Sunteți pe pagina 1din 5

166 AUTOMATIC

VOL.
CONTROL,
TRANSACTIONS
ON
IEEE AC-32, NO. 2. FEBRUARY 1987

Step 2 of the main loop consists of iteratively performing the above complex root convergence (from Soukup [4, Theorem II]) with a
operations (computing { b , } ; , { x , } : , q(z,,,) and zm+ until I~(Z,,,-~)I < E. polynomial deflation technique developed by the author. The algorithm
Then this latest value of the root estimate z, = zm-I is taken to be the has guaranteed convergence, works wellwith multiple roots. and is
correct estimate of the true root sk. straightforward to implement. This procedure should prove useful to
Step 4 of the main loop, the polynomial deflation. is accomplished by those researchers doing control system design or analysis who require a
synthetically dividing the factor (z - sn) out of the previous form of the stable numerical method for complex-coefficient polynomials.
polynomial
ACKNOWLEDGMEYT

The author wishes to thank two reviewers for their helpful comments,
as well as Dr. N. H. McClamroch for his excellent editorial assistance.
to produce the new reduced form of

REFERENCES

,=O D. H. Lehmer, "A machine method for solving polynomial equations." J . Asroc.
Comput. Machinery. vol. 8, pp. 151-162, 1961.
The net effect is to compute new values of the coefficients { nc } ,"- I so that R. W. Hamming, Numerical Methods f o r Scienlists and Engineers, 2nd Ed.
they may be used for f(z) in the determination of the next root. This is New York: McGraw-Hill. 1973.
D. M. Young and R. T. Gregory, A Survey of .Vumerical Marhemafics, Vol.
done as follows (for order k > 1): assign a ; _ , = u k and for i = 1, . . ., I. Reading. MA: Addison-Wesley, 1972.
k , form a,.-, = U W + I + S A , U ; _ , + ~ . J. Soukup, "A method for finding the roots of a polynomial," .Vumerische
This completes the description of the proposed algorithm. Math.. vol. 13, pp. 349-353, 1969.

m. EXAMPLEAPPLICATION
This section of the note illustrates the application of this root-finding
method to a specific polynomial. The algorithm has been extensively
tested, and it performs very well. This portion of the note is not intended
as a proof or verification of the proposed procedure. but is merely an State Observation of Nonlinear Uncertain Dynamical
explanation of the typical performance of the algorithm. Systems
The polynomial for this example is

f(z)=zs+(l.2+j0.9)z7+(-0.74+j0.83)z6+(1.502-j0.535)z5 B. L. WALCOTT AND S. H. ZAK


+(-0.7795+j1.8515)~~+(-0.111-jl.5095)~' Abstract-This note proposes new types of observers for nonlinear
+ (0.88775 +jO.86925)z2+ ( - 0.49175 +j0.07175)~ dynamical systems subjected to bounded nonlinearities or uncertainties.
The design of these observers utilizes techniques related to variable
+ (0.0735 -j0.0735). structure systems theory. A measure for the rate at which the estimates
converge to the actual states is derived.
I t has eight roots, at 0.5 f j0.5, -0.5 - j0.5, 0.3, -2, + j , and a
double root at -j0.7. The termination tolerance E is set at IO-'*. about
lo3times greater than the relative precision of the computer being used. I. INTRODUCTION
The algorithm was run for six values of the convergence factor {: 0.5,
State observation of nonlinear dynamical systems is a topic of recent
0.6, 0.7, 0.8, 0.9, and 1.0, and it successfully converged to the correct
discussion in the literature [1]-[7]. Thau [l] incorporates the nonlineari-
root values in each case. The algorithm required the same number of
ties of the plant into the dynamics of his observer design plus requires that
iterations in Soukup's procedure to obtain a root for each value of j- 2
the nonlinearities be Lipschitz in the states. Thau's results were
0.6, with anaverage number of seven iterations for each root; the
generalized by Kou el a/. [5], yet the nonlinearities of the plant still
maximum number of iterations was 19, and the fastest convergence
appear in the observer dynamics. Bestle and Zeitz [2] and Krener and
occurred for the last root found (with two iterations necessary).
Respondek [4] transform the nonlinear plant into "observable canonical
Several comments maybe made about the use of the proposed
form" from where observer design is performed. However. finding an
algorithm.
appropriate nonlinear. time variable, one-to-one transformation is nontri-
1) The value of the tolerance E should be several decades larger than the
vial or often impossible by their own admission. Moreover, knowledge of
maximum relative precision of the computer running this software, to
the nonlinearities of the plant must be exact since itis needed in the
ensure that roundoff effects do not alter the convergence of the routine.
computation of the observer dynamics. Baumann and Rugh [3] utilize an
As long as this condition is met, convergence to each root is guaranteed,
extended linearization technique to produce anobserver which, when
as pointed out in [4].
linearized about any of a family of equilibrium points. has locally
2) With regard to the convergence factor 5: any value from 0.5 to 1 .O
invariant eigenvalues. Hereagain, exact knouledge of the plant's
works, but operation is most stable and rapid for a value in the middle of
nonlinearities and, furthermore, the first derivatives of these nonlineari-
this range. This factor is not critical, even though it plays an important
ties must be known in order to calculate the gain function of the observer.
role in Soukup's procedure.
The goal of this note is to present an observer design which does not
3) In [4], Soukup mentions three aspects of his procedure which
necessitate exact knowledge of the system nonlinearides. This aim is
deserve repeating: the algorithm performs much better for the case of
accomplished by utilizing techniques prevalent in variable structure
multiple roots than other procedures; convergence is faster for roots
systems ( V S S ) theory.
inside the unit circle: and the algorithm is capable of correct convergence
in spite of numerical errors which may be made in the iteration process.
Manuscnpt received Februae 26, 1986: revised September 2, 1986. Paper recorn-
mended by Associate Editor. T. J. Tam. This work was supported in pan by the David
Iv. CONCLUSION
Ross Grant of Purdue University.
The authors are with the School of Electrical Engineering, Purdue Vniversity. West
This note has presented an algorithm for determining the roots of a Lafayette, IN 47907.
polynomial with complex coefficients. It combines an iterative method of IEEE Log Number 8612207.

0018-9286/87/0200-0166$01.00 0 1987 IEEE


IEEE TRANSACTIONS ON AUTOMATIC CONTROL. VOL. AC-32. NO. 2. FEBRU.4RY 1987 167

STATEMENT
11. PROBLEM Lyapunov function candidate is given by

Consider the class of systems modeled by the following equations:

i ( t ) = A x ( t ) + f ( t x,
, u)
y(t)=C.x(t) (1) which simplifies to

where x E R", u E EP"', y E jP', the matrices A and C are of V ( e ) =-erQe-211Cellp-2erCrh(t, x). (6)
appropriate dimension, and the matrix C is of full rank. The functionf(t,
x , u ) can be construed as the uncertainties or nonlinearities in the plant. Taking the Euclidean norm of the last term of (6) and noting assumption
For existence purposes. we require that f(r, x, u ) be continuous in x . A3. yields
Our problem is to design an observer with inputs y and u whose output
f will converge to x (Le., lim (2(t) - x ( t ) ) = 0 ) . V ( e ) s -eroe-211cellpi2iiceI.p<o. (7)
I-m
Therefore, the Iimr-- e(t) = 0. Q.E.D.
~ nPRELIMINARY
. ASSUMPTIONS Theorem 1 shows that error difference between the estimate and the
true state asymptotically tends to zero. However, it is desirable to know
Consider the following three assumptions pertaining tothe system
the rate at which the estimate converges since, if the time response of the
denoted in (1). observer is of the same order or greater than the system's response time,
A 1: The pair ( A , C) is detectable which implies that we can find a
the observer is of little use in an observer-controller configuration.
matrix K E a n x p such that o[Ao] C C- where A0 = A - KC and C-
Dividing inequality (7) by (4) yields:
is the open left-half plane.
A2: There exists a symmetric, positive definite matrix Q E ;I""", and
function h where h ( .. ., .):;i'+ X F i n X Fi"' B P such that
+

f ( t , x, u ) = P - ' C r h ( f ,x, U )

where P is the unique, positive definite solution to the Lyapunov equation

A,TP+PAo= - Q , where 9 is the minimum eigenvalue of P - ' Q [SI. Thus, if we consider


eTPe to be a measure of the magnitude of the error, then the error will
A3: There exists a positive scalar valued function, p such that approach zero in magnitude exponentially, with a rate of decay that is at
least as fast as e-?'.

V. INSERTION OF A BOUNDARYLAYER

A shortcoming of the observer described by (2) is that the term S(2,y ,


IV. MAIN RESULT p ) is discontinuous. Thus, analysis of (2) must be performed via treating
the observer as a generalized dynamic system [9]. Additionally, such a
Let the error difference betwen the observer estimate and the truestate term will cause the undesirable phenomenon known as "chattering" in
be denoted by VSS systems. Hence, it is desirable to find a gain which is continuous in
the error e and also assures us that the estimates will converge at least
e(t)=i(t)-x(f).
asymptotically to some arbitrarily small neighborhood of the true states.
Consider the following nonlinear observer dynamical equation: To fulfill these requirements wewill borrow a technique from VSS
control design known as the "boundary layer" control strategy [lo], [12],
(2) p)+Ky
i=A,P+S(P,y, 1131.
The boundary layer observer which will satisfy the aforementioned
stipulations is
where
P=Ao.?+Ky+S(P, y , p )
where An and K are as described in assumptions AI and A2 and

and N = { e:Ce = 0 ) . Note that this observer design incorporates only


the bound of the nonlinearities and/or uncertainties, p ( t , u ) , and does not
require exact knowledge concerning the structure of the plant nonlineari-
ties except that they satisfy assumption A2. Wenow state the ensuing
theorem. Note that if E = 0, then our boundary layer observer is identical to the
Theorem I : Given system ( I ) and the observer governed by (2): if observer denoted by (2). Thus, the error difference between the observer
assumptions AI-A3 are valid, then limr-a (2(t) - x ( t ) ) = 1imrdme ( t ) estimate obeying ( I O ) and the true state of (1) is governed by
= 0.
Proof: The error difference between the output of the observer and e = A ope)+-yP
S, (-i' ,C r h ( r , x). (12)
the true state obeys the following differential equation:
Prior to proving that this observer satisfies our stipulations we must
introduce a few facts and one additional assumption which are essential to
the proof.
Firstly, since V ( e ) = e7Pe is a continuous, positive definite Lyapunov
Consider the following positive definite Lyapunov function candidate function for the error equation (3). we assume the following 1111. There
exist continuous: strictly increasing functions, y l ( . ) : : l - --t 3 + fori = I ,
V(e)=erPe (4) 2, 3 such that

where P isdefinedin assumption A2. The time derivative of this yi(0)= 0 i = 1, 2 , 3


168 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. AC-32, NO. 2, FEBRUARY 1987

lim y , ( r ) = w i=l, 2 shaft, g is the gravitational constant, and 0 is the angular displacement of
1-03
the plumb bob from vertical. If we let I = 10 m, m = 1 kg, and g =
10 m/s2 we obtain the following differential equation:
rl(llell)s Ve)Ir2(lleII)
ij= -sin 8. (21)
P(e)I -Y3(llell).
For the boundary layer observer governed by (lo), (1 l), we will again
Letting x I = 0 and x2 = 6 yields the ensuing state variable description:
employ the Lyapunov function candidate V(e) = eTPe. To avoid
confusion with V ( e )as defined in ( 3 , let L(e) denote the time derivative
of V(e) for our boundary layer observer. Note that if IlCep )I > E , then
We will choose the scalar output:
U e ) I -n(llell).
However, if 11 Cep 11 IE , then

We note that this system is compatible with (1). Also, observe that it is not
necessary to expand the sinxI term in a power series and extract the linear
I -erQe-211CepII-2e'CTh(t, x)+21(CepII portion provided that the normal system in its present form is detectable.
One may readily confirm that the pair ( A , C) is observable thus, we may
= -Ys(llel0+2f (13) arbitrarily assign the spectrum of A . = A - KC. Ifwe select the
spectrum { - 1, - 1 }, the corresponding gain matrix K which will satisfy
which implies that for all e E W" this requirement is K = [ 1 11 and the corresponding A0 matrix is
1
U e ) s -ydllel0+2E. (14)
We further assert that since the functions yI and y~ are strictly increasing, J
then yrl(.):W+ L;3+ and y?'(.):W+ + a+ exist and are also strictly
+
Next, we must find a matrix Q and function h ( . ) to satisfy assumption
increasing. Additionally, if we let I E lim,,,y3(rJr theny,'(.):[O, I ) +
A2. One may easily verify that
R+ exists, is continuous, and is strictly increasing. If we dictatethe r

assumption that
A4: 2~ < I, then we are ready to state the following theorem.
Theorem 2: Consider the observer error obeying (12). If assumptions
Al-A4 are valid, then the error is uniformly bounded. That is, if e(.):[to, and
t,] -, W" and e(to) = eo is a solution to (12), then ((eo\(Ir implies that
Ile(t)ll s d(r) for all t E [to, tl] where - h ( f , x , u)=sin xI

meet the stipulations of A2. The corresponding positive definite matrix P


which satisfies the Lyapunov equation A,TP + PAo = - Q is
r l
and

R=y;1(2+ (16)
.=I: : I
L A

Therefore, in the notation of assumption A2, h ( t , x, u ) = -sin X I and


Furthermore,the error is uniformly ultimately bounded. That is, if Ilsin x111 I1. Hence, for example we can choose p = 2. It is important to
e(.):[to,0 0 3 R" and e(t0) = eois a solution to (12) with lleoll 5 rthen, interject that although many mathematical models of systems have
for a given d > -yc1(y2(R)) nonlinearities which are radially unbounded in the states, in practice. most
state variables represent physical entities which are bounded and thus, our
Ile(t)llsd for dl t = t , + T ( d , r) (17) observer design can be applied. The differential equation which governs
the dynamics of our observer is
where

if r < R

where
r

and
R=Y;I(Yl(d)). (19) An estimate for the speed of convergence of our observer is obtained
from the minimum eigenvalue of P - I Q = min [I, 31 = 1. Therefore, the
Proo;f: The proof is nearly identical to the proof of the theorem observer output denoted by (26) will converge to the true estimate with a
presented in Corless and Leitmann [ 121. The concept of uniform stability rate which is at least as fast as e - ' . Computer simulation of the system
of a neighborhood about e = 0 also follows immediately from [12]. Thus, described by (22) as well as the observer obeying (26) can be found in
by decreasing the value of E we can attain a boundary layer observer Walcott and Zak [7]. Simulation of the boundary layer observer for (22) is
response which is arbitrarily close tothat of our original observer denoted also contained in [7].
by (2). Q.E.D.
w. STATE OBSERVATION IN THE PRESENCE OF INPUT
VI. EXAMPLE: THE SIMPLE PENDULUM DISTURBANCES
Consider the ideal simple pendulum governed by the equation Consider the class of systems governed by
mIe+mg sin e=o (20) x=Ax+af(r, x ) + B ( u ( t ) + v ( t ) )
where m is the mass of the plumb bob, I is the length of the pendulum y= cx
IEEE TRANSACTIONS
AUTOMATIC
CONTROL.
ON VOL. AC-32, NO. 2, 1987
FEBRUARY 169

where a f ( t , x ) represents the uncertainties of the plant and u ( f ) depicts VIII. EXAMPLE: BOUNDEDINPUTDISTURBANCE
input disturbances. In this less general case of system (1) we will first
assume that the so-called matching conditions [12] are satisfied. That is, In this section we will apply our observer design methods to the system
there exists a function w where w( .. -):fly
X R” 8“ such that
+
described by

B w ( t , x ) = B u ( t ) + d f ( Xf ,) .

We maynow modify the three assumptions introduced in the main


section to accommodate the problem of observing the states of system with output
(27).
A I ’: The pair (A, C ) in (27) is detectable. Thus, again we may find a
matrix K such that a[Ao] E C - where A0 A - KC. e
A 2 ’: There exists Q symmetric and positive definite such that for some
matrix F E Rmxp, where

~-~cTFT=B (28)

where P i s the unique, positive definite solution to the Lyapunov equation


and o ( r ) is normal with zero mean and a variance of 1. Hence, u ( t )
represents the output of a limiter which has an input of white noise. Since
a limiter behaves much like a low-pass filter, u ( t ) may be construed as
A3’: There exists a positive scalar-valued function, p such that band limited or so-called “pink noise.”
Since the control u ( t ) is known and appears in the system and observer
)Iw ( r , x)ll s p ( t ) vz E W!+and Vx E 3”. equations, it has no effect upon the error difference between the estimated
and actual states. Thus, without loss of generality we will assume u ( t ) to
Consider the following nonlinear state observer described by the be zero. Note that under this assumption the plant described by (31) is
following equation: equivalent to the simple pendulum considered earlier except that the
nonlinear function of the states, f ( t , x, u ) = -sin x ! , has now been
i=Ao.f+S(P,y, p ) + K y + B u (29) replaced by u ( t ) which is also bounded by 1 in magnitude. Therefore, this
system satisfies assumptions Al’-A3’and we may employ thesame
where observer design as was introduced in Section VII. Digital computer
simulation of both the discontinuous and boundary layer observer for (3 1)
-P-I CrFrFCe may be found in 171.
p(t, u ) for all e $ N g { e : FCe=O}

for all e E N . IX. SUMMARY

This note has presented a variable structure observer design methodol-


Again, let the error between the observer output and the true state be
ogy for systems containing completely observable linear parts and
given by
bounded nonlinearities or uncertainties. A minimum estimate for the rate
e(r)=f(t)-x(f). of convergence of theobserver error tozero was also given. An
alternative boundary layer observerstrategy which offered a gain function
We may now state the following theorem. which was continuous and contained less energy was shown to provide a
Theorem 3: Given the system obeying (27) and the observer governed response that could be made arbitrarily closeto that of the original
by (29), if the assumptions AI’-A3’ are valid, then lim1-- e ( t ) = 0. observer design. A new type of asymptotic state observer for uncertain
Proof: The proof of Theorem 3 follows the same arguments as the systems satisfying matching conditions was also proffered. Finally, the
proof of Theorem 1, thus, it has been omitted. results of two substantiating simulation examples can be found in [7].
Here again, an estimate for the rate of convergence of the distance
measure, e T P e , to zero is e-7‘ where q is the minimum eigenvalue of ACKNOWLEDGMENT
P-’Q.
We may also obtain a system theoretic interpretation of assumption A2’ The authors are grateful to Prof. M. J. Corless for fruitful discussions
by employing the results of [14]. Let us introduce the function on the subject and are indebted to G . R. Widmann for his helpful
comments.
G(s)=FC[sZ-Ao]-’B (30)

and the following definition concerning strictly positive real functions. REFERENCES
Definition: G(s) is strictly positive real if there exists some E > 0 such F. E. Thau. “Observing the state of non-linear dynamicsystems,” Inr. J. Contr.,
that G(s - E ) is positive real. vol. 17?pp. 4 7 1 4 7 9 , 1973.
We may now state the following fact. D. Bestle and M. Zeitz, “Canonical form observer design for npn-lineartime-
Facr: If the triple (Ao,B , FC) is controllable and observable, and there variable systems,” Inr. J. Conrr., vol. 38, no. 2, pp. 419-431, 1983.
W . Baumann and W. Rugb, “Feedback control of nonlinear systems by extended
exists a matrix Fsuch that G(s)as defined in (30) is strictly positive real, linearization,” IEEE Trans. Automat. Conrr., vol. AC-31, no. I , pp. 40-47,
then there exists Q such that assumption A 2 ’ is valid. 1986.
Proof: See Steinberg and Corless [14). A. I. Krennerand W. Respondek, “Nonlinear observers and linearizableerror
For the case when the number of inputs and outputs are equal (i.e,, m dynamics.” SIAM J. Conrr. Optimiz., vol. 23, no. 2, pp. 197-216, 1985.
S . R. Kou, D. L. Elliott, and T. I. Tam, “Exponentialobservers for nonlinear
= p ) , we may set F = I,,,. Thus, if the triple (Ao,B , C) is controllable dynamic systems,” Inf. and Contr., vol. 29, no. 3, pp. 204-216, 1975.
and observable and the transfer function matrix formed by this triple J-I. Slotine, J. K. Hedrick, and E. A. Misawa, “On sliding observersfor nonlinear
systems,” in Proc. 1986 Amer. Contr. Conf.., Seattle, WA, 1986, pp. 17%-
G ( s )= C ( s I - A o ) - ’ B 1800.
B. L. Walcott and S . H. &k, “Observation of dynamical systems in the presence
of bounded nonlinearitiesluncerinti~,“ in Proc. 25lh Conf. Decision Contr.,
is strictly positive real. then there exist symmetric and positive definite Q
Athens,Greece, Dec. 1986.
such that assumption A 2 ’ is valid with F = ,Z (i.e., condition (28) K. Ogata, Modern Control Engineering. Englewocd Cliffs, NJ: Prentice-Hall,
becomes PB = C r ) . 1970.
170 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. AC-32. NO. 2, FEBRUARY 1987

[9] 0 . Hajek, "Discontinuous differential equations I. 11." J . Differenrial Eq., vol. is an output vector: A ; , B,, and C are real matrices of appropriate
32, no. 2. 1979. dimensions. We introduce the concepts of causal (c) observability and
[IO] J . J . Slotine. "Sliding controller design for nonlinear systems." Inr. J . Contr.. reconstructability of the system employing the notion of /-state of the
vol. 40. no. 2 . pp. 421434. 1984.
[II ] J . C. Willems. Stubflit), Theoryof Dynumicul Systents. Yew York: Wiley. system. The importance of these properties is caused by the fact that they
1970. imply the system's g-observability and g-reconstructability, which can be
LIZ] M.J. C o r k s and G. Leitmann, "Continuous state feedback guaranteeing uniform easily defined similarly to its g-obsenfability, for any definition of g-state
ultimate boundedness for uncertain dynamic systems." IEEE Trans. Auromat. or BCSofthe system.However,for completeness, we start with the
Conrr., vol. AC-26. no. 5. pp. 1139-1143. 1981.
[I31 A. S. Vostrikov. Dynamicul Syslerns Control. Novosibirsk Institute of Elec. system's /- and m-observability, which are also based on the /-state of the
Eng. (in Russian) Novosibirsk. 1979. system.
[ I d ] A. Steinberg and M. Corless, "Output feedback stabilization of uncertain
dynamical systems," IEEE Trans. Automat. Contr., vol. AC-30. no. I O , pp.
1025-1027. 1985. 11. OBSERVABILITY

For a description of the I-obsen~abilityof the system. we adopt the


appropriate definition given in [IO].
Definition I : The 2-DGM is /-observable iff any /-state of the system
x(k, t ) can be determined from the knowledge of the future outputs and
inputs of the system {y(i,j ) , u(i,j ) ; i 2 k a n d j 2 t } under assumption
x(k + i, t ) = x(k, t + i ) = 0 , i = 1, 2, .... rn
Observability and Reconstructability of 2-D Linear Then. based on the Cayley-Hamilton theorem, generalization [8], the
Digital Systems following theorem may be proved.
Theorem I: The 2-DGM is /-observable iff
J. E. KUREK

Abstmcr-New definitions of observable and reconstructable 2-D


linear digital systems are presented. Then, necessary and sufficient
conditions for observability and reconstructability of the system de-
scribed by the 2-D general state-space model are given.
where ~ ( +k 1, t + 1) = Ao@(k,t ) + A @ ( k + 1, t ) + Ay$(k, f + 1),
I. INTRODUCTION b(0, 0) = Z and @ ( k ,t ) = 0 for k < 0 or t < 0.
Proof of the theorem, similar to ones given in [ 5 ] . [lo], is omitted.
The observability problem for systems described by two-dimensional
(2-D) linear state-space digital models was considered in a number of Remark: Matrix b ( k , 1) defined as above is called the state-transition
papers, e.g., [I], [2], [ 5 ] . [7], [lo]. There were presented notions of local matrix of the system [SI. rn
( I ) . modal ( m ) , and global (g) observabilities of the system. The The notion of m-observability for 2-D system is a generalization of the
conditions of I- and m-observabilities for systems described by different approach to the observability problem for a I-D system based on relative
7-D models were also given in the papers. Unfortunately. there was no primeness of polynomial matrices. It is defined asa factor right
theorem concerned with g-observability, although it seemed to be the coprimeness of the system's operator matrices. Thus, we have, following
most important and the most interesting. [7] and [12].
The concept of g-observability is based on the notion of an infinitely Definition 2: The 2-DGM is rn-observable iff any decomposition of
dimensional g-state vector of the system X , [7]. [3]. In short, the system is the multinomial matrix
g-observable iff its g-state can be calculated using future outputs and
inputs of the system. However. more common 2-D state-space models are
based on the notion of a finitely dimensional /-state vector of the system
[3]. Hence. the g-state vector X;should be assumed arbitrary. It is rather
natural to define that X , is equal to the given boundary conditions set such that Q z ( w , z ) is square, implies det @(w, 2 ) # 0 for all w and z . .
(BCS). However, it suggests to define g-observability of the system as the Then. [9. Theorem 7.21 states the following condition of m-observabil-
observability (or distinguishability) of its BCS. Nevertheless. this notion ity .
is still not effective enough because one can use the BCS of a different Theorem 2: The 2-DGM is m-observable iff
shape for one equation of the 2-D model. e.g.. compare [j.ch. 1.11 and
PI.
Next, it should be noted that a similar important problem. namely the
reconstructability of the system described by the 2-D model. was never
for any generic point (wi,zi) of Vi where V, is an irreducible curve defined
considered before.
by irreducible multinomial u,(w, a ) and det (Iwz - A. - Alw - A g ) =
In the present note we consider the observability and reconstructability
of a system described by the 2-D state-space model (2-DGM) [8]. II,a,(w, z). rn
Remark: One can find more tests for checking m-observability of the
x ( k + 1, t + l)=Aox(k, t ) + A l x ( k + I , r ) + A , x ( k , t + l ) + B o u ( k , t ) system in [9].
Now we introduce the definition of causal observability of the 2-DGM.
+ B , u ( k + 1, t ) + B > U ( k . t + 1) (1) Definition 3: The 2-DGM is c-observable (or simple observable) iff
any /-state state ofthe system x ( k , t ) can be determined from the
y(k, t)=C.u(k, f ) knowledge of thefuture outputs and inputs of the system { y(i,j ) , u(i, j ) ;
i 2 kandj 2 t}.
where x E R" is an /-state vector, u E R"' is an input vector, andy E RP
Then the following theorem can be established.
Theorem 3: The 2-DGM is c-obsenrabl,e (or simple observable) iff
Manusript received May 22. 1985: revlsed Febmary 24, 1986 and September 25.
r 1
1986.
The author is with the Instytut Automatyki Przemysiowej. Politechnika Warszaaska.
Warsaw. Poland. on leave at the Lehrstuhl fiir Elektrische Steueruns und Regelung.
Riihr-Universitat. Bochum, Germany.
lEEE Log Number 861 1955. for all w and z.

0018-9286/87/0200-0170$01.00 0 1987 IEEE

S-ar putea să vă placă și