Professional Documents
Culture Documents
Abstract: In quantitative feedback theory, plant parameter and disturbance uncertainty are the reasons for
using feedback. They are defined by means of a set J 8 = \P}oi plant operators and a set 3 = {b}of disturb-
ances. The desired system performance is defined by sets of acceptable outputs J^U in response to an input
M, to be achieved for all P e & . If any design freedom remains in the achievement of the design specifications,
it is used to minimise the effect of sensor noise at the plant input. Rigorous, exact quantitative synthesis
theories have been established to a fair extent for highly uncertain linear, nonlinear and time-varying single-input
single-output, single-loop and some multiple-loop structures; also for multiple-input multiple-output plants
with output feedback and with internal variable feedback, both linear and nonlinear. There have been many
design examples vindicating the theory. Frequency-response methods have been found to be especially useful
and transparent, enabling the designer to see the trade-off between conflicting design factors. The key tool in
dealing with uncertain nonlinear and multiple-input multiple-output plants is their conversion into equivalent
uncertain linear time-invariant single-input single-output plants. Schauder's fixed-point theorem justifies the
equivalence. Modern control theory, in particular singular-value theory, is examined and judged to be com-
paratively inadequate for dealing with plant parameter uncertainties.
List of principal symbols and abbreviations disturbance d not measurable, and command input r(t)
measurable. Suppose that p and d are precisely known, that
set of acceptable outputs in response to the output yr due to r is to be yr = r * h (?) (* denotes con-
input u volution), and that the disturbance component \yd\ <m(f)
lower bound is given. Then simply insert a prefilter fx between r andp, with
Bh universal high-frequency boundary of p *fi = h or transform Fx (s) = H(s)/P(s), and inject a signal
acceptable L0(joS), applies for all CJ > coh z in Fig. 1A such that \(z + d) *p\ <m(t). This is so even if
b,b(co) upper bound p is unstable, for a little thought shows that uncertainty in
&= {D} set of disturbances p and/or d must be invoked to render this method invalid.
dB decibels, 20 log10 Uncertainty necessitates feedback around the plant, and a
L(ju)),L0 loop transmission and its nominal value suitable canonic two-degrees-of-freedom structure (see section
LQR linear quadratic regulator 6.1 of Reference 1) is shown in Fig. IB.
LQG linear quadratic Gaussian
LTI linear time invariant
MIMO multiple-input multiple-output
& = {P} set of LTI plant operators, or plant
matrices
P = \Pij] plant matrix
Q = Uu\ withP" 1 = [l/Qu],P a matrix Fig. 1A No feedback needed in absence of uncertainty
QFT quantitative feedback theory
resistor-capacitor oD
RC
single-input single-output G X P
SISO C(orY)
SVT singular-value theory
N
set of acceptable system response func-
tions (matrices in MIMO systems) to com- -1
mand inputs Fig. 1B Canonic feedback structure
= {Td) = as above, but for disturbance inputs C = TR
plant template, set of complex numbers T = (/ + PGY*PGF
L = PG
= set of nonlinear plant operators
Most of the feedback control literature, both classical and
modern, concentrates on realising a desired input-output
1 Introduction relation under the constraint of a feedback structure around
the plant, as if feedback is a tool in filter synthesis. If so, then
This paper is a survey of our work in feedback systems,
feedback theory is just a branch of active network synthesis,
denoted as quantitative feedback theory (QFT), and its com-
and its merits should be compared with other techniques such
parsion with the modern control approach, in particular with
as active RC etc., active RC elements against the transducers
singular-value theory. It is our view that feedback around the
and the constant-gain infinite-bandwidth amplifiers needed in
constrained 'plant' is mandatory only because of uncertainty
in its parameters and/or in disturbances entering the plant. modern control theory. Such comparisons have never ap-
Fig. 1A depicts one of the simplest problems: a linear time- peared.
invariant (LTI) plant operator p with transfer function P(s), The true importance of feedback is in 'achieving desired
whose output is the system output and can be measured, performance despite uncertainty'. If so, then obviously the
actual design and the 'cost of feedback' should be closely
related to the extent of the uncertainty and to the narrow-
Paper 2206D, first received 2nd June and in revised form 2nd ness of the performance tolerances. In short, it should be
September 1982
quantitative. But feedback theory has not been quantitative:
The author is with the Department of Applied Mathematics, Weizmann
Institute of Science, Rehovot, Israel, and the Department of Electrical one hardly finds in the voluminous feedback literature any
Engineering, University of Colorado, Boulder, CO80309, USA quantitative design techniques, or any quantitative problem
IEEPROC, Vol. 129, Pt. D, No. 6, NOVEMBER 1982 0143-7054/82/060215 +12 $01.50/0 215
statements. This is a fantastic phenomenon: so much teaching I T(jto)\ suffice, in the form
and research effort, such a huge literature, but the heart of the
problem is almost ignored. Even the graduate Ph.D. often does
not know the real reason why feedback is used in control. Two examples are shown in Figs. 2a and b, where the final
simulation results fill the envelope (but not when the dis-
2 Linear time-invariant single-input single-output turbance response specifications are more severe than the
feedback systems command response specifications [3]). At any co ojj,
In quantitative feedback theory, the uncertainties formulated the set of points {P(/coi)} is a region in the complex plane,
as a set of plants &= {P} and a set of disturbances J^ = {D}. called the plant template Jr'p(P(j<^>\))- The first step is to
There are specified a set of acceptable command response find these 3~p for a reasonable number of GJ values; see
transfer functions ^~= {T} and a set of acceptable disturbance Reference 3 for examples. In Fig. IB,
FL
T= (2a)
1 +L
(40,10,30,100)0,10,60,200)
(40,10,60,200X40,1,60,200) (2b)
(1,1.60,200 ) (1,1,30,100)
Alog T = A log
(40,1,60,200) (1,1,1,200)
1+L
because there is negligible uncertainty in F (and G). Since
L = GP, the variation in P(jcS) generates via eqn. 2b a
variation in log T(jto). The function of G in L = GP is to
guarantee that the variation in log T is within the amount
allowed by the specifications. Let Lo = GP0 be a nominal
loop transmission at a nominal plant P0-v It is convenient
to find the bounds on Lo in the Nichols chart which achieves
this. Fig. 3 is an example of some bounds on Lo (ju>) for the
plant shown and specifications of Fig. 2a:
Pi = ki/s i = 1,2, b (3)
Pc = K kj<E[aj,bj] aj = 20,50,1,1000
bj = 800,500,60,200000
for/ = l,2,b,c
(4)
30 r
b time, s
Fig. 2 Time envelopes and their u> equivalents
-340 degrees
-180
2x10 7
_X_| (single loop
Nls scale 2) _
o
o I?!
bound on L0(j22) in
o Nichols chart
\P3lfllgl +P32f2\g2
du = + (7)
\P23P\2g2g3 " ( }/
{(1 + Plll) [(1 is any member of the n generated by the and
the 01/in alj,j = 2,
~P23P32g2g3\ ~P2\g\ [Pl 2 2 Q
d
d
33C
G3
tolerances for input u (t)/2 w \ ^33
y3T
1 G3 G3 \ Q33
y33
(nonlinearlyrelated tou(t)
tolerances)
S R "32 I -1 r
33
time
Simulation results
The design details are given in Reference 25. Typical and
extreme step responses are shown in Fig. 9 over the set &.
-0.3 -
ii iii iv
1.0
time s
Fig. 14 Outputs due to simultaneous step inputs, four runs
A 5 5
B 5 5
C 5 5
D 5 5
r, = M,(f)
r2 = M2u(t)
f1 .
0.5
0.5
2 u
time s 0
time s
time S time s
0 ' 5 0 " 5
time s 5 0
-0.5
0
-0.5 -0.5
-0.5 h