Professional Documents
Culture Documents
H Infinity Optimal Degital Redesign Method
H Infinity Optimal Degital Redesign Method
Abstract
The plant input mapping (PIM) method is a digital redesign technique which guarantees the
closed-loop stability for all non-pathological sampling periods. The PIM method, however, has some
possible drawbacks observed in the closed-loop step response: (1) a steady-state error, and (2) a pure
time delay. The first problem might occur when the plant is unstable, and the second one occurs,
when the analog controller is not bi-proper. In the present study, we first modify the original PIM
method such that these two drawbacks are resolved. Although the PIM method does not, in general,
lead to undesirable inter-sample ripples, yet, its formulation does not explicitly address the inter-
sample behavior of the resulting digital control system. Therefore, the main contribution of the paper
is to extend the idea of the PIM method in the context of sampled-data systems, so that the inter-
sample behavior can be taken into account explicitly. For this purpose, we consider a standard
sampled-data setting, in which the H1 norm of the discretization error is minimized. The
effectiveness of the proposed methods is demonstrated through benchmark examples.
r 2006 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.
1. Introduction
0016-0032/$30.00 r 2006 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.jfranklin.2006.02.028
ARTICLE IN PRESS
2 R. Dargahi, A.H.D. Markazi / Journal of the Franklin Institute ] (]]]]) ]]]–]]]
In the indirect or digital redesign approach, a continuous controller is designed first, and
discretized afterwards to obtain an equivalent digital controller. The advantages are that
the existing continuous-time methods are more intuitive and better established than their
discrete time counterparts and, furthermore, the sampling period can be selected based on
the known continuous-time closed-loop bandwidth. A number of classical indirect
methods exist, namely, Tustin’s, step-invariant and matched pole-zero techniques. Such
methods are called local in the sense that they disregard where the controller is actually
going to be used. In other words, the closed-loop performance is not taken into account.
The advantage, however, is that they do not need any information about the plant and can
be applied for digital redesign of already existing industrial analog control systems. A
common difficulty with local methods is that, in order to preserve the closed-loop stability
and performance, a rather fast sampling frequency is usually required. There are occasions,
however, that a sampling rate cannot be selected freely (e.g., when a system is controlled
through the internet). Newer indirect approaches take the closed-loop behavior into
account [3–7], and one of them, even guarantees the closed-loop stability for arbitrarily
selected (non-pathological) sampling periods [4].
The third approach is the, so-called, SD method, in which the digital controller design is
performed in the continuous time framework, taking the inter-sample behavior into
account [8–10]. Despite the fact that this approach is the most realistic way of designing a
digital controller, its drawback is that the sampling period must be decided before knowing
the closed-loop bandwidth.
The proposed method is partly motivated by the idea of a previously existing indirect
method, called the plant input mapping (PIM) method [4]. Another variation of the PIM
method has been studied in [5,6]. Later studies on the PIM method are the order-reduction
methods and a set of patterns in solving the Diophantine equations associated with the
PIM method and the reduction in the number and orders of the required controllers in
[11,12], and a sufficient condition for guaranteeing the uniform-in-time convergence of the
controller blocks and loop signals to their continuous-time counterparts as the sampling
period is reduced, in [13]. Finally, an alternative method was proposed in [14] for reduction
of disturbance effects in PIM digital redesign method.
While the original PIM approach guarantees the closed-loop stability of the discretized
control system for any non-pathological sampling periods, it does not directly address the
inter-sample behavior. Even the methods introduced in [5,6], do not take the continuous
nature of digitally controlled systems into account, because they are casted in a purely
discrete-time H1 setup. Our goal in this paper is two fold; first, we improve the original
PIM method do that the problems with the steady-state error and delayed closed-loop step
response are resolved, and second we introduce a new formulation for the PIM method
which takes the inter-sample behavior of the digital control loop into account. For this
purpose, we minimize the H1 -norm of a particular SD system, representing the
discretization error. The method presented in this paper, can be regarded as a mixture of
the indirect and SD approaches, and takes the advantage of both methods.
required that the closed-loop stability is preserved for all non-pathological sampling
periods. In Fig. 1, M c ðsÞ is the transfer function from the command input to the plant
input (continuous-time plant-input transfer function: CT-PITF). Accordingly, in Fig. 2,
M d ðzÞ is called the DT-PITF.
The original PIM method was first introduced in [4]. The basic idea of the PIM method
was to relate the mentioned analog and digital control systems through their PITFs.
Extensive studies has revealed the effectiveness of the method in preserving the closed-loop
stability of the digital control system, when the analog one is so. The original PIM method,
however, has two drawbacks: (1) The steady-state step responses of the digital and
continuous control loops are not matched when the plant is unstable, and (2) The closed-
loop response of the digital loop shows a delay of one sampling period when M c ðsÞ is not
bi-proper. In this section, we present a slightly modified version of the PIM method to
resolve the two mentioned problems. These modifications are important from practical
point of view, and are explained in steps 3 and 4 of the following PIM design procedure:
Step 1: Select a non-pathological sampling period h with respect to the eigenvalues of the
plant.
Step 2: Calculate the CT-PITF, M c ðsÞ, of the continuous-time loop.
Step 3: Discretize M c ðsÞ into an equivalent DT-PITF, M d ðzÞ, using the classical matched
pole-zero technique. Unlike the original PIM method, and in order to prevent any time
delay in the closed-loop response, we add a number of extra zeros at z ¼ 1 such that
M d ðzÞ becomes bi-proper. It should be noted that in the traditional matched pole-zero
technique, the relative degree of M d ðzÞ is forced to be equal to 1, when the M c ðsÞ is strictly
proper. It is, then, easy to show that
N d ðzÞ :¼ Pd ðzÞM d ðzÞ 2 RH1 , (1)
where N d ðzÞ is the closed-loop transfer function and Pd ðzÞ is the hold-equivalent discrete-
time model of the plant. As usual, RH1 denotes the space of rational, proper and stable
transfer functions.
ARTICLE IN PRESS
4 R. Dargahi, A.H.D. Markazi / Journal of the Franklin Institute ] (]]]]) ]]]–]]]
Step 4: In order to make sure that the steady-state step response of the outputs are the
same, M d ðzÞ is scaled so that limz!1 N d ðzÞ ¼ lims!0 N c ðsÞ.
Step 5: Solve the following Diophantine equation for the polynomials aðzÞ and bðzÞ:
aðzÞbðzÞ þ bðzÞaðzÞ ¼ dðzÞ, (2)
where aðzÞ and bðzÞ, respectively, are the numerator and denominator of the hold-
equivalent discrete-time model of the plant and dðzÞ is the denominator polynomial of
M d ðzÞ. We then will have q½aðzÞ ¼ p n and q½bðzÞpp n where p is the order of the DT-
PITF and n is the numerator degree of Pd ðzÞ. Here, q denotes the degree of its polynomial
argument.
Step 6: Let uðzÞ and vðzÞ be arbitrary Hurwitz polynomials where q½uðzÞ ¼
q½vðzÞ ¼ p n, and mðzÞ is the numerator of M d ðzÞ, then the digital controllers are
obtained as
mðzÞ bðzÞ uðzÞ
Rd ðzÞ ¼ ; F d ðzÞ ¼ ; K d ðzÞ ¼ . (3)
vðzÞ uðzÞ aðzÞ
3. Example 1
Katz [15] has shown that the SISO analog control system in Fig. 4 with
863:3 2940s þ 86436
Pc ðsÞ ¼ and K c ðsÞ ¼ (4)
s2 ðs þ 294Þ2
is hard to discretize. In fact, except for very small sampling periods, all of the classical
discretization methods lead to unsatisfactory and/or unstable results. In this example, we
demonstrate the effectiveness of the modified PIM method, by selecting a sampling period
of h ¼ 0:039 s which is large compared with the closed-loop settling time of 0:25 s. Also
note that the Pc ðsÞ is unstable and K c ðsÞ is not bi-proper. Hence, the original PIM method
leads to a delayed output, compared to the continuous system, together with a
considerable steady-state error.
The controllers obtained by the modified PIM methods are shown in Table 1. The
improvement of the closed-loop response with the modified PIM method is clearly depicted
in Fig. 3.
Table 1
Comparison between the digital controllers obtained using different methods
h ¼ 0:039
1.8
1.6
1.4
1.2
Plant Output
0.8
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6
Time (Seconds)
Consider a pre-designed 1-DOF analog control system shown in Fig. 4, and a given non-
pathological sampling period h. Our objective is to discretize the controller K c ðsÞ into an
equivalent digital controller K d ðzÞ, by minimizing the H1 -norm of the discretization error.
In order to take the inter-sample behavior into account, the discretization problem is re-
casted in a SD setting, depicted in Fig. 5. The general idea behind this new formulation is
explained below:
Let r and d, respectively, denote the command input and disturbance applied to the
closed-loop system. Also suppose e1 and e2 are the weighted error signals, due to the inputs
r and d. In fact, e1 and e2 , represent the closed-loop discretization error for some given
inputs.
Now, this problem can be put into a standard SD framework as shown in Fig. 6, where
G is the continuous-time generalized plant, w is the exogenous input, u is the control input,
z is the controlled output, and y is the measured output. Furthermore, the sampler S and
the zero-order hold H are assumed to be synchronized with the sampling period h.
ARTICLE IN PRESS
6 R. Dargahi, A.H.D. Markazi / Journal of the Franklin Institute ] (]]]]) ]]]–]]]
The H1 -norm of a continuous-time transfer function equals the L2 ðRþ Þ-induced norm
of the corresponding linear system. For SD systems, there exist no transfer functions in the
normal sense; thus the H1 -norm of a SD-system is defined as the L2 ðRþ Þ-induced norm
of a linear time-varying operator. This definition captures the system’s behavior, in
between the sampling instants.
By defining the vectors of exogenous inputs w (consisting of r and d), and the error
vector signal z (consisting of e1 and e2 ), and assuming that w is in L2 ðRþ Þ, the asymptotic
closed-loop stability implies that z will be also in L2 ðRþ Þ. Our objective, then, is to find a
digital controller K d ðzÞ such that the H1 -norm of the SD system of Fig. 6, from w to z, is
minimized.
Compared with the standard setup of Fig. 6, the generalized plant of Fig. 5 is given by
2 3
F 2N cF 1 F 2 Pc F 2 Pc
6 F 3 7
G ¼ 4 F 3M cF 1 0 5. (5)
F1 Pc Pc
Since Pc ðsÞ and F 1 are strictly proper, it can be easily seen that the generalized plant will
be of the form
(6)
ARTICLE IN PRESS
R. Dargahi, A.H.D. Markazi / Journal of the Franklin Institute ] (]]]]) ]]]–]]] 7
This particular form of the D-matrix, ensures that the problem is technically solvable
using the standard SD algorithms (see Appendix A).
1. F 1 is a rational, strictly proper and asymptotically stable pre-filter. The reason for
inclusion of F 1 is to make the induced-norm of the sample-and-hold operation finite
[10].
2. F 1 also attenuates the higher frequencies of the command input, to make it compatible
with respect to the selected sampling period h. More precisely, in order to make the
magnitude of F 1 steeper around the bandwidth, we choose F 1 to be in the form of a
second order low-pass filter with the bandwidth ob ¼ lp=h.
1
F 1 ðsÞ ¼ . (7)
ð1=o2b Þs2 þ ð2=ob Þs þ 1
The role of l is to act as a design parameter, which modifies the bandwidth of F 1 . In
general, by selecting l to be less than one, a better low frequency tracking can be
attained, and by selecting l around 1, or somewhat larger, tracking over a larger range
of frequency is possible, yet with less accuracy. In general, one needs to perform some
trials and errors on l, in order to obtain a reasonable result (e.g., a better step response).
3. F 2 and F 3 are also rational, proper and asymptotically stable filters that serve as
weighting functions in the closed-loop minimization process. It is customary to select
such weighting functions in the form of
ðM 0:5 s þ ob Þ2
F 2 ðsÞ; F 3 ðsÞ ¼ , (8)
ðs þ A0:5 ob Þ2
which provides high and low frequency gains being equal to 20 log10 M, and
20 log10 A, respectively. By increasing the magnitude of F 2 and F 3 at certain frequency
ranges, one forces the closed-loop error be smaller in those frequencies. One might
prefer to use the same weighting functions for e1 and e2 , i.e., F 2 ¼ F 3 . By selecting M as
a relatively large value (say, M ¼ 100), and A small (say, A ¼ 1, one can force more
penalty on the discretization error at frequencies lower than ob .
The optimal PIM method directly addresses the minimization of the inter-sample
behavior by proposing a SD formulation. This does not necessarily mean that the original
and modified PIM methods suffer from undesirable inter-sample behavior. In fact, the
following discussion shows that, under natural conditions, one may expect a ripple-free
response even from the non-optimal PIM methods. In general, there are three causes for a
possible undesirable oscillation in the response of any LTI digital control system, namely,
(1) the excessive oscillations of the control input to the plant, (2) low-damped modes of the
continuous-time plant, excited by the successive step change signals from the zoh element,
ARTICLE IN PRESS
8 R. Dargahi, A.H.D. Markazi / Journal of the Franklin Institute ] (]]]]) ]]]–]]]
(3) those frequency contents of the external disturbances which are within the closed-loop
bandwidth.
In order to reduce the effects of the last two sources of oscillation, if exist, the sampling
period must be selected to be sufficiently small. This applies no matter what design method
is used, direct or indirect, PIM or non-PIM. The first source of oscillation, however, can be
avoided by a thoughtful design so that the resulting DT-PITF does not bear any unstable
or low-damped poles. Unlike conventional discretization techniques, the PIM approach
provides a direct control on the location of these poles by discretizing the original CT-
PITF into a desired DT-PITF with well-damped poles. Therefore, even in the case of non-
optimal PIM methods, the control input to the plant is guaranteed to be free of excessive
oscillations and no undesirable inter-sample behavior will exist.
7. Example 2
Here, we demonstrate the effectiveness of the optimal PIM method using the analog
control system given in Example 3. With the selected sampling period of 0:039 s, the
modified PIM method provided three controllers, shown in Table 1. For the optimal PIM
procedure, we select the weighting functions F 1 and F 2 ¼ F 3 with the parameters M ¼ 100
and A ¼ 1. Based on a few simple trial and errors, it can be seen that, l ¼ 1:2 provides a
desirable step response for the digital control loop. Furthermore, the optimal PIM method
gives rise to a single high order controller, whose order can be reduced by cancelling the
very close poles and zeros of the controller (see Fig. 9). The resulting reduced order
controller is shown in Table 1. Figs. 7 and 8, show the unit step responses of the original
analog system and the digital control systems obtained using Tustin, modified, and optimal
PIM methods. It can be seen that the method of Tustin leads to a totally unacceptable
result, while the modified and optimal PIM methods provide results close to the
Optimal PIM
1.6
Modified PIM
1.4
Plant Output
1.2
0.8
0.6
0.4
0.2
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time (Seconds)
Tustin
3
Optimal PIM
2 Modified PIM
1
Plant Input
-1
-2
-3
-4
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time (Seconds)
Pole−Zero Map
1
0.8
0.6
0.4
Imaginary Axis
0.2
-0.2
-0.4
-0.6
-0.8
-1
1 0.8 0.6 0.4 0.2 0 0.2 0.4 0.6 0.8 1
Real Axis
Fig. 9. Pole-zero map of the digital controller obtained using optimal PIM.
continuous case (Fig. 9). The main advantage of the PIM method over the modified PIM
method is a better step response, together with a less number of digital controllers.
8. Conclusions
PIM, which is known to be efficient in discretizing analog control systems. The main
objective of this paper is to modify the original PIM method, so that two of its drawbacks,
i.e., steady-state error and delayed closed-loop step response are resolved and,
furthermore, extend the idea of the PIM method so that the inter-sample behavior is
explicitly taken into account.
The main idea of the optimal method was to find a digital controller such that the H1 -
norm of a particular SD system, representing the discretization error, is minimized (Fig. 6).
With the proposed optimal PIM method the number of digital controllers are reduced to
one, instead of three. Such improvements lead to more efficiency and simplicity in the
implementation stage.
It was shown, through examples, that the original, modified and optimal PIM methods
provide extremely better results compared against the conventional Tustin’s method. In
particular, the optimal PIM method gives rise to a better closed-loop response with less
number of controllers.
(9)
Here, the first assumption is D21 ¼ 0, because the sampler must be lowpass filtered for kTk
to be finite and the second is D11 ¼ 0; this is for technical simplification [10]. Let L denote
the continuous lifting operator mapping w to w, then the relation between the lifted input,
w ¼ Lw, and the lifted output, z ¼ Lz, is z ¼ T w, where T :¼LTL1 is the lifted system
(Fig. 10); this is exactly the closed-loop map w 7! z of Fig. 6. Now, G is LTI in discrete-
(10)
Rh
with Ad ¼ ehA , B2d ¼ 0etA dtB2 , and the operators given by
Z h
B1 : K ! E; B1 w ¼ eðhtÞA B1 wðtÞ dt, ð11Þ
0
C 1 : E ! E; ðC 1 xÞðtÞ ¼ C 1 etA x, ð12Þ
Z t
D11 : K ! K; ðD11 wÞðtÞ ¼ C 1 eðttÞA B1 wðtÞ dt, ð13Þ
0
Z t
D12 : E ! K; ðD12 vÞðtÞ ¼ D11 vðtÞ þ C 1 eðttÞA B2 dt, ð14Þ
0
where E denotes any finite-dimensional Euclidean space, and K ¼ L2 ½0; hÞ. Since L and
L1 are isometries, we have that kTk ¼ k T k, that is, the L2 ðRþ Þ-induced norm of T
equals the ‘2 ðZþ ; KÞ-induced norm of k T k. The advantage of this lifting construction lies
in that T is now LTI (albeit with infinite-dimensional input and output spaces).
Step 2: Under the assumption g4kD11 k, it can be shown that SD H1 control problem,
kTkog, is equivalent to the discrete-time H1 problem, kT eq;d kog, where T eq;d is the
closed-loop system composed of a discrete-time generalized plant G eq;d and the same K d .
The H1 discretization G eq;d Fig. 11 is given by
(15)
with
Step 3: Let T eq;d be the closed-loop map with discrete-time plant G eq;d and controller K d .
The solvability condition for the standard discrete-time H1 problem should be tested:
There exists a K eq;d to achieve internal stability and kT eq;d k1 og. If the test fails, g should
be increased, otherwise, g should be decreased together with retrogress to step 2.
Step 4: The discrete-time H1 problem kT eq;d kog should be solved for a stabilizing K d
when a satisfactory g is found. This K d also achieves kT zw kog.
References
[1] G.F. Franklin, J.D. Powell, M.L. Workman, Digital Control of Dynamic Systems, Addison-Wesley,
Reading, MA, 1998.
[2] K. Ogata, Discrete-Time Control Systems, Prentice-Hall, Englewood Cliffs, NJ, 1987.
[3] B.D.O. Anderson, J.P. Keller, Discretization techniques in control systems, Control and Dynamics System,
vol. 66, in: C.T. Loendes, (Ed.), Academic Press, New York, 1994, pp. 47–112.
[4] A.H.D. Markazi, N. Hori, A new method with guaranteed stability for discretization of continuous-time
control systems, Proceedings of American Control Conference, Chicago, 1992, pp. 1397–1402.
[5] A.H.D. Markazi, N. Hori, Discretisation of continuous-time control systems with gauranteed stability, IEE
Proc. Control Theory Appl. 142 (4) (1995).
[6] A.H.D. Markazi, N. Hori, Reply to comments on discretization of continuous-time control systems with
guaranteed stability, IEE Proc. Control Theory Appl. 144 (3) (1997) 277–279.
[7] N. Rafee, T. Chen, O.P. Malik, A technique for optimal digital redesign of analog controllers, IEEE Trans.
Control Systems Technol. 5 (1) (1997) 89–99.
[8] B. Bamieh, J.B. Pearson, The H2 problem for sampled-data systems, Systems Control lett. 19 (1992) 1–12.
[9] B. Bamieh, J.B. Pearson, A general framework for linear periodic systems with application to H1 sampled-
data control, IEEE Trans. Automat. Control 37 (1992) 418–435.
[10] T. Chen, B. Francis, Optimal Sampled-Data Control Systems, Springer, Berlin, 1996.
[11] C.A. Rabbath, N. Hori, On a comparative study of digital redesigned methods, Proceedings of American
Control Conference, Chicago, Illinois, 2000, pp. 1154–1158.
[12] C.A. Rabbath, N. Hori, Reduced-order PIM methods for digital redesign, IEE Proc. Control Theory Appl.
150 (4) (2003).
[13] C.A. Rabbath, N. Hori, N. Lechevin, Convergence of sampled-data models in digital redesign, IEEE Trans.
Automat. Control 49 (5) (2004).
[14] H. Shimamura, N. Hori, S. Nazhand, Reduction of a disturbance effect in PIM digital redesign, IEEE
International Conference on Methods and Models in Automation and Robotics, Poland, August 2003, pp.
1291–1296.
[15] P. Katz, Digital Control Using Microprocessors, Prentice-Hall, Englewood Cliffs, NJ, 1981.