Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Internationa l Journal of Systems Science, 2002, volume 33, number 1, pages 6779

Modied internal model control approach to the design and tuning of linear digital controllers
and D. B. D a C IC M. R. M atauSEK *, A. D. MICIC

A new approach to the design and tuning of two degrees of freedom linear digital controllers is presented. The important properties of the proposed method are that the controller structure is directly obtained from the model used and that a small number of adjustable parameters can be easily tuned manually to obtain high closedloop system performanc e both in the set-point response and in the load disturbance rejection. The method is applicable on both type-0 and -1 plants and is e ective in the presence of a poorly damped complexconjugate pole pair, a non-minimum phase zero or a dead-time. The proposed design and tuning procedures are illustrated by simulation examples, including the analysis of the IFAC 93 benchmark control problem.

1.

Introduction

Recently, a straightforward modication of the Internal Model Control (IMC) (Morari and Zariou 1989), named Modied Internal Model Control (MIMC) (Mataus ek and Stipanovic 1998, Mataus ek et al. 1998), was found to be more e ective than IMC when designing non-linear neural network digital controllers. Both structures, in their basic form, are presented in gure 1. The non-linear operators P, M and Cm (or Cp ) are the plant, the plant model and the main controller, respectively, while F is the rst-order IMC lter. In reality, the non-linear plant can never be perfectly represented by the model used, i.e. M 6 P. Then, if IMC is applied, for the o set free control, the inverse nonlinear model, used as the main controller Cm , must be a perfect representation of the inverse of the non-linear plant model (Cm M 1 ) (Mataus ek and Stipanovic 1998). This is a very restrictive condition when the experimental neural network modelling technique is used to determine the main controller Cm . Supposing that Cm M 1 , MIMC structure in gure 1b is directly obtained from the IMC structure in gure 1a. Thus, the condition Cm M 1 is inherent to the MIMC structure. The main controller in gure 1b is denoted as Cp , since it represents now an approximation of P1 , which can be
Received 17 March 2000. Revised 27 April 2001. Accepted 31 May 2001. *Author for correspondence. University of Belgrade, Faculty of Electrical Engineering, 11120 Belgrade, PO Box 35-54, Yugoslavia. e-mail: matausek@etf.bg.ac.yu

determined experimentally by the inverse modelling technique proposed in Psaltis et al. (1988). In Mataus ek and Stipanovic (1998), by simulation, and in Mataus ek et al. (1998), by analyzing a real non-linear plant from Mataus ek et al. (1996), it was demonstrated that a high closed-loop system performance, with respect to the disturbance rejection and tracking of the reference input, can be easily obtained by applying an MIMC structure. Besides, a di cult non-linear control problem from Narendra and Parthasarath y (1990) was e ciently solved using even a static main controller, i.e. dening the inverse non-linear model by a polynomial 1998). The method proposed (Mataus ek and Stipanovic by Narendra and Parthasarath y (1990) fails on this example. The present paper is a further development of the modied internal model control. The problem of designing and tuning high-performanc e linear digital controllers for the Single-Input-Single-Outpu t (SISO) plants is studied in detail. A two degrees of freedom MIMC structure is dened. The controller structure consists of a rst-order reference prelter, the rstorder IMC lter, the main controller and a compensation transfer function, introduced to make possible an e ective solution of the di cult control problem when the plant contains a poorly damped complexconjugate pole pair. Both the main controller and the compensation transfer function are directly obtained from the model used to design the controller. The maximal number of tuning parameters is three for type-0 plants or four for type-1 plants. Adjustable parameters of the

International Journal of Systems Science ISSN 00207721 print/ISSN 14645319 online # 2002 Taylor & Francis Ltd http://www.tandf.co.uk /journals DOI: 10.1080/00207720110071994

68
Y
ref

M. R. Matausek et al. + R F Cm U P Y

Ym

+ -

(a)

Y ref

+ -

R F C
p

U P

+ -

(b)

Figure 1.

Basic form of (a) IMC; (b) MIMC.

proposed controller have a clear meaning and can be easily tuned manually. The idea to develop this new method comes from the fact that the nal step of a controller design is always a closed-loop simulation, used to analyse the resulting control system to see if the performance specications are satised. This holds true also for all the powerful analytical methods, such as Output Pole-Placement (OPP), two-Degrees-of-Freedom Loop-Shape Design (2-DOF LSD), LQG, H1 . As a result of this nal step, almost by rule, some modications should be done, and previous steps in designing and tuning must be repeated. Consequently, there is a strong motivation to develop a simple design and tuning procedure based entirely on the simulation. To be e ective, such a method must satisfy the following requirements: to be well suited to exploit the benets of prior knowledge and experience gained from the open-loop dynamics of the plant; the control system structure must be directly obtainable from the model used to design the controller; the controller must contain only a small number of tuning parameters, with clear meaning, followed by simple tuning rules, fairly easy to apply. Finally, the method must guarantee the high closedloop system performance for a large class of the type0 and type-1 plants, characterized either by the welldamped or the oscillatory impulse responses, minimum or non-minimum phase characteristics, and the dead-time. The proposed method satises these objec-

tives and denes an e ective and easy to apply experimental procedure for designing and tuning digital controllers. The paper is organized as follows. The basic properties of the proposed controller are presented in Section 2. Design and tuning are discussed in Sections 3 and 4, respectively. In Section 5.1, a fourth-order type-1 plant with a poorly damped complexconjugate pole pair is used to clarify the main ideas and to analyse in detail the proposed design and tuning method, compared with the stro OPP method (A m and Wittenmark 1990). In Section 5.2, the IFAC 93 benchmark problem (Graebe 1994) is used to analyse, by simulation, the robustness of the MIMC controller compared with the H1 controller (Whidborne et al. 1995). In Section 5.3 it is demonstrated how to design and tune the MIMC controller if the model, used to approximate a higher-order nonminimum phase plant, is represented by a rst-order transfer function plus dead-time. Finally, in appendix A, the relationship to the two degrees of freedom IMC (Morari and Zariou 1989) and to the output pole stro placement method (A m and Wittenmark 1990) is briey discussed.

2.

Control system structure and basic properties

The control system structure for a SISO plant with transfer function Ps is dened as in gure 2, where

Design and tuning of linear digital controllers


D Yref Fr (z)

69

+ -

R F(z)

1 (z ) Pm 0 z -

Ur

Uc

ZOH

Y P(s)

G(z)

1 z

G L (z)

Figure 2.

Control system structure.

F r z F z

1 r z ; z r 1 z ; z

1 2 3

Closed-loop system transfer functions Gr z, relating Y zto Yref z, and Gd z, relating Y z to Dz, for GL z 1, are dened by G r z
1 PzP m 0 z P z G z F z F r z = z ; Q z

G z g 0 g c G c z :

6 7 8

The general form of the plant model, used to design the controller, is given by Pm s exp sPm0 s. In this case, in the presence of the dead-time 6 0, transfer function GL z is dened by: GL z z
L

G d z

Pz1 F z=z ; Q z

1 Q z 1 G z P z P z P m 0 z 1 F z = z:

where delay L is determined from =t and t is the sampling period. If not stated otherwise, it is assumed that 0, i.e. GL z 1. Adjustable parameters of the controller are 0 < < 1, 0 r < 1, gc and g0 . 1 The main controller Cp P m0 z=z is a stable and causal inverse of Pm0 z. Transfer function Gc z is introduced to make possible an e cient active compensation of oscillations of the closed-loop system output caused by the load or impulse disturbance Dz. Parameter gc must be di erent from zero if the plant Ps contains a complexconjugate pole pair with a small damping ratio . If the plant is of type-1 parameter g0 must be di erent from zero to reject the load disturbance. As will be shown in Section 3, transfer functions 1 P m0 z=z and Gc z are obtained directly from Pm0 z. The largest number of adjustable parameters of the proposed controller is four (; r ; gc , g0 ). These parameters can be easily tuned manually to satisfy stringent timedomain closed-loop performance requirements, even when the control signal ut is constrained i.e. jut 4 U max , where and the transfer function Gh0 s 1 ets =s denes the zero-order hold (ZOH). ut L1 Gh0 sU *s 5

The basic properties of the proposed controller will be discussed considering the case of perfect inverse modelling,
1 1 P m0 z P z:

In this case, from (6) to (8) it follows: G r z Fr z G d z F z ; z 10 11

P z F z z 1 ; 1 P z G z 1 z 1 F z D z : 1 P z G z z

and, for Yref z 0 , signal Ur z is given by U r z 12

Thus, in the case of perfect inverse modeling and GL z 1, it follows from (10) that the set-point response is a damped second-order system step response, for both the type-0 and type-1 plants. From (11), it follows that the condition lim Gd z 0
z !1

13

is satised for type-0 plants if Gz 0. However, for type-1 plants the condition (13) is satised only if

70 lim Gz g0 ; g0 6 0;
z!1

M. R. Matausek et al. 14 W 0 z z 1 : tz

assuming that lim z!1 Gc z 0. In other words, to reject the load disturbance, the parameter g0 must satisfy the relation g0 6 0 when considering type-1 plants. Finally, let g0 0 and consider the closed-loop system step or impulse disturbance response. It follows from (12) that Ur z represents the step or impulse response of the rst-order system F z=z if Gc z 0. In this case, according to (3) and (11), a step or impulse disturbance Dz will cause strong oscillations of the closed-loop system output Y z, if the plant Ps contains a complex-conjugat e pole pair with small damping ratio .

1 (3) P m0 z =z is obtained as a cascade of rst-order, W i z, Wj z, W0 z, and second-order, Wp z, transfer functions, dened above, and the inverse gain (static or velocity) of Pm0 s i.e. Y Y 1 1 P W 0 z W p z W i z W j z : m0 z=z K i j

21

3.

Design

Design of the proposed controller requires specication 1 of two transfer functions: P m0 z=z and Gc z. For the sampling period t, the stable and causal inverse 1 P m0 z=z is obtained by applying the partial (cascade) discretization. The procedure is as follows: (1) Represent both the nominator and denominator of Pm0 s in the factored form, e.g. K Pm0 s s
j 1 m1 Q

Tj s 1
i 1 n1 Q

: T i s 1

T 2 s2

2 Ts 1

15

(2) Then, represent all real continuous-time poles and only stable zeros by unity gain transfer functions of the form W i z za ; a expt=Ti ; 1 a z 16

The reason why throughout the paper the inverse 1 model is denoted by P m0 z=z is the following. Representing the transfer function Pm0 s in the form Pm0 s Bc s=Ac s, deg Bc m; deg Ac n, m < n, and by applying the z transform directly on Gh0 sPm0 s, one obtains Pm0 z Bd z=Ad z, where deg Bd deg Ad 1. In this case, the causal inverse of Pm0 z is dened by Ad z=zBd z 1 P m0 z=z. Also, it should be mentioned here that, for a small sampling period t, some roots of the polynomial Bd z would have negative real parts, even in the case when Pm0 s denes a minimum phase system. To avoid ringing of the control signal, these zeros must be sub1 stituted in P m0 z=z Ad z=zBd z with zeros at the origin. By applying the partial (cascade) discretization 1 this additional step in dening P m0 z=z and some numerical imperfections of the discretization procedures used in MATLABSIMULINK are avoided. The transfer function Gc z is dened also in a straightforward manner. If Ps Pm0 s is a secondorder transfer function dened by only one poorly damped complexconjugate pole pair, the transfer function G c z z1 z 22

for poles, Gpi s 1=Ti s 1, and 1 b z W j z ; b expt=Tj ; zb 17

for zeros, Gzj s Tj s 1. Continuous-time complexconjugate pole pair, Gp s 1=T 2 s2 2 Ts 1, is represented by unity gain transfer function Wp z of the form: W p z z2 a 1 z a 2 ; 1 a1 a2 z2 18 q 1 2 ; 19 a2 exp2 t=T : 20 The pole at the origin, G0 s 1=s, is represented by

t a1 2 exp t=T cos T

was found to be an adequate choice of Gc z to make possible compensation of strong oscillations caused by the disturbance. To obtain an adequate compensation signal uc t for a higher-order transfer function Pm0 s, e.g. as the one dened by (15), all the additional poles and stable zeros (except the complexconjugate pole pair with a small damping ratio ) must be included into Gc z, i.e. must be cancelled in the path relating Y z to Uc z, for g0 0. This means that the transfer function Gc z, for Pm0 s as in (15), is given by Y Y z 1 Gc z W 0 z W i z W j z : 23 z i j MIMC controller, with Gc z dened in this way and adequately tuned parameter gc , generates a pulse-like signal uc t, such that the oscillations caused by disturbance are compensated by the oscillations of the plant

Design and tuning of linear digital controllers output induced by uc t. This procedure, discussed in detail in Section 5.1, is named active compensation in comparison with the active damping, used in the stro output pole-placement method (A and m Wittenmark 1990) for the same purpose, i.e. to compensate strong oscillations of the output of the closed-loop system, caused by the disturbance in the presence of a poorly damped complex-conjugat e pole pair of Ps. 4. Tuning

71

For a type-1 plant with well-damped poles only three parameters (; r ; g0 ) must be adjusted. Type-1 oscillatory plant requires the tuning of four parameters (; r ; gc ; g0 ). However, even for an oscillatory plant, e cient rejection of the impulse disturbance requires adjustment of only three parameters (; r ; gc ) for both types of plants. For type-0 plants the number of adjustable parameters can be reduced by 1, i.e. g0 can be taken as zero. However, this statement is not correct if the control objective is to obtain the closed-loop system rise-time TrCL considerably smaller then the dominant time constant of the open-loop system, e.g. as in Skogestad and Postlethwaite (1996, example 2.7) or in Mataus ek et al . (1998). For example, in Skogestad and Postlethwaite (1996), it is required to obtain TrCL < 0:3 s (overshoot <5%) for the plant dened by the transfer function Ps 200 =10s 10:05s 12 . In this case, as seen by the controller, the plant behaves like a type-1 plant dened by the transfer function Ps 20=s0:05s 12 and parameter g0 must be different from zero, to obtain fast rejection of the load disturbance. From (10) and (11), it follows that smaller , 0<<1, result in smaller closed-loop system risetimes and better rejection of the load disturbance. However, since the model used to design the controller is never a perfect representation of the plant, parameter must be adjusted, as in IMC, in a trade-o between the stability/robustness and the closed-loop system risetime. Parameter 0 r < 1 is used to adjust the overshoot and to prevent appearance of the high values of the control signal, when the set-point is changed. The sampling period t is determined in the usual way. When the desired closed-loop system rise-time is TrCL , then t is taken to be t TrCL =Nr ; N r 4 10, as stro suggested by A m and Wittenmark (1990), or smaller then this value. Initial estimates of and r are dened by in rin exp2t=TrCL , since, according to (10), the set-point response is dened by cascade connection of two rst-order transfer functions F z and Fr z. Obviously, if r 0, initial estimate of is dened by in expt=TrCL . Parameters and r are adjusted rst. Then, in the case of an oscillatory plant the parameter gc is adjusted,

setting g0 0 in (3). gc is increased from zero to the value when oscillations of the closed-loop system output, caused by the disturbance, disappear. Then, if the plant is of type-1, parameter g0 is increased from zero to a value that guarantees fast rejection of the load disturbance , not deteriorating signicantly the set-point response. It is recommended to adjust normaln ized values (gn c ; g0 ) of the parameters gc and g0 , i.e. n gc gc =K and g0 gn 0 =K , respectively, where the parameter K is the velocity or the static gain of the plant model. A great number of simulations, and some experiments on a real plant, conrm that the proposed manual tuning can be easily performed.

5.

Simulation examples

To demonstrate the design and tuning procedures and to analyse the performances of the closed-loop system in detail, three examples will be considered. The rst is the problem of controlling a type-1 SISO plant with a poorly damped complexconjugate pole pair. The second example is dened by the IFAC 93 benchmark problem. A non-minimum phase plant used to demonstrate the role of the transfer function GL z zL denes the third example. 5.1. Type-1 SISO plant with a poorly damped complex conjugate pole pair This plant is dened by the transfer function P 1 s K T3 s 1 ; 2 s2 2 T s 1 sT1 s 1 T2 2 24

where the nominal values of the parameters q K ; T 1 ; T 2 ; T 3 ; are given by q q1 0:2736 ; 0:05, 0.087, 0.8, 0.0546). Tr 0:1 s is obtained from the open-loop impulse response of P1 s. For TrCL Tr , i.e. for t 0:01 s and Pm s P1 s, according to steps (2) and (3), one obtains
1 P 1 z1 m 0 z z 0:2736 0:01z

0:0124 z 0:8187 z2 1:9744 z 0:9875 : 0:1813 z 0:9876 0:0131 z2 Transfer function Gc z is dened by G c z z 1 z 1 0:0124 z 0:8187 : z 0:01z 0:1813 z 0:9876

25

26

Initial estimates of and r are in rin 0:82. In order to obtain faster load rejection, is adjusted to 0:7 and r was set to r 0:85. Set-point and load disturbance responses presented in gure 3ab are obtained by applying the MIMC controller dened by (25) and (26) and gc g0 0. The

72

M. R. Matausek et al.

Figure 3.

Set-point and load disturbance responses of P (s) 5 P1 (s) for q 5 q1 . MIMC controller, L 5 gc 5 g0 5 0; (c, d) and (e, f) gc 5 50, g0 5 0; (g, h) gc 5 50, g0 5 7.

0, a 5

0:7, ar 5

0:85: (a, b)

results presented in gures 3cd are obtained for gc 50 and g0 0. In gure 3ef two signals, yr=d t and yc t, are presented to illustrate why the strong oscillations in gure 3a are eliminated in gure 3c by adjusting the parameter gc to 50. Signal yr=d t is the response of P1 s to the input signal ur t d t, while yc t is the

response of P1 s to the signal uc t. The signal yt in gure 3c is the superposition of the signals yr=d t and yc t, i.e. yt yr=d t yc t according to gure 2. Obviously, signal uc t perfectly compensates strong oscillations of the system output induced by the disturbance, as demonstrate d in gure 3c. This is the origin of

Design and tuning of linear digital controllers the name active compensation. Finally, to obtain the o set free control, the parameter g0 was set to 7. For gc 50 and g0 7, set-point and load disturbance responses are presented in gure 3gh. The results obtained conrm entirely the proposed design and tuning procedure. This analysis shows that a high-performance digital controller can be easily designed and tuned for a very `unpleasant plant, as indicated by the impulse response of P1 s, presented in gure 4 for q q1 . In table 1, the Phase Margin (PM) , the Gain Margin (GM) and the Maximum sensitivity (Ms) of the so designed and tuned controller are presented. PM and GM are calculated for the both nominal plant (q q1 ) and perturbed plant q q2 0:5085, 0.05, 0.0548, 0.342, 0.08). As seen from table 1, the closed-loop system is stable even in the case of strong perturbation (q q2 ). Since the desired set-point response, dened by (10), is perfectly realized, to compare MIMC design to the stro Output Pole-Placement (OPP) design (A m and Wittenmark 1990), let the polynomial Am z in the OPP design be dened by Am z z z r , where 0:7 and r 0:85. Also, let the polynomial Ad z, used in the OPP design to obtain active damping, be dened by the two dominant poles of the transfer function (11), i.e. Ad z z2 1:7636 z 0:803 . It is interesting to note that almost the same Ad z is obtained by discretizing the continuous-time pole pair p 2 p1=2 !n j !n 1 , where !n 11:49rad =s and 1. Then, the OPP controller in gure 10 (appendix A) is given by the relations (2729), 85:94 z4 1:7636 z3 0:803 z2 Q r z ; R c z 27

73

Table 1. Phase margin, gain margin and maximum sensitivity for MIMC controller and process P (s) 5 P 1 (s): PM(8) MIMC (nominal plant) MIMC (perturbed plant) 36 9.4 GM (dB) 7.7 1.5 Ms 2.1

Rc z z4 0:8601 z3 1:0222 z2 0:6540z 0:2283 ;

29

which are di erent from (A34). The set-point, load disturbance and output disturbance responses, obtained by the OPP controller and the MIMC controller are presented in gure 5. It is evident that the set-point responses are the same for both controllers, while the load disturbance rejection is better for the OPP controller. This is conrmed by gure 6, where the magnitude of the return di erence function is presented for both controllers, in the region around the natural frequency !n 11:49rad =s. However, the performance of the OPP controller is less impressive in the presence of the output disturbance, resulting in a very high actuator e ort. Moreover, contrary to the MIMC controller, an unstable closed-loop system was obtained when this OPP controller was applied to the perturbed plant, i.e. when q q2 . 5.2. IFAC 93 benchmark problem The design and tuning problem will be considered using the SISO model of a process with parameters known only within a certain range (Graebe 1994, Whidborne et al . 1995). Depending on various operating regimes, the loop operates at three stress levels, with higher stress levels inducing larger parameter variations. The set-point to the loop is a square wave varying between 1 and 1 with a period of 20 s. For each stress level, it is required to design a controller to achieve as fast a rise-time TrCL as possible (preferably between 2 and 3 s), subject to the following conditions. . Plant output must be within 1:5 and 1:5 at all times. . Zero steady-state tracking error (modulo high frequency noise). . Preferable if under/overshoot is 0:2 most of the time (occasional large under/overshoots are acceptable as long as the output is within 1:5). . Fast settling time.

3839 :4387 z4 13364 :2815z3 17520 :0456z2 10244 :1085 z 2252 :2918 ; Q d z R c z
3 2 y(t) 1 0 -1 -2 0 1 2 3 4 5 6 Tim e (s) 7 8 9 10

28

. Plant input saturates at 5:0 and 5:0. The complete transfer function Psof the plant is given by

Figure 4.

Impulse response of P (s) 5

P1 (s) for q 5

q1 .

74

M. R. Matausek et al.

Figure 5.

Set-point, load disturbance and output disturbance responses of P (s) 5 controller.

P1 (s) for q 5

q1 : (a, b) MIMC controller, (c, d) OPP

10 3

10 2

10 1
n

10 0 0 10

(rad/s)

101

Figure 6.

Magnitude of the return di erence function: MIMC controller (solid): OPP controller (dotted).

K T 2 s 1 ! 2 0 P 2 s 2 s 2!0 s !2 T 1 s 1 0

where T1 1=8; T 2 1=12 ; ! 15; 0:6. The model Pm s, used to design the controller, is given by

!2 ; s2 2 ! s !2 T1 s 1 T2 s 1

30

where the nominal values qn of the parameters q q1 q2 q3 q4 q5 T1 T2 !0 K are qn 5 0.4 5 0.3 1). Thus, the control problem considered includes the e ect of unmodelled dynamics. For each stress level, the complete transfer function is dened by (30) with the parameter variation intervals q q1 q2 q3 q4 q5 dened in table 2. Besides the high range of stochastic variations of the plant parameters, abrupt changes of the static gain K occasionally take place. This e ect, especially pronounced when the closed-loop simulation of the IFAC 93 benchmark problem is performed for the stress level 3, may cause occasional violation of the hard constraint jytj 1:5. The sampling period t 0:2 s is adopted, corresponding to the desired closed-loop rise-time TrCL 2 s and Nr 10. Discretizing the nominal plant model (31) for q qn , it is obtained that MIMC controller is dened by the relations: R z 1 z 1 r z Yref z Y z z1 Rz ; z z r 32

K T 2 s 1 ! 2 0 P m s 2 ; s 2!0 s !2 T s 1 1 0

31

|1 +W|

Design and tuning of linear digital controllers


Table 2. Stress level 1 2 3 IFAC 93 Benchmark parameter variation intervals. T 1 0.20 0.30 0.30 T 2 0.05 0.10 0.15 !o 1.50 2.50 3.00 0.10 0.15 0.15 K 0 0.15 0.50

75

U z

z2 0:8571 z 0:5488 z 0:9608 R z : 0:0392 z 0:6917 z2

33

In this analysis r 0 and gc g0 0. Since larger parameter variations are induced with larger stress levels, greater values of the parameter should be used for higher stress levels. Accordingly, parameter was set to expt=2 for stress level 1; expt=2:3 for stress level 2; and exp t=3:5 for stress level 3. Using the so designed and tuned MIMC controller and plant Ps P2 s, closed-loop system responses, presented in gure 7, are obtained for q qn . Then, the analysis is performed using a `black box simulation of the IFAC 93 benchmark control problem, dened by a scrambled computer code (Graebe 1994) in which the supplied unity feedback controller is replaced with the MIMC controller (3233). The results are compared with those obtained, on the same simulator, using the H1 (mixed sensitivity) controller. Taking into account its modest complexity and good robust performance, the H1 (mixed sensitivity) controller is chosen as an adequate representative of the analysis performed in Whidborne et al. (1995), where, combining the method of inequalities with analytical optimisation techniques, several robust controllers (PI/PID, LQG, H1 , 1-DOF LSD, 2-DOF LSD and ) were designed and analysed.
1,5 1,0 0 ,5 0 ,0 -0 ,5 -1,0 -1,5 0 2 4 6 8 10 12 14 16 18 20
Time (s)

Figure 8 illustrates the performance of the MIMC controller. For comparison, the results obtained by the H1 (mixed-sensitivity) controller are also presented. Both controllers are investigated by running the simulator over 300 s, separately for both controllers and for each of the three stress levels. The results are presented by superimposing the 15 responses of the simulator output over periods [0, 20); . . . ; [280, 300) in a window scaled from 0 to 20 s on the horizontal and 1:5 to 1.5 on the vertical axis. Then, for the MIMC controller the simulator was run 75 times over 600 s, yielding 2250 realizations of the 20-s duration for each of the three stress levels. If the per cent of violations of the hard constraint jytj 1:5 within these 2250 realizations is denoted by , one obtains: 0, 0.6 and 1:5%, respectively for stress levels 1, 2 and 3. These results and the results presented in gure 8 are very encouraging, having in mind the simplicity of both the design and tuning procedures of the MIMC controller. 5.3. Type-0 SISO non-minimum phase third-order plant In this example the role of the transfer function GL z zL will be demonstrated. The plant Ps P3 s is dened by the transfer function P 3 s Ts 13 1 s ; 34

where the nominal values of the parameters T and are T 1, 1:4. Two models are used to design the MIMC controller. The rst model Pm s Pm1 s is an approximation of the plant dynamics, dened by Pm1 s exp2:51s : 1:79s 1 35

Figure 7. Set-point responses of P (s) 5 P2 (s) for q 5 MIMC controller, L 5 0, G (z) 5 0, ar 5 0: a 5 e 2 (solid), a 5 e 2 0:2=2:3 (dotted), a 5 e 2 0:2=3:5 (dashed).

qn . 0:2=2

The rst-order plus dead-time model (35) is obtained by using an e ective experimental procedure (Mataus ek ev 2000). The equivalent dead-time and Kvas c ( 2:51 s) and the equivalent time constant (Tm 1:79 s) are obtained from the step response of the plant as follows: is the time t1 to go to 10% of the steady-state yss of the step response of the plant and Tm is the time Tm t2 t1 , where t2 is the time to go to 63% of the yss . The second model is based on the assumption that the plant dynamics are perfectly modelled, i.e. Pm s P3 s. In this example gc g0 0 and the sampling period is t 0:2 s. For this sampling period one obtains that the delay L, in the transfer function GL z, should be given by L 12 if the model Pm s Pm1 s is used to design the MIMC controller. In gure 9a, set-point and load disturbance responses are presented for an adequately tuned analogue PI controller. The results presented in gure 9b are obtained by the MIMC controller based on the model (35), but neglecting the information about the dead-time

76

M. R. Matausek et al.

y(t)

y(t)

(a)

Time (s)

(b)

Time (s)

y(t)

y(t)

(c)

Time (s)

(d)

Time (s)

y(t)

y(t)

(e)

Time (s)

(f)

Time (s)

Figure 8. Responses of the IFAC 93 benchmark simulator. MIMC controller, L 5 0, G (z) 5 0, ar 5 0: (a) stress level 1, a 5 e 2 0:2=2 ; (c) stress level 2, a 5 e 2 0:2=2:3 ; (e) stress level 3, a 5 e 2 0:2=3:5 . H1 (mixed sensitivity) controller: (b) stress level 1; (d) stress level 2; (f) stress level 3.

2:51s. This means that the parameter L was set to 0. For Pm s Pm1 s and L 0 the MIMC controller is equivalent to a digital PI controller. To obtain almost the same responses as in gure 9a, parameters and r were adjusted to 0:985 and r 0. The results presented in gure 9c are obtained for Pm s Pm1 s and L 12, 0:8943 and r 0:85. In both experiments, gure 9bc, the main controller is dened by
1 P m0 z=z

for L 0, 0:94 and r 0:9. In this case the main controller is dened by z 0:8187 3 1 P z z 37 = : m0 1 0:8187 z Responses in gure 9ad are presented for the nominal plant, T 1, 1:4 (solid), and for the perturbed plant, T 1:3, 1:68 (dotted). The impressive improvement of the closed-loop system performance is obtained using a priori information that the delay L is given by L 12 if t 0:2 s and the model Pm s Pm1 s is used to design the MIMC controller. Moreover, the closed-loop system performances obtained by using the model Pm s Pm1 s and

z 0:8943 : 1 0:8943 z

36

The results presented in gure 9d are obtained by the MIMC controller based on the model Pm s P3 s and

Design and tuning of linear digital controllers

77

Figure 9. Set-point and load disturbance responses of P(s) 5 P3 (s). Nominal plant (solid), perturbed plant (dotted): (a) analogue PI controller; (b)(d) MIMC controller, G (z) 5 0. (b) Pm (s) 5 Pm1 (s), L 5 0, ar 5 0, a 5 0:985; (c) Pm (s) 5 Pm1 (s), L 5 12, ar 5 0:85, a 5 0:8943; (d) Pm (s) 5 P3 (s), L 5 0, ar 5 0:9, a 5 0:94.

L 12 are very close to those obtained by using the model Pm s P3 s and L 0. Concluding this section, it should be noted here that the delay L must be di erent from zero if the plant is dened by Ps exp sP0 s and the dead-time is di erent from zero. However, when discussing the deadtime compensation, the proposed controller should be compared with the Smith Predictor (SP) (Ha gglund 1992), for type-0 plants, or to the Modied Smith 1996, 1999), for Predictor (MSP) (Mataus ek and Micic type-1 plants, since both methods guarante e high closedloop system performance. In Ha gglund (1992), a simple and e ective tuning of SP is proposed for type-0 plants, while in Mataus ek and Micic (1996, 1999), a new deadtime compensator with a simple and e ective tuning is proposed for type-1 plants. Owing to the space limitation, such a comparative analysis could not be given here. 6. Conclusion

The maximal number of tuning parameters is four for type-1 oscillatory plants, in the presence of the load disturbance. In this case, type-0 oscillatory plants require the adjustment of only three parameters. The minimal number of tuning parameters that have a clear physical interpretation is two. Simulation results show that the proposed design and tuning procedures are a promising tool for dening high-performanc e digital control systems for single-input-single-outpu t plants. A straightforward design, followed by an e ective and simple manual tuning of a small number of adjustable parameters, even in the presence of the constraint on the manipulated variable, is a very useful property of the proposed method for many practical applications. Acknowledgements The authors are grateful to J. F. Whidborne, who was kind enough to give them the scrambled C code of the IFAC 93 benchmark, with a code for H1 controller, and to clarify some performance criteria. Also, the authors thankfully acknowledge the comments and suggestions made by the referees. A1. Appendix A: Relationship to other methods

The proposed method of designing and tuning two degrees of freedom linear digital controllers can be successfully applied to the minimum and non-minimum phase single-input-single-outpu t systems. The method is applicable on both type-0 and type-1 plants and it is e ective in the presence of a poorly damped complex conjugate pole pair. The controller structure is directly obtained from the model used.

Two design methods will be considered: the output pole stro placement method (A m and Wittenmark 1990) and

78

M. R. Matausek et al.
D Y ref Cr(z)

+ -

U ZOH

+
P(s)

Pm(z)

Cd(z) (a) D Y ref Qr(z)

+ -

U ZOH

+
P(s)

Qd(z)

(b)

Figure 10.

Relationship of MIMC to (a) two-degrees of freedom IMC structure; (b) output pole-placement structure.

the two degrees of freedom IMC (Morari and Zariou 1989). From gure 2, in the case when GL z 1, by simple transformation s one obtains two structures, presented in gure 10ab, where C r z Cd z
1 P m0 z F z F r z ; z

A 1 A 2 A 3 A 4

1 P m 0 z F z G z ; z " # 1 P Gz 1 z m0 z Q r z F z ; z z z1 r

Cd z, or Qr z and Qd z, by using di erent closed-loop system performance criteria. On the other hand, the MIMC approach o ers a straightforward design, followed by an e ective and simple manual tuning of a small number of adjustable parameters, even in the presence of the constraint on the manipulated variable. For many practical applications this is a very important property of the proposed method.

References
stro m, K. J., and W ittenmark, B., 1990, Computer-Controlled Systems, 2nd edn (Englewood Cli s: Prentice-Hall). Graebe, S. F., 1994, Robust and adaptive control of an unknown plant: a benchmark of new format. Automatica , 30 , 567 575. gglund, T., 1992, A predictive PI controller for processes with Ha long dead-times. IEEE Control System Magazine, 12, 5760. , B. I., MILJKOVIC , D. M., and B ebic , M ATAUSEK, M. R., JEFTENIC M. Z., 1996, Gain scheduling control of DC motor drive with eld weakening. IEEE Transactions on Industria l Electronics, 43, 153 162. EV, G., 2000, Step response tuning of the PI M ATAUSEK, M., and KVASC controller revisited. Internal report, Faculty of Electrical Engineering, Belgrade, pp. 111. , A. D., 1996, A modied Smith predicM ATAUSEK, M. R., and M icic tor for controlling a process with an integrator and long dead-time. IEEE Transactions on Automatic Control, 41, 11991203. , A. D., 1999, On the modied Smith M ATAUSEK, M. R., and Micic predictor for controlling a process with an integrator and long deadtime. IEEE Transactions on Automatic Control, 44 , 16031606. , D. M., and J eftenic , B. I. , 1998, M ATAUSEK, M. R., MILJKOVIC Nonlinear multi-input-multi-output neural network control of DC

Q d z C d z

It should be mentioned here that when GL z 1 and Gz 0, the MIMC is directly related to the inverse dynamic control, i.e. to the `classic control system structure, where the controller C z is dened by 1 C z F z=1 F z=zP m 0 z = z. When GL z 1 and Gz 6 0, the MIMC in gure 2 can be interpreted as a two degrees of freedom IMC design or as the output pole-placement design, with transfer functions Cr z and Cd z, or Qr z and Qd z, dened by (A12), or (A34), respectively. It should be mentioned here that both designs, the IMC and the output pole-placement design, o er more freedom to determine the transfer functions Cr z and

z : z1

Design and tuning of linear digital controllers


motor drive with eld weakening. IEEE Transactions on Industria l Electronics, 45 , 185187. , D., 1998, Modied nonlinear M ATAUSEK, M. R., and Stipanovic internal model control, Control and Intelligent Systems, 26, 5763. M orari, M. , and Zafiriou, E. , 1989, Robust Process Control (Englewood Cli s: Prentice-Hall). Narendra, K. S., and P arthasarathy , K., 1990, Identication and control of dynamical systems using neural networks, IEEE Transactions on Neural Networks, 1, 427.

79

P saltis, D., Sideris, A., and Y amamura, A. A., 1988, Multilayered neural network controller. IEEE Control System Magazine, 8, 17 21. Skogestad, S., and P ostlethwaite , I. , 1996, Multivariable Feedback Control: Analysis and Design (New York: Wiley). W hidborne, J. F., Murad, G., Gu , D. W. , and P ostlethwaite , I. , 1995, Robust control of an unknown plant the IFAC 93 benchmark. International Journal of Control, 61, 589640.

You might also like