure suggested by Chen and Seeborg in [15] (last row in Table II), based on modication of the ZieglerNichols rules, is shown in Figs. 10 and 11. V. C ONCLUSION In this brief, a method for decentralized PI controller design of MIMO systems was presented. The method is applicable to a plant described by an arbitrary transfer matrix with time delays. The proposed controller satises the phase margin specication for the Gershgorin bands. The control experiments carried out on the laboratory model of quadruple-tank system conrmed that the phase margin is a suitable tuning parameter closely related to the damping of the closed-loop response. Moreover, the experiments revealed that if appropriately tuned, the proposed design achieves better results than a popular method based on modication of ZieglerNichols rules. R EFERENCES
[1] K. H. Ang, G. Chong, and Y. Li, PID control system analysis, design and technology, IEEE Trans. Control Syst. Technol., vol. 13, no. 4, pp. 559576, Jul. 2005. [2] K. J. strm and T. Hgglund, PID Controllers: Theory, Design and Tuning. Research Triangle Park, NC, USA: ISA, 1995. [3] K. J. strm and T. Hgglund, Advanced PID Control. Research Triangle Park, NC, USA: ISA, 2005. [4] Q. G. Wang and Y. S. Yang, Transfer function matrix approach to decoupling problem with stability, Syst. Control Lett., vol. 47, no. 2, pp. 103110, 2002. [5] Q. G. Wang, Y. Zhang, and M. Chiu, Non-interacting control design for multivariable industrial processes, J. Process Control, vol. 13, no. 3, pp. 253265, 2003. [6] Q. Xiong, W. J. Cai, and M. J. He, Equivalent transfer function method for PI/PID controller design of MIMO processes, J. Process Control, vol. 17, no. 8, pp. 665673, 2007. [7] S. Tavakoli, I. Grifn, and P. J. Fleming, Decentralised PI (PID) controllers for TITO processes, Control Eng. Pract., vol. 14, no. 9, pp. 10691080, 2006. [8] Q. G. Wang, Z. Ye, W. J. Cai, and C. C. Hang, PID Control for Multivariable Processes. New York, USA: Springer-Verlag, 2008. [9] D. E. Seborg, T. F. Edgar, and D. A. Mellichamp, Process Dynamics and Control, 2nd ed. New York, USA: Wiley, 2004. [10] W. Luyben, Simple method for tuning SISO controllers in multivariable systems, Ind. Eng. Chem. Design Develop., vol. 25, no. 3, pp. 654660, 1986. [11] A. Loh, C. Hang, C. Quek, and V. Vasnani, Autotuning of multiloop proportional-integral controllers using relay feedback, Ind. Eng. Chem. Res., vol. 32, no. 6, pp. 11021107, 1993.

[12] S. H. Shen and C. C. Yu, Use of relay-feedback test for automatic tuning of multivariable systems, AIChE J., vol. 40, no. 4, pp. 627646, 1994. [13] Y. Halevi, Z. J. Palmor, and T. Efrati, Automatic tuning of decentralized PID controllers for MIMO processes, J. Process Control, vol. 7, no. 2, pp. 119128, 1997. [14] A. P. Loh and V. Vasnani, Describing function matrix for multivariable systems and its use in multiloop PI design, J. Process Control, vol. 4, no. 3, pp. 115120, 1994. [15] D. Chen and D. E. Seborg, Design of decentralized PI control systems based on Nyquist stability analysis, J. Process Control, vol. 13, no. 1, pp. 2739, 2003. [16] L. Campestrini, L. C. S. Filho, and A. S. Bazanella, Tuning of multivariable decentralized controllers through the ultimate-point method, IEEE Trans. Control Syst. Technol., vol. 17, no. 6, pp. 12701281, Nov. 2009. [17] H. H. Rosenbrock, State-Space and Multivariable Theory. New York, USA: Wiley, 1970. [18] J. M. Maciejowski, Multivariable Feedback Design. Reading, MA, USA: Addison-Wesley, 1989. [19] W. K. Ho, T. H. Lee, W. Xu, J. R. Zhou, and E. B. Tay, The direct Nyquist array design of PID controllers, IEEE Trans. Ind. Electron., vol. 47, no. 1, pp. 175185, Feb. 2000. [20] W. K. Ho, T. H. Lee, and O. P. Gan, Tuning of multiloop proportionalintegral-derivative controllers based on gain and phase margin specications, Ind. Eng. Chem. Res., vol. 54, no. 4, pp. 18981906, 2007. [21] F. Vzquez and F. Morilla, Tuning decentralized PID controllers for MIMO systems with decouplers, in Proc. 15th IFAC World Congr., Barcelona, Spain, 2002, p. 137. [22] Q. Xiong and G. Jin, Iterative decentralized PID tuning based on gain and phase margins for TITO systems, Int. J. Innov. Comput. I., vol. 3, no. 3, pp. 673684, 2007. [23] D. Rosinov and A. Kozkov, Robust decentralized pid controller design, in Introduction to PID ControllersTuning and Application to Frontier Areas. Rijeka, Croatia: Intech, 2012, pp. 133168. [24] D. Garcia, A. Karimi, and R. Longchamp, PID controller design for multivariable systems using Gershgorin bands, in Proc. 16th IFAC World Congr., 2005, pp. 183188. [25] K. H. Johansson, The quadruple-tank process: A multivariable laboratory process with an adjustable zero, IEEE Trans. Control Syst. Technol., vol. 8, no. 3, pp. 456465, May 2000. [26] D. Shneiderman and Z. J. Palmor, Properties and control of the quadruple tank process with multivariable dead-times, J. Process Control, vol. 20, no. 1, pp. 1828, 2010. [27] W. K. Ho, T. H. Lee, and O. P. Gan, Tuning of multiloop PID controllers based on gain and phase margin specications, Ind. Eng. Chem. Res., vol. 36, no. 6, pp. 22312238, 1997. [28] J. R. Bar-on and E. A. Jonckheere, Phase margins for multivariable control systems, Int. J. Control, vol. 52, no. 2, pp. 485498, 1990. [29] Q. G. Wang, Y. He, Z. Ye, C. Lin, and C. C. Hang, On loop phase margins of multivariable control systems, J. Process Control, vol. 18, no. 2, pp. 202211, 2008. [30] Z. Shaei and A. T. Shenton, Frequency-domain design of PID controllers for stable and unstable systems with time delay, Automatica, vol. 33, no. 12, pp. 22232232, 1997.