00 0 1997 AACC
S.P. Bhattacharyya Department of Electrical Engineering Texas A&M University College Station, TX 77843 bhat t @eesunl.tamu.edu
sign requires readjustment because no scalar index can capture all the performance requirements of a control system. This means that any useful design procedure should generate a controller which also has sufficient room for readjustment of its coefficients. This translates to the requirement that an adequate stability margin be available around the transfer function coefficients of the designed nominal controller. With the above background as motivation we study in this paper the parametric stability margin of several controller designs from the published literature, obtained by , , 1 and p approaches. The parametric using the H2,H stability margin calculation is a standard computation in the robust parametric control literature [lo, 11, 12, 131. For the calculations given in this paper we used the software package Robust Parametric Control Toolbox developed in conjunction with the text [lo]. In each of the examples treated we obtain the somewhat surprising conclusion that the parametric stability margin of the controller designed is vanishingly small. This means that extremely small perturbations of the coefficients of the controller designed will succeed in destabilizing the loop; in other words the controller itself is frsgile and so is the control system. We also compute the gain and phase margins of the systems designed and observe that these too are very poor, except in one case. In the last section of the paper we briefly discuss some of the issues raised by the calculations presented in the paper. It is obvious that it would be unwise to place a controller that is fragile with respect to perturbations of its coefficientsin an actual control system, without further precautions and analysis. A quick fixto the design procedure is to include the parametric stability margin with respect to the controller designed as a side constraint in the optimization algorithm. However our hope is that this paper will open up research into some of the fundamental pitfalls of optimal and robust designs as currently advocated in the control literature.
1. Introduction
Over the last 15 years control theory has developed several techniques for designing linear time-invariant control systems that are optimum and robust. These techniques rely on the YJBK parametrization [l]of all stabilizing controllers for a fixed linear time-invariant plant, which provides a free parameter over which an appropriate function of a closed loop transfer function may be minimized. Elegant techniques for minimizing the H2 [2], Ho0 [l,3,4] and l1 [6] norms of various closed loop transfer functions as well as efficient numerical approaches to minimizing p, a structured uncertainty measure [5] have been developed using this technique, the state space DGKF approach [7], and the MUSYN ToolBox [SI. These formulations reflect robust stability and/or robust performance. An implicit assumption that is inherent to this type of methodology is that the controller that is designed will be implemented exactly. Relatively speaking this assumption is valid, in the sense that plant uncertainty is clearly the most significant type of uncertainty in a control system, while controllers are generally implemented with high precision hardware. On the other hand it is necessary that any controller that is part of a closed loop system be able to tolerate some uncertainty in its coefficients. There are at least two reasons for this. First controller implementation is subject to the imprecision inherent in analog-digital and digital-analog conversion, finite word-length and finite resolution measuring instruments and roundoff errors in numerical computations. Thus it is required that there exist a nonzero (although possibly small) margin of tolerance around the controller designed. Second, every paper de-
2. Examples
We analyze four examples of optimal designs taken from the control literature. Each example is a single-
1307
Authorized licensed use limited to: Texas A M University. Downloaded on September 1, 2009 at 14:57 from IEEE Xplore. Restrictions apply.
input single-output system and the controller is designed to optimize some closed loop performance function. In Examples 1, 3, and 6, the procedure consists of fixing the nominal plant model, parametrizing all proper feedback controllers that stabilize the nominal model through the YJBK parameter Q(s), which is only required to be stable and proper, and optimally selecting Q ( s ) . In cases where the optimal Q(s) turns out to be improper it is divided by a factor (ST l)k to make it proper, and T is chosen to be suitably small and positive. For our analysis we compute in each case, the parametric stability margin around the coefficients of the designed optimal controller, as well as the gain and phase margins of the closed loop system.
0 . 5
0.4
0 . 30.2 -
.-
o)
0.1 -
P -
0 -. ............
-0.1
::I
4.3 0 . 2
ExampEe 1. (H, based optimum gain margin controller) This example is found in ([9], p.200). It uses , the YJBK parametrizaton and the machinery of the H
Model Matching Problem to optimize the upper gain margin. The plant to be controlled is
-1
This means roughly that a reduction in gain of one part in one thousand will destabilize the closed loop sysP(5) = 9-s-2 tem! Likewise a vanishingly small phase perturbation is destabilizing. and the controller, designed to give an upper gain margin To continue with our analysis let us consider the transof 3.5, (the closed loop is stable for the gain interval [l, fer function coefficients of the controller to be a parameter 3.51) is obtained by optimizing the H , norm of a complevector p with its nominal value being mentary sensitivity function. The controller found is
s-1
...
...,...
......................
1
0
-0.8
Real Axis
-0.6
-0.4
-0.2
and let Ap be the vector representing perturbations in p. We compute the Z2 parametric stability margin around where __
q$ = 379 q i = 39383 q4 = 192306 q ! = 382993 q$ = 383284 qy = 192175 qg = 38582
=3
-328 = -38048 = -179760 = -314330 = -239911 = -67626.
p = 0.15813903109631.
The normalized ratio of change in controller coefficients required to destabilize the closed loop is
-IIP0Il2
- 2.103407115900516 x
The poles of this nominal controller are: 174.70, -65.99, -1.86, -1.04, -0.98 fj0.03
and the poles of the closed loop system are: -0.4666 fj14.2299, -5.5334 fj11.3290, -1.0000 fj0.0002, -0.9998 -1.0002,
This shows that a change in the controller coefficients of less than 1 part in a million destabilizes the closed loop. This controller is anything but robust; in fact we are certainly justified in labeling it as a fragile controller. In order to verify this rather surprising result, we construct the destabilizing controller whose parameters are obtained by setting p = po Ap and are:
913=
45
379.000285811 39382.999231141
192305.999998597
p6 p5
p3
= =
3.158134748 -327.999718909
-38048.000776386
and this verifies that the controller is indeed stabilizing. The Nyquist plot of P(s)C(s)is shown in Figure 1 and verifies that the desired upper gain margin is achieved. On the other hand, we see from Figure 1 that the lower gain margin and phase margin are: 0.99921, Phase Margin = IOo, 0.1681] Gain Margin = [l,
44 =
43
p4 =
p2
= -179760.000001380
The closed loop poles of the system with this controller are: 0.000 fj14.2717, -5.5745 fj10.9187,
1308
Authorized licensed use limited to: Texas A M University. Downloaded on September 1, 2009 at 14:57 from IEEE Xplore. Restrictions apply.
-0.9820
0.8
0.6 0.4 I I I
.*
- - - e - - _-
which shows that the roots crossover to the right-half plane at w = 14.27, and the perturbed controller is indeed destabilizing.
.
\ \ \
I
I
\\
Introduce the parameter vector p corresponding to the controller coefficients and with nominal value:
0 0 P O = [Q10 Qo Pol
P ( s )=
The normalized ratio of change in controller coefficients required to destabilize the loop is
-llP0112
- 0.07219317556675087.
which is to be compared to the previous value. This controller can tolerate a change in coefficient values of 7.2% compared to the value of lom4% for the optimum controller. The Nyquist plot of the system with this controller is shown in Figure 2 and gives the lower gain and phase margins: Gain Margin = [l,0.794039744513461 Phase Margin = [ O O , -9.88729274575800]. The system can tolerate gain reduction of about 21%. This is an improvement over the previous controller by a factor of about 20, OOO! The phase margin is improved by a factor of about 60. We have already shown the drastic improvement in the parametric stability margin. Therefore this non-optimal controller is far less fragile than the optimal controller, on all counts.
The normalized ratio of change in controller coefficients required for destabilization is:
- 0.01167214151733 llP0112 which shows that the controller, which by design is maximally robust with respect to H , perturbations, is quite fragile with respect to controller coefficient perturbations.
--
1309
Authorized licensed use limited to: Texas A M University. Downloaded on September 1, 2009 at 14:57 from IEEE Xplore. Restrictions apply.
To continue our analysis the Nyquist plot of the system with this controller is drawn in Figure 3. From this we obtain lower gain and phase margin: Gain Margin = [l,0.916666666666671 Phase Margin = IOo, 12.91170327761722]. which are quite poor and would probably be unacceptable
= 1.17938672900662x lo3.
The normalized ratio of change in controller coefficients required to destabilize the closed loop is
= 1.455352715525003x
l P O 1 1 2
-0.4'
-1.2 -1
-0.8
Real IWS
-0.6
-0.4
-0.2
which indicates that closed loop stability is very fragile with respect to controller parameter perturbations. To continue, we draw the Nyquist plot of the plant P(s) with this controller C(s). This is shown in Figure 4 which shows that Gain Margin = [0.5745485,2.585314] Phase Margin = [Oo f 2 4 . 0 5 5 5 ' 1
P(s)=
and the controller designed to tolerate prescribed structured plant perturbations is given as
where
qg = -5.220000000 x lo8, p! qg = -1.190629800 x loll, pg
= 1.468170000 x = 8.153914724x = -1.089211902 x p! = 2.268680248x = -5.104622252 x pg = 1.818763428 x 98 = -1.285270261 x px = 5.698409038x q! = -1.629532689 x loi7, p i = 6.284542925x g o o = -7.937217972 x loi7, po = 6.227740485 x
-i 1;
Real Axis
-1
-557.780980781351 -424.062740635737
1310
Authorized licensed use limited to: Texas A M University. Downloaded on September 1, 2009 at 14:57 from IEEE Xplore. Restrictions apply.
Example 5. optimal control) This example is taken from ([6], p. 51, p. 342). The given plant which is a discrete time model of an X-29 aircraft has transfer function:
(e'
with
The optimal controller designed to minimize the tl norm of a disturbance transfer function is
46z6 45z5 44z4 43z3 42z2 q1.z 40 C(z) = p6Z6 p5-Z -k p4Z4 4- p3Z3 4- p2Z2 -k pi% po where
+ +
+ +
3' -5
-4
-3
-2
Real Axis
41
0.96739955458900, 0.90462192702824, 0.71656397178049, -0.46249566183404, 0.51179070536440, 0.42346983479152, -0.03470824179828 f j 0.08323893024208, 0.07832845106251f j 0.06923045644543
and therefore the controller does stabilize the nominal closed loop system. For our analysis we took the controller transfer function coefficients its a parameter vector and computed the l2 parametric stability margin around the nominal controller. Here we found that
p = 0.01796866210822.
and the optimal controller is determined by minimizing a weighted H 2 norm of the disturbance transfer function IIW(s)P(s)S(s)llz where S(s) is the sensitivity function. In this example the optimal YJBK parameter &(s) is improper and a suboptimal controller is picked after dividing &(s) by the factor (ST+ l)kwith 7 = 0.01 and IC = 2. The controller designed is
C(S)= 46s'
45S5 44S4 4 3 S 3 4-42S2 -k 41s 40 p6ss +p5s5 +p4s4 +p3s3 +p2s2 +pis
where
= 1.0002, 45 = 3.0406, 44 = 8.1210, 43 = 13.2010, 42 = 15.2004, 4 1 = 12.08, 40 = 4.0 p6 = 0.0001, p 5 = 1.0205, p4 = 2.1007, p3 = 5.1403, ~2 = 6.06, pl = 2.0
46
The normalized ratio of change in controller coefficients required to destabilize the closed loop is:
? -!l l P 0 1 1 2
= 3.231207641519448x
This shows how fragile the system is with respect to controller parameter perturbations: perturbations of less than 1 part in 10,000 will destabilize the closed loop. Next, the Nyquist plot of P(z)C(z)is drawn in Figure 5. This gives the closed loop gain and phase margins: Gain Margin = [0.5463916,1.38153348] Phase Margin = [O', f23.03169'].
131 1
Authorized licensed use limited to: Texas A M University. Downloaded on September 1, 2009 at 14:57 from IEEE Xplore. Restrictions apply.
The normalized ratio of change in controller coefficients required to destabilize the closed loop is
P = 3.737066131643626 x
coefficients, we have the following closed loop poles with the perturbed controller: -58362838.6564950, -49.0701313, -1.0093632, -0.9910823 -0.5053510 fj 1.3259448, -0.4947516 f j 1.3200071. This shows that one pole is far in the left half plane. Now we select another small perturbation Ap: -0.3210910254683 x lop4 -0.0090763098117 x 0.0017436867673 x lop4 0.0000492889549 x lop4 -0.0000094691023 x lop4 -0.0000002676642 x A p = 0.0000000514218 x lop4 -1.0000020385893 x 0.3321104789004 x 0.0054305171549 x -0.0018035279761 x -0.0000294904567 x lop4 0.0000097940693 x whose 12 norm is llApli2 = 1.101598411660255 x which is slightly larger than the previous one. If we add this perturbation to the nominal controller coefficients, we have the following new set of closed loop poles
llPO 112
\ \
1.5 1-
, ,
0.5 -
3.
o-
-0.5
-1
- -2 -1.6
-1.4
-1.2 '
-1
-0.8
Real Axis
-0.6 .
-0.4
-0.2
99899133.5177260, -49.0700608, ~ 1.0093632, -0.9910823 -0.5053510 fj 1.3259448, -0.4947516 fj 1.3200071. which includes an RHP pole at about 100 x lo6! This example shows that a slight perturbation in controller coefficients of the optimal controller will result in very large perturbations of the closed loop poles. This is due to the degree dropping of the characteristic polynomial at the perturbation in the highest coefficient.
3. Concluding Remarks
The calculations presented in this paper show that H,, Hz, p and l1 designs can lead to fragile controllers. This means that very small perturbations of the controller coefficients can result in instability. This fragility also shows up usually as extremely small gain and/or phase margins of the closed loop system. Moreover these margins were calculated at the nominal plant; the worst case margins over the set of uncertain plants would certainly be even poorer! The numbers obtained for the parametric margin around the controller coefficients means that there is practically no freedom left to readjust or tune the controller. Thus the control engineer who opts for such an optimal design is forced to either accept a fragile design or reject it altogether. These calculations therefore raise some questions about the usefulness of optimal design as
1312
developed over the last few decades and also the fundamental reasons for this fragility. We make some speculative remarks on the latter. Norm optimization uses up the free YJBK parameter to bring down the norm, resulting in higher order &(s)s. This results in higher order controllers. As we know the stability regions in the parameter space of higher order systems have instability holes and the optimization algorithm can stuff the controller parameter into tight spots close to these holes, since no margin is asked for with respect to the controller in the optimization procedure. Thus in a sense the design transfers the sensitivity from the plant to the controller. These considerations suggest that we look more deeply into some of the following topics:
Prentice-Hall Publishing Company, Englewood Cliffs, New Jersey, 1995. [7] J.C. Doyle, K. Glover, P.P. Khargonekar, and B.A. Francis, State space solution to standard HZ and H , control problems, IEEE Transactions on Automatic Control, Vol. AC-34, no. 8, pp. 831 - 847, August, 1989.
[8] G.J. Balas, J.C. Doyle, K. Glover, A. Packard, and R. Smith, p Analysis and Synthesis ToolBox, The Mathworks, Inc., Natick, MA, 1993.
[9] J.C. Doyle, B.A. Rancis, and A.R. Tannenbaum, Feedback Control Theory, Macmillan Publishing Company, New York, 1992.
1) Parametrize lower and fixed order controllers in a useful way 2) Develop optimization and design algorithms with parametric stability margin of the controller as a side constraint
[lo] S.P. Bhattacharyya, H. Chapellat, and L.H. Keel, Robust Control: The Parametric Approach, Prentice-Hall, Englewood Cliffs, NJ, 1995.
[ll] D. Hinrichsen and B. Martensson, Eds. Control of Uncertain Systems, Bitkhaiiuser, Berlin, 1990.
[12] M. Mansour, S.Balemi, and W. Truol, Eds. Robustness of Dynamic Systems with Parameter Uncertainties, Bitkhaauser, Berlin, 1992. [13] N.K. Bose, A system-theoretic approach to stability of sets of polynomials, Contemporary Mathematics, vol. 47, pp. 25 - 34, 1985. [14] M. Fujita, T. Namerikawa, F. Matsumura, and K. Uchida, p-synthesis of an electromagnetic suspension system, IEEE Fransactions on Automatic Control, vol. 40, no. 3, pp. 530 - 536, March 1995. [15] I. Horowitz, Synthesis of Feedback Control Systems, Academic Press, New York, NY, 1963. [16] S. Jayasuriya, Frequency domain design for maximizing tolerance to disturbances in uncertain systems, in Control of Uncertain Dynamic Systems (S.P. Bhattacharyya and L.H. Keel, Eds.), CRC Press, Boca Raton, FL, 1991. [17] W.M. Wonham, Linear Multivariable Control: a Geometric Approach, Springer-Verlag, Berlin, 1979. [18] K.S. Narendra and R.V. Monopoli, Applications of Adaptive Control, Academic Press, New York, 1980. [19] P.A. Ioannou and P.V. Kokotovic, Adaptive Systems with Reduced Models, Springer-Verlag, New York, 1983. [20] A.C. Antoulas and J.C. Willems, A behavioral approach to linear exact modeling, IEEE Transactions on Automatic Control, vol. 39, no. 12, pp. 1776 - 1802, December 1993.
3) Investigate robust parametric realizations and the robustness of the closed loop system with respect to alternative controller realizations, such as balanced realizations
4) Investigate redundant structures and notions such as design centering in the VLSI literature for realizing the controller 5) Explore the robustness of other design methods such as mixed norm or multiobjective optimization, QFT [15, 161, geometric methods [17], adaptive control methods [18, 191 and the behavioral approach [20].
We are hopeful that some of these lines of research will shed light on the problem discussed here and its remedies.
References
[l] M. Vidyasagar, Control System Synthesis, The MIT Press, Cambridge, MA, 1985.
[2] B.D.0 Anderson and J.B. Moore, Optimal Control: Linear Quadratic Methods, Prentice-Hall, Englewood Cliffs, NJ, 1989. , Control Problem, [3] A.A. Stoorvogel, The H Prentice-Hall, Englewood Cliffs, NJ, 1992.
[4] H. Kimura, Y. Lu, and R. Kawatani, On the struc, control systems and related questions, IEEE ture of H Dansactions on Automatic Control, vol. 36, no. 6, pp. 653 - 667, June 1991. [5] K. Zhou, J.C. Doyle, and K. Glover, Robust Optimal Control, Prentice-Hall, Englewood Cliffs, NJ, 1995. [6] M.A. Dahleh and I.J. Diaz-Bobillo, Control of Uncertain Systems: A Linear Programming Approach,
1313
Authorized licensed use limited to: Texas A M University. Downloaded on September 1, 2009 at 14:57 from IEEE Xplore. Restrictions apply.