ada094~~~~ 41 txsaadmuvcolesttotlcm muiaon-t o lee tc ation unclassified tcsl … · 2014. 9....
TRANSCRIPT
ADA094~~~~ 41 TXSAADMUVCOLESTTOTLCM MUIAON-T F/ 1/NONLIN11 EAR N M O LEE TC ATION R LEOMMER ATION E TCOR F/6U 2/
NOVLINAS MOE IDTFCALTUR N RO ROPCRAT N E 46C-ORS
UNCLASSIFIED TCSL-80-01 AFSRTR81-0076 NL
EoommohhhhmhllomEEEmhhhEEEEEEEEohhhhhoh
"III.'.."mo
AFOSR -T 8 ' 0 7 6 '- .
TELECOMMUNICATION and CONTROL SYSTEMSLABORATORY
r--4
92 OTIC
ELECRICA ENGNEERNG ETENTr
___E 81 2 2 065vi 7H7 if7
IjI
NONLINEAR MODEL IDENTIFICATIONfFROM OPERATING RECORDS
Bruce K. ColburnAlvin R. Schatte
TCSL- REPQ 80-01
November 1980
I d' I f~f '/ - ;7f- (-oT
Assion For
ITS GRAVII 1 4 T 1 S
DTIC T'i3UnanouneedI Just trcatio a .
By._ttia~__ _Di~stributioni ..Availability Codes
)Avail and/or*Dist Speoal
AIR FORCE OFFICE OF SCIENTIFIC RESEARCH (AJSC)NOTICE OF T AYW:LT2"L TO DDC
This tech; ' . t L; beou reviewed and isappoi ! :,.. A'~,,)ne IAW A.FR 190-12 (7b).
', Distributioa iii i!zlmited.
A. D. BLOSErechnioal Information Officer
Approve for reei ft.distrbutio U11iimite
4 .
.. . *1
SECURITY CLASSIFICATION OF THIS PAGE (Wheon Does Entered)
READ INSTRUCTIONS( REPORT DOCUMENTATION PAGE BFR OPEIGFR=REPORT2. GOVT ACCESSION NO. 3. RECIPIENT'S CATALOG NUMBER
tAF A T -81-. 67 __ ________I
4. TITLE (and Subtie, -... - 5 TYPE OF REPORT & PERIOD COVERED
Nonlinear Model Identification /rom Operating Final, 6/79 - 12/79Records, ------
TCSL 8-i
Bjruce K. .tolburn1-7Alin . SchatteF46 79C09
9. PERFORMING ORGANIZATION NAME AND ADDRESS I0. PROGRAM ELEMENT. PROJECT. TASKTea AMRsen h ordai 1 AREA& WDRK NIT NUMBERS
C ollege Station -T--7764-3214 4;I I. CONTROLLING OFFICE NAME AND ADDRESSgF r0 _AX&
AFOSR(NL)IUSAF Boiling AFB, D.C. 20332 92
14. MONITORING AGENCY NAME A AODRESS(if different fromt Controlling Office) I5. SECURITY CLASS. (of this report)
UnclassifiedSCHEDULE
IA. DISTRIBUTION STATEMENT (of this Repot)
Approved for public release; distribution unlimited
IS. SUPPLEMENTARY NOTES
* 19. KEY WORDS (Continue on reverse side /I necessary and identify by block num.ber)
MRAS, Nonlinear Identification, Adaptive Observer, Human Operator
20. ABSTRA.~'oilnuo on reverse side it necessry end Idenify by block number) ietfctinwt-
Two basi c response error (RE) model-reference adaptive system
out dvace nowedg ofthenoninerites.Themethods are developed Msand shown to be applicable for systems with input and output measurementnoise. Design consideration, comparison of linear and nonlinear modelfits to a nonlinear plant are given. The applicability of such anapproach for the human operator model are discussed.
I
i TABLE OF CONTENTS
) page
1. I. NTRODUCTION 1
I
SIII. HYPERSTABLE NONLINEAR RESPONSE 32
m IV. DESIGN REQUIREMENTS 47
V. RESULTS 56
3 VI. SUMMARY AND CONCLUSIONS 75
m APPENDIX 80
REFERENCES 91
.ITI~I
IIII'_ _ _ _ _ _ _ _
I - -. -
III
NONLINEAR MODEL IDENTIFICATION
I FROM OPERATING RECORDS
If ABSTRACT
Practical considerations are given for the use of MRAS identification
techniques for the identification of linear and nonlinear parameters of
dynamic systems. The response error ("parallel MRAS") was used to identify
single and/or multivalued nonlinear parameters whose forms and "transfer
function equivalents" were unknown a priori. All parameter adaptation
occurred simultaneously. This paper develops the problem, discusses dif-
ficulties unique to the nonlinearity approach, and outlines possible solu-
~tions.
:1i
it
' I
NONLINEAR MODEL IDENTIFICATION
FROM OPERATING RECORDS
I I. INTRODUCTION
System identification has become a refined art and science in the
last twenty years [l-5], at least as regards linear system parameter iden-
tification. The area of nonlinear identification, however, has been some-
what neglected. As system operation requirements are tightened, efficien-
cy and energy concerns increase, and more complex systems (large-scale
systems) are contemplated, the need for detailed accurate knowledge of the
full system structure, including nonlinear and time-varying effects, is
increased. Some examples of dynamic systems where nonlinear effects are
significant include power plants, high-performance aircraft, human opera-
tor pilot-models, and transformers.
The objectives of the present work are to develop a model reference
adaptive system (MRAS) identifier for estimating parameters of nonlinear
pilot models to ascertain if information better than traditional (Klelnman
Optimal Control Model, crossover frequency response models) linear models
can result. The application of these results would be to identify human
dynamic models so
(1) stress lists/danger vehicle maneuvers can be investigated safelyby applying the results to the model, not the human
(2) as to use the information in the design of aircraft control sys-tems and for predicting closed-loop flying qualities
(3) as to determine effects of environmental inputs such as tempera-ture, vibration, acceleration, and psychological variables suchas motivation, training, and task difficulty
(4) to parameterize changing pilot characteristics during glide-slopeto flare-up; air to air tracking in unsteady maneuver, and hybrid(loose/tight) control situations
The concept is shown in Figure i-1.
. ...* _ ,,* 1. -2 2 -T 2 Z ... .... .. ..... ... .... .. .. .. .
2
The control of physical systems using modern control theory has become
: increasingly important with the advent of the space age and the energy cri-
sis. The need to operate systems at their peak efficiency and obtain the
utmost in performance from then requires that the control system have ac-
curate knowledge in the form system models of what the system is doing and
what it will do given a set of inputs to it. A special section of control
r theory known as system identification theory is responsible for developing
accurate models of physical systems using the input into the plant and the
output of the plant for use by a control engineer in designing control sys-
tems and to facilitate greater understanding of the physical system.
System identification can be divided into two major catagories: (1)
output tracking, and (2) parametric identification. Output tracking iden-
tification develops models for systems which for a given, fixed input will
approximate the output of the physical system or plant. Since the model is
good only for a particular input, output tracking is of little significance
for identification.
Parametric identification is of greater importance since it yields the
parameters of the plant. Thus, output tracking occurs for any type of in-
put, yielding a much more versatile and accurate model given the least
amount of a priori knowledge of the plant. Parametric system Identifica-
tion consists primarily of three classes: 1) linear, time-invariant, 2)
linear, time-varying, and 3) nonlinear, time-invariant identification.
The linear, time-invariant identification is best understood and most
widely used because such approaches are simple and can be applied easily.
However, in many physical systems, major nonlinearities occur (i.e. hyster-
esis in power systems, saturation in electron devices, etc.) for which
linear, time-invariant identifiers may not supply an accurate description.
Thus, nonlinear parametric system identification has become a subject of
Iincreasing importance as more accurate models become necessary.
Two key approaches exist for systems with one or more nonlinearities.
m These are 1) when the nonlinear form is known a priori [17] and 2)
l when the nonlinear form is unknown a priori [7]. The first approach is
predicated on knowledge of the nonlinear form in advance (e.g. quadratic
j curve, symmetrical saturation, etc.) or where the nonlinearity is one in
which the parameters to be identified enter linearly (e.g. y + ay + cy 2= u).The second approach is based on a series expansion concept (Taylor series,
power series, sine-cosine, etc.) for piecewise continuous curve fittings,
when no a priori knowledge of the nonlinearity is available. This latter
approach is of the greatest interest, especially in regards to its ability
to admit memory nonlinearities (i.e. hysteresis), Figure 1-2.
Several different forms and parameterizations exist for modelling non-
linearities in dynamic systems. These include 1) the Hammerstein model
(8], 2) Volteira series (9], 3) Uryson models [10], 4) memoryless non-
linearity peicewise series fit, 5) memory nonlinearity piecewise series
fit, and 6) known nonlinear model structure. A methodology using forms
4) and 5) which has been developed over the last several years is known as
Model Reference Adaptive System (MRAS) Identification [5]. MRAS identifi-
cation techniques can be broken down into two general catagories: 1) e-
quation error or "series parallel" and 2) response error or "parallel"
-- methods. Block diagrams of these methods are shown in Fig.I-3. The pri-
mary differences between these two methods are the way the error is formed
and the way stability of the composite system is assured. In the equation
error MRAS, the error is formed as a difference between weighted system
output states and input states. In the response error formulation, the
error is formed as the difference between the system output and the ad-
justable model output. The equation error system does not guarantee sta-
K E
I~~ Qo9UCE
LOA OMUE
DYNAMIC MODEL OF
PILOT
Figurer.1.
1 5
N NNLI NE A R ITY CH A RAC T ER IZ A TIO N
S SI NG LE- VA L UE
h 8 hZ h_ h)hkh hfO^x [hlp+ah2p X +a p p +X
I p
M U LTI -VAL UE D
F(w A11l + A12 p p-v +
h hf h(X )X
Figure 1-2. Example of usinga first order Taylor series expansion of f(x)
abutteW-ntx
C66
x
CDC
Il-0
U- 0
a.0
C) 0-a
~~(J%C3 LAC
LAJC-,C c.
It
Noo w
7
bility of the system while the response error system can guarantee a form
of stability. It is this form of stability, known as hyperstability, that
has motivated the recent interest in the response error or parallel MRAS
that this thesis is concerned with.
Hyperstability theory was developed by Popov [21] as an extension of
linear, asymptotic and absolute stability. Its use in linear MRAS iden-
tification assured these algorithms were stable in the sense of hypersta-
bility. Thus, the parametric identification of linear dynamic systems un-
der the hyperstability conditions is assured; a very strong statement which
non of the previous methods could guarantee. The extension of hyperstabil-
ity assurance for nonlinear systems, then, is of interest in order to as-
sure exact, nonlinear, parametric identification.
BACKGROUND
Dynamic system identification had its inception in statistics [1,3],
and numerous such ad-hoc identification algorithms exist. However, with
the resurgence of Lyapanov's stability theory, the field of MRAS was formed
[5]. Lyapanov designs produced several algorithms, but these were of lim-
ited value. In the early 1960's, though, V. M. Popov developed hypersta-
bility theory which brought about a radical change in MRAS identification.
Hyperstable MRAS identification algorithms were develooed by Hang [18],
Landau , and Johnson [12], amongst others, which accomplished linear,
time-invariant identification of dynamic systems. Early systems were con-Ustrained by the requirement of the existence of a strictly, positive realfunction (SPRF), but in 1978, Landau and Johnson removed this condition for
Shyperstability.
Developments of the response error MRAS technique into nonlinear sys-
tem parametric identification, however, have been scarce. In 1977, Tomi-
.-aka [20] applied hyperstability theory successfully to a very special
.1. -'
1 8
I class of nonlinearities. More recently, Colburn and Schatte [22] have
applied hyperstability with limited success to nonlinearitles of a more
I general type.
9
II. MRAS IDENTIFICATION THEORY
The fundamental concepts and theory underlying MRAS identification to
the present will be discussed in this section. The development of nonlin-
ear identification is a straightforward extension of these basic concepts.
Two major MRAS methods will be developed. This is followed by an expose
on hyperstability as it is applied to the MRAS identification process, and,
concluding, with the proof of the hyperstability of a particularly impor-
tant linear, MRAS identification technique.
MRAS IDENTIFICATION TECHNIQUES
There exis two primary MRAS identification techniques: (1) the equa-
tion error (EE) or series-parallel MRAS, and (2) the response error (RE)
or parallel MRAS. The EE method shown in Fig. II-1 was expanded on by
Lion [17], and Kudra and Narendra [6]. It was adapted for use with nonlin-
ear systems by Sprague and Kohr [7] and Sehitoglu and Klein [193. Consider
the nonlinear dynamic system:
anx(n) f (xCe) )x(n-) + .. + fi(xh) )x(i+l) +Fi(x i))
f(x(d) )x( i - l) + + fo(x(S))x u + gl(u(V))u ( 1) + ... +
gm (g)1(m) (2.1)
where an - constant,
(.)(i) dtdt I
u is the input or forcing function, x is the state, and fk' Fi, and g are
nonlinear functions of the indicated argument. Equation (2.l) contains
two types of nonlinear functions, f and F. Functions of the form fk and
gk are assumed to be single-valued and piecewise continuous with fk =,
such as shown in Fig. 11-2. The function, Fi, of which only one may occur
10
'00
4-4-
41
EUL
tm
1 1
I due to uniqueness problems with the constant term associated with the func-
tion is assumed to be piecewise continuous and possible multivalued. Also,
i Fi(0) 0 0 necessarily, see Figure 11-3.
Several assumptions on (2.1) must occur:
I 1) n > m, that is, the order n of x must be greater than the order m
of u
2) the system input, u, and output, x, must be measurable,
3) the input, u, is selectable and must be sufficiently frequencyrich [7] to insure parameter adaptation, and
4) the system may possess at most one multivalued nonlinear function, F.
For the identification process, x and u are filtered by state variable
filters (SVF) which develop "pseudo-states" which are approximations, x and
u, to the system's actual states, x and u. These "pseudo-states" are then
used to form the tracking error or "equation error":
x(n) + (-(e))X(n-1) A (h ) +(Xn)X + ... + -il (h) i +1)
n ^n -i ^ . .. + )
AAl)+ (d)A^(il) A + (s)A A
i(x) + .. + f0(x )x - u +
Ai(u(V u(A) + + g((g u(m) (2.2)
which is formed from (2.1):
(n) (n-l(xCe) )x(n-1) + (x(h) )x ( +
,+ f. 1 (x (d)x (i-l ) + ... + f (x )x - u - gl(u(V))u ( ) -
gm(u (g))u (m ) = 0 (2.3)
of the actual system and where ( ) denotes the approximation to the original
.4 function.
'U Exact identification of the above functions entails the estimation of
an infinite number of points which describe the curve, f vs. x. This is
not practical; therefore, an approximation is sought. Describing each of
.!'[.
I 12
Figure 11-2. Example of fkfunctions.
Figure 11-3. Example of F(') functions.
1 13
the functions fi, gi, and Fi in (2.2) by a Taylor series expansion in in-
tervals over the expected operating domain, piecewise continuous approxi-
I mations to the functions can be obtained. For the single-valued functions,
and g, these expansions take the form:(k) A
)x k l + ak2p(x h p) pfk(x~h ) +ak p " -' +-'
(x(h) r- + (k)
akrp x Xhp) +.. ]x
J where
k denotes the function being expanded,
r denotes the position in the series expansion
(.)r-1 indicates raising the argument to the r-l st power as opposed
to (.)(h) which indicates the hth derivative of (.) with respect totime
p is the interval index
h is the argument index, and
xhp is the expansion point in the pth interval.
For the multivalued function, Fi' the expansion appears as:
A i) A +
Fi (x ) A + A12p(x(i) - Ip) + ... +
A A~) -rlirp x iP + (2.4)
where the same notation as above applies. The nonlinear identification
process has now been reduced to the estimation of the parameters, akrp ,
Airp . and b where the b are the parameters associated with the ex-
pansion of the g (ur)U functions. The error equation now becomes
A n;(n rn- + ^ () ... + +E an + ,l,p +an-l,2,p (^ () e
an~l~r~p~x~e ]A - . .. (n -1) ^
a ep r-l + + [a1+l,lp
A (xh" hp) + + ai+,rp(x(h ) - xhp) + ... (+l)al+l ,2p ...... ~lrp x h
Il
1 14
Ail + A1i2p(x(i) - j)+ +. ip( x(i) x ) +Ca~~~ ~ ^-~l +p ..+a (()-1~ r-Ii)+- ) irp x l
X
+. I + +[al + ao( - +
(xp 02-xS
* Orpx sp Ix u - [bllp + blp ( u - vp +
... +b ir(u(V) -)r-1 A) A
rp Urp + "'"" [b mlp +SA Au(g) A A(g) - r-I (m)
)u + ...+b ( u ) +.Jm2p gp mrp gp ...
(2.6)
The EE method is derived from the minimization of the cost functional
j= 1 2 (2.7)
By taking the partial derivatives of (2.6) with respect to the parameters
y (i.e. akrp), the steepest descent algorithm is obtained:
A T =K AE a (2.8)
where K is a positive constant. This yields the parameter adaptation equa-
tions: ^(n)an =-G xna A A(n ' l)
an-l,lp = "Gn-l,lp E x
t ^( n)r-nGrp n-rp E x ep
ai+lrp -Gi+l)p E x
al~ir p " -Gi+irp E x l(x ) rp r-I
A AA Allp -G il p E
Ii 15
* Ai rp = Gi rp E(x''' - xi -itp
ai-lrp = -iIpE x I~(^X~ (d xdp r
ao = -G O E x'( '
a Orp GOrp ;()^() SP
blp HU E u
Ar 21 HlpEu (v) - Pr-
b =H Eu (u -iE u
it (m) '(g) r1bmrp = mrp Eu W gp
G..r = 0 r p
H.. = 0 t tp (2.9)
The other MRAS approach is known as parallel MRAS or response error
(RE), Figure 11-4. Consider the following linear dynamic system:
x(n) = nE ax (i)X + mr b~u Ci) (2.10)
P0O 1=0
and corresponding model:
xm E a ix m + z b u (.11=0 1-0
Define the tracking error, e, analogous to E in (2.2), as
einx -xm. (2.12)
1 16
.11
II
II
0
+ Ig
CL.
C In
a. L
S. La.
_ I:a,
tL
.eM.
17
This error is then processed thorugh a compensator, C, to yield the response
error, v,: n-I
(n) Mv en) + ?- ciei)(2.13)i=O
Hyperstability theory requires that the transfer functionn-liI1 + E
G(s) i=O (2.14)n-I1 - Z a s1
i=O
be a strictly positive real function (SPRF) for asymptotic hyperstability
of the error system (2.13). An SPRF is defined as:
Re{G(s)} > 0
for all (2.15)
Re{s} > 0.
Hyperstability also yields the following parameter adaptation equations:A (i)[On]
a i = ivx V i c [0,n-l]
(2.16)
it Mbi = Bi V i C € M-1
Because the derivation of this method is governed by exact stability theory
(i.e. hyperstability) parameter identification is guaranteed provided cer-
tain conditions (to be discussed in the next section) are met.DOI3 railLandau and Johnsonrremoved the SPRF requirement by insuring that it
always occurred. They did this by adapting the ci terms in (2.13) by:Y M.
c1 i v e i . (2.17)
In doing this, the transfer function, G(s), became unity and, therefore,
always an SPRF.
Tomiuka [29] applied hyperstability and RE identification to a class
1 i of nonlinear systems of the form:
"1
18
f Dh(m-l) I +". + f0 + fhOx(t) n (u(t)) (2.18)
D n + an-iDnl + ... + aID +h
for m < n and 0 where
nh(u(t)) = nl(u(t)) + n2(u(t)) + ... + n,(u(t)) (2.19)
are known static functions of u(t) and the constants f hi's and ai s must
be identified. Tomi-luka showed that using the parameter adaptation equa-
tions:A
a. = -kai v D x i [O,n-l], kai > 0I 1 a(2.20)
Afhj = kfhj v Di n j e [O,m], h E [1,Z], kfhj > 0
the transfer function (2.14) remained unchanged and that this algorithm
was, therefore, hyperstable.
RE and EE identification methods yield parameter adaptation algorithms
similar in structure. However, there are several fundamental differences
which exist between the two methods. Foremost is that the RE method is
governed by hyperstability theory, while the EE method is not assured to
be convergent. The RE method was developed through hyperstability so that
v - 0 as t - ®. The EE method uses gradient following techniques to min-imize E , with E in equation error analogous to v.
Landau shows that the EE identification yields biased parametric
estimates in the presence of measurement noise, while RE yields unbiased
results. In the nonlinear case, Colburn and Schatte ['(] demonstrated that
the RE state estimate (i.e. xm( )) is much more noise free than the EE
state approximations, (i.e. X (), Figure 11-5. This is because the EE
develops the state information by filtering the noisy input and output
while the RE develops state approximations through its own dynamics, giv-
ing a more noise free approximation.
'I .
-.
Ii 19
U L&J
%.-0
C
06
ci
1 20
I f()
Figure 11-6. EE Identification of saturation plus dead-zone.
'p Figure I1-7. EE Identification of hysteresis.
21
g In the presence of good (i.e. no noise) state information, though,
the EE will identify single-valued and multivalued memory type (e.g. hys-
teresis) nonlinearities very well, see Fig. II_6,7 .This is the chief advan-
5 tage of the EE method over the RE method because the RE method heretofore,
has not been applicable to nonlinear identification.
HYPERSTABILITY CONCEPTS
Hyperstability plays a key note in guaranteeing the stability of the
RE identification technique. It is a generalization of linear asymptotic
and absolute stability. Consider the system shown in Figure 11-8,
Figure 11-8 Basic Hyperstability System
block L is assumed to be composed of linear differential equations. If
block N is a linear feedback, then stability bounds may be derived:
1 < X < X2 (2.21)
where -w(t) = Xv(t). This is the linear asymptotic stability case. If
block N is nonlinear and obeys:
- (t) = 4(v(t)) (2.22)
p(v)v > 0 V v 0 0 (2.23)
(0) = 0 (2.24)
then the limits on block L and (v) can be obtained. This is the absolute
stability problem. Now suppose (2.23) is rewritten
1Ii I t0 -(lv(-) dT 0 V> t (2.25)0 (225
22
this, then, is the hyperstability problem. The problem can be stated as
N find the conditions on block L given a block N such that condition (2.25)
occurs.
To begin the study of hyperstability, suppose that block L may be de-
scribed as:
dx = Ax + bw (2.26)TT
v = cx + dw (2.27)
Several definitions must now be made:
Definition 2.1
Let the hyperstability integral, n, be defined as
n(0,t) = ft - wt!) v(2) dT (2.28)0
Definition 2.2 Minimal Stability
The system (2.26-2.28) is said to posses the property of minimal stabi-
lity if for any initial condition x(0) = x0 there exists a pair of continu-
ous functions x and w, defined for t > 0, and satisfying a) equations (2.26)
and (2.27), b) the restriction
n(0,t) f - (T) V(T) dT < 0 V t > 0, (2.29)
0
and c) the condition
lim x(t) = 0 (2.30)t-tDefinition 2.3 System Transfer Function
The transfer function of system (2.26)-(2.27) is defined as:
g(s) = C(sI-A)- b + d. (2.31)
In order to apply hyperstability to a system (2.26)-(2.28) it is nec-
essary that three conditions for this system (2.26)-(2.28) be satisfied.
I
f 23
f The first condition is that the system (2.26)-(2.28) be minimally stable
in the sense of Def. 2.2. This condition is acceptable since asymptotic
stability is of prime importance and, if the system (2.26)-(2.28) is not
minimally stable, then asymptotic stability in the presence of an inte-
gral restriction of the form (2.28) is impossible. The second condition
is that b # 0. This condition is obviously necessary since the solution
of (2.26)-(2.27) is not a function of u and the problem can be solved
4directly. Finally, the third restriction is that the transfer functionof system (2.26)-(2.27) given in (2.31) be not identically equal to zero.
This is necessary because if (2.31) was identically zero then n(O,t1 )
would be identically zero and (2.28) would be meaningless.
Given, that the above conditions are satisfied, hyperstability may
I now be defined.
Definition 2.4 Hyperstability
System (2.26)-(2.28) is said to be hyperstable if there exist posi-
tive constants (, a, y, and such that the following two properties are
satisfied:
1) Property Hs
For every t e [to,T 0] and every solution of the system (2.26)-
(2.28) in the interval [to,T 0 , if for every constant B0 > 0 for which
r(tot) < 2 V t E [to,T O] (2.32)>4 1 then_
X Hx(t)I I B0 + iIx(t0)I (2.33)
for all t c [to,T 0].
" 2) Property HA
j For every solution of the system (2.26)-(2.28) in the intervalLTTj
1 24
I "yLt~t) > -Yjjx(t 0) 1 - Y I x(t0 )11 SUP aI~X(T)HI (2.34)
Y t E[tOST 01.
Simply put, if
-a < ri(0,t) < 2(2.35)
then the system (2.26)-(2.28) is hyperstable, and if 62=0, then the sys-
I tern is asymptotically hyperstable [21].
The results of Popov show that if, for a linear system (2.26)-(2.28),
the transfer function from w(t) to v(t) given by (2.31) is a strictly pos-
itive real function (SPRF) and the inequality (2.34) holds then the system
(2.26)-(2.28) is asymptotically hyperstable. Willerns [26-27] showed that
the SPRF insures that
n(O,t) = ft- WW'~ V(T) dT < 0 YVt > 0. (2.36)0
RESPONSE ERROR IDENTIFICATION - APPLICATION OF HYPERSTABILITY
Consider the linear dynamic plant:
x(n) n- a (i) + m- bi u(i) (2.37)i=O i=O1
where
x - state variable
u - input or forcing function
ab. - constants
i) dt1
S and the associated model:
x(n) n- z1ax(i) rn+ M1b (i) (2.38)rn = ixr +z 0
1 25
I1 Define the tracking error as:
I e = x - xm (2.39)
and the error dynamics become:
I (n) n-I M n-Ie (n) a e + Z (ai - a +
1 (bi - b)uM i) (2.40)
i=O 1
Let
v(t) = e (n) - n ci e~') (2.41)i=O
thenn-i ^ n-i
v(t) = E (a. - ci)e + z (a1 - ai)Xm i) +i=O i=0
mn-i
(b - b)u(i) (2.42)i=O i
and let
* W(t) = v(t) (2.43)
The hyperstability system now is:
j= a e + E (ai - ai)xm +i=O i=O
rn-i)Em (b -b)u('i:0
(n n-l .b M
v(t) e - Ece2 i (2.44)
W(t) - v(t)
where e is defined in (2.39).
Theorem 2.1
I j If the system (2.44) is minimally stable and the input, u, into the
j 26
plant and the model satisfies a frequency richness criterion, then the
following parameter adaptation equations will yield the hyperstability of
system (2.44) and the parameters a1i ai and bi - bi as t ®:
a. = ),xm V i > 0 i e [0,n-l] (2.45)
iA Bi(i)vb = B. > 0 i C [0,m-l] (2.46)
= y" e(i)v y, >0 le [O,n-l] (2.47)
Proof
Since w(t) = v(t) the transfer function of system (2.44) is
G(s) = 1 (2.48)
which is an SPRF, thus
n(0,t) ft - W(T) v(T) dT < 0 V t > 0 (2.49)0
The other bound on the hyperstability integral must now be determined, that
is
n(0,t) = ft - W(T) v(T) dT > y2 (2.50)00
Substituting (2.42) into (2.50) one obtains
n-i 1 (i)n(0,t) = ft z (a. - ci)e vdT +
0 i=0
n-l (in-ft aM - at )vdT + - (b- vd (2.51)
0 i=0 0 i=0
It is sufficient to look at each of these terms and bound them individu-
ally. Thusn-ift E . (at - cY)e vdT > -y, (2.52)
S0 i=O Y
-tn-l )X(i)v _ft(a - a)x vdT > , and (2.53)
0 i=0
1I
I 27
ft m-1u i vd _,f Z - (bi b)u AT > (2.54)
0 i=O
I The integration and sumation may be interchanged in (2.52)-(2.54)
because they are independent of one another, and since the resultant is a
sum of integrals, it is sufficient that each of these be bounded, yielding:
I f - (ai - ci)e(i)vdr > 2 , i £ [0,n-1) (2.55)0 1 Yi
ft _ (ai - a )Xm(i)vdT > -y2 , E [0,n-l] (2.56)0 -i L,
ft - (b - )u(i) vdT > .y2 , i C [O,m-1] (2.57)0 1 )i
Using (2.55), let
fi(t) = -(ai - c.). (2.58)
Differentiating with respect to t yields
i(t) = ci (2.59)
Define the ci parameter adaptation as
= yi e (i) v , > 0 (2.60)
Substituting (2.60) into (2.59) and solving for fi(t) yields
fi(t) = f ft e ()vdT + fi(O) (2.61)0
where fi(0) = -(a1 - ci(0)). Substituting (2.61) into (2.55) yields
• I ft C fr e(i)vdT,]e(i)vdT + fi(0) ft e(i)vdr >
0 0 0
-Y (2.62)Yj
,I
=~
|228
I Using Leibnitz's rule:
fT f(T'r)dt' fT dT' + f(,T) dd'=T (2.63)
dt 0 0 d " T
I The first term of (2.62) may be integrated. Let
f(T',T) = e(i) (T')V(T') (2.64)
and use Leibnitz's rule yielding:
f 0 and (2.65)T
f(T,T) = e(i) (T)V(T). (2.66)
Thus, if
Fi = fT e(i)vdT' (2.67)0
then from Leibnitz's rule
dFi e(i)(T)V(T) (2.68)
and
dF i = e(i) (T)V(T)dT. (2.69)
This substitution, (2.67), puts (2.62) into the formyi ft Fi(T)dFi(T)d + f (0 ) ft e(i)vdT > _y2 (2.70)
0 0
'I which yields,
'L [ft e(i)vdT]2 + fi(O)fT e(i)vdT > -Y2 (2.71)2 0 0 Yi
Completing the square in (2.71) yields:
Y_ t (i) fi(0) 2 fi2(0) fi2 (0) = 2 . (2.72)
2L0 Y2 Yi
Thus, the boundedness of (2.55) has been proved with the given parameter
'I|
29
I adaptation (2.47). Using similar substitutions:
gi(t) = -(ai - ai ) and (2.73)
hi(t) = -(b i - bi) (2.74)
the boundedness of (2.56) and (2.57) can be proved using the parameter
adaptation (2.45) and (2.46), respectively. Since, all the integrals
(2.55)-(2.57) are bounded the boundedness of (2.51) is assured and is equal
to2 n-l f.2 (0) n- gi2(0) m- 1 h2(0)Y1 = E i + E + E (0 (2.75)
i=0 2yi i=0 2 i= 28i
or, substituting, for fi(0), gi(0), hi(0)
n- A - 2 nl2n-1 [ci(0) - a.] n-I [ai(0) - a.]Y0 = E + +
i=0 2yi izo 1
^ bl]2m-l [bi(0) - b 2(.6
i (2.76)i=0 2i
and the theorem is proved.
Theorem 2.1 was proved under the assumption of minimal stability of
the system (2.44). The following theorem gives the requirements on the
plant (2.37) in order for minimal stability to hold.
Theorem 2.2
If the plant (2.37) is asymptotically stable, then the error system
(2.44) is minimally stable.
Proof
Minimal stability requires that for any initial condition e(O) =e 0
a pair of functions w(t) and e(t) exists such that
1) n(0,t) < 0 and
2) lim e(t) = 0.
Define''I
30
n-I ( m- u'(t) Z Z (aj - ai ) + E (bi - b ) (2.76)
i=O I=0
then (2.40) becomes
(n) n-i (i)e = Z aie + u'(t) (2.77)
Let* n-I ( )
w(t) = (a. - c.)e (2.78)i=O
(this can be done by setting ai = ai, i E [O,n-1] and bi = bi, I E (0,m-l]
in (2.42)-(2.43), and is certainly an allowable w(t)). By using this W(t),
u'(t) = 0 and (2.77) becomese(n) n-I l( i )e = ni e (2.79)
1=0
which is asymptotically stable only if the plant is asymptotically stable
for any initial condition x(O) = xO. Since the w(t) was chosen from a set
for which
-y2 < n(Ot) < 0 (2.80)-Y
then the first condition of minimal stability is maintained and the theorem
is proved.
The frequency richness criterion in Theorem 2.1 was discussed by
Sprague and Kohr [7]. They showed that for parametric identification to
occur when the error was identically equal to zero the number of distinct
frequency components of the input, u, to the plant and model is given byn+m+l
N= 2 ' (2.81)
when the parameters to be identified were the ai i e [O,n-l] and bi,
i [O,m-l], and the plant was linear. They showed this mathematically
for the linear plant and no zero case and inferred it experimentally in
the general linear case. For the nonlinear case it was found to be atI
31
m least as large as (2.81). For the case in which there are also adaptive
ci terms the number of distinct frequencies was found to be equal to
N. 2n + m + 1 (2.82)N= 2I
using experimental results and empirical inferences.
The development of nonlinear RE identification algorithms is based on
the linear algorithm presented here. Its proof is very similar to the
I proof just presented.
I!)
i
32
III. HYPERSTABLE NONLINEAR RESPONSE ERROR IDENTIFICATION (HNREI)
g The development of linear MRAS identifiers and hyperstability prin-
ciples along with the nonlinear concepts of the EE method lead to the de-
velopment of nonlinear RE identification techniques. Heretofore, the ex-
tension to nonlinear identification using the RE method has been thwarted
by the requirement of the SPRF which is not defined for general nonlinear
systems. However, with Landau's [11] and Johnson's [12] papers elimina-
ting the SPRF difficulty through problem reformulation, an approach for
HNREI algorithms now exists. In this chapter two different nonlinear
identifier formulations will be presented and two nonlinear identification
algorithm proved to be hyperstable are given.
Nonlinear Formulations
For the development of HNREI algorithms, two nonlinear formulations
were considered: 1) piecewise linear expansions in intervals of the
state space and 2) power series expansions over the entire state space.
Both of these have distinct advantages and disadvantages. The piecewise
linear expansion, Figure 111-1, allows the nonlinear function to be divided
into a series of intervals in the state space and a linear approximation to
the nonlinearity formed. This method has a serious disadvantage, though,
when perfect state information is lacking, Figure 111-2. It is possible
that the identification algorithm could be adapting a parameter in the wrong
interval yielding erroneous results. Thus, the degree of accuracy obtain-
able by the piecewise linear formulation decreases with the loss of per-
fect state information.
The power series expansion technique, Figure 111-3, uses a continuous
approach and, hence, does not suffer from the intervalization problem as-
sociated with the piecewise linear method. The power series method attempts
to approximate the actual nonlinear function as a finite power series:I''
IFI
I ok + al k (x(h) x kh)
X(h)
7kh
Figure III-1. Piecewise Linear Expansion
noisy state
stateintervals
time
Figure 111-2. Interval Selection Fromi States.
I 34
I~ f(x)
do..
.00,
Fiur 11-. PwrSrisEpnin
INf(x) Z N aix'. (3.1)
i= l *
The principal disadvantage of this method is that for some nonlinearities
which can be approximated with a small number of intervals using the piece-
wise linear expansion may require many terms using a power series expansion.
Noisy state measurements also present problems because of the possible
high degree of the polynomial and noise biasing.
Nonlinear Identification Algorithms
The nonlinear expansion techniques just discussed were used to de-
velop HNREI. The piecewise linear HNREI required a special interval con-
dition to insure hyperstability while the power series expansion technique
required no special conditions.
Piecewise Linear HNREI
Consider the following nonlinear plant:
x(n = nE1 fi(x(i)) + g(x (Z)) + Z hi(u(i)) (3.2)i=O i=O
with a corresponding model:
x(n) n-I 0i)) + ̂ (xm(o) + ME, NO (3.3)
i m 9z m i=O
where (.(i) , and Z 6 [O,n-l]. Leti dti
fi(x (i )) = ailX(i)(34
.igt(x ( ) = a a ( (91) - p) (3.5))) ato p U ~p(
hi(u(i)) = bilpU(i) (3.6)
which have corresponding model nonlinearities
I ! I
36
fi(xm () = ilpX(i) (3.7)
1t~ m (Z il= tpm+aZP( Z z 38(x ( )() - m2 (3.8)g (x )= a 0p + a~lp xm - mp
t h~) ()lp (1(.g
where p denotes the pth interval of the state space x i) or u( ). Thus,
the plant and model may be rewritten, respectively, as:
(n) = n-I ailX(i) + a + a i(X(Z) ) +
m-I bilpZIPi)
E il (3.10)i=O
=m(n ailp mi) + a Op + a .lp ( Z xi z
ii=0
i=0ini (i)
where the parameters bilp' ailp' atop, and a~1p must be identified in each
interval p. Define the tracking error as
e = x - xm (3.12)
The error dynamics which are to be made hyperstable are:
n-i-A
) e(n) n-I p1 e (i) -a _ .i~ _m~ .+2 0 ^a
~i=0
A W,, nE )x (i)+ (a - aP )(Xm -XmZp) = (aiip -a )
i=0
+ - (b b )u(i) (3.13)
i=0 P UP
Let/ . e( n ) n -Iv(t) = Z c( le (3.14)
I i lp• I
I.
. 9
37
thenn-1 A p()
v(t) E (ai - )e -a p(X -X ) +i=O ip lp ZIP XP mp
atop " op + (a ip a zlp)(Xm C Xmzp +
mE (a - a )x (i) + m (bilp b ip)Ui (3.15)
i=0 lp lP m i=0 i
Define
S= v(t) (3.16)
The hyperstability system is given by (3.13), (3.14) and (3.16).
Theorem 3.1
If the system (3.13)-(3.16) is minimally stable, the input, u, into
the plant and the model is sufficiently frequency rich, and the condition
X =p XMIp are satisfied, then the following parameter adaptation equations
guarantee the hyperstability of system (3.13)-(3.16) and the parametric
identification: a t0p - a top, a ip and b lpo b as t
Aa Lip a zip v a mp > 0 (3.17)
" V(pm( p
a Ip a IP Vx - Xm~p)' aLIp>U (3.18)
a lp = lp V m i [0,n-l], i L UP > 0 (3.19)
bA (i) ,b = i UP vu U > 0, i E [0,m-l] (3.20)
Clip = yx ve i) i > 0, i e [0,n-lJ (3.21)
Proof
Since w(t) = v(t),
/, , (0,t) - -t.W(T)V(T)d < 0 (3.22)
0
"I
- I I • I " " 'mm- " n . . . . . .. . .. I l m m i l
38
and, thus, one limit on n(0,t) is established. To establish hyperstability
the other limit must be satisfied:
n(0,t) = ft - W(T)V(T)dT > -Y (3.23)0Y(
Substituting (3.15) into (3.23) for w(T) using (3.16) yields
n(O,t) ft n-I l p e +i
E z - (all - ci )e vdT + ft0 i=0 P 0
(-X )vd + ft (a O- a Lp)vd + ft
. -a a op)vd,]+ Z - (rn- i)u vdT > -d 0 (3.24)
i It is sufficient to bound each of the terms in (3.24) in order to insure
the boundedness of (3.24) and by exchanging the integration and summation
i of terms 1, 5, and 6 and using the same sufficiency argument (3.24) yields:
f t - ci )e(i)vdT > Y2 i 6 [O,n-l] (3.25)
0 (ail cp Yilp
l ft - (a - a 0op)vdT > y2 (3.26)
0 Lop
(a,1 - ip)(Xm(4) - mp)VdT > C (3.27)
I A Ift - (all - a p)Xm ivdT >-Y i C [0,n-l], is2. (3.28)
S0ip p -- P
tMf- (b 11p ~b 1 p)u 'vdT > - a i I [0,m-i] (3.29]
Using (3.25), let
I Afp (t) -(ailp cilp). (3.30)
I Differentiating (3.30) with respect to time yields:
I fuip(t) = cilp (3.31)
1. ..
39
Define the cil p parameter adaptation as:
Aleiv, 0 (3.32)Cilp= Yi p 0
Substituting (3.32) into (3.31) and solving for filp(t) yields
f (t ~ t f~ i vdt + p()(3.33)Sfilp ( t ) Yilp 0e filp(0)
0
where f ilp(O) =-(ailp -Cilp(O)). (3.34)
Subs tituting (3.33) into (3.25) yields
ft [f T (i)vd']e(i)vd + f11l 1t e(i)vd (3.35)
0 0 l 0
By using Leibnitz's rule and completing the square, (3.35) yields:
1. f 2 (0)il[ft e(i)vdr + >( - fip 0 = 2 (3.36)2 1l 0- - 2yilp Yilp
Thus, on the first integral has been obtained using the parameter adapta-
tion (3.21).
Similarly, using (3.26), let
g9op(t) = -(ato p - atop). (3.37)
Following a development similar to (3.30)-(3.36), (3.26) can be bounded
using the parameter adaptation (3.17) obtaining:
g g(0) 2 . (0)l g2 (0)njZO 2 Eft vd £ ODj 2 tl tp0 p r atOpp
(3.38)
The other integrals (3.27)-(3.29) can be bounded using (3.8)-(3.20), re-
spectively, and obtaining, respectively,
IiI
ii 40(0
!L-'R [ft (xm ('1) )vdT + 2..1DLID -
p0 1 az"p 1 p
> - = _ Y2 (3.39)- i1p i p
Ili [ft Xm(i)vdT + g~ip(Oi ]2 >
+il(0) 2.2al
2a _ a (3.40)ip ip
8i_ [ft u(i)vd h2il(0) ]2 h211p(O)2 ip 2 lip
h2ilpO .Y 2 (3.41)" 2iip = ilp
where
hilp(t) = -(bil p - bilp). (3.42)
Thus, all of the integrals (3.25)-(3.29) have been bounded, and,
therefore, the integral (3.24) is bounded by -y' where
n-if1 n-Ig 2 i(0) m()___ 2iZ + li +O 0 iO ilp 10 2ailp i=0 20ilp
g2o(0) g2 (0)t + 11 (3.43)
b or, substituting for filp(0) gilp(O), hilp(O)
n-I (a. (O)-a n-I (a -a ' 1Y.. 2yiO i=O 2aiilp 0 =0
(b lp'-blIp + (aOD-alop) + (a.p-a,,I) 2 (3.44)2 il p 2a top 2aLIP
Thus,
1 41
I 0 < n(ot) < 0 (3.45)
Iand hyperstability is proved.The conditions for minimal stability of the error system (3.13)-(3.16)
Iare given in the following theorem.Theorem 3.2
If the plant (3.10) is asymptotically stable, then the error system
(3.13)-(3.16) is minimally stable.
I Proof
The proof of Theorem 3.2 parallels the proof of Theorem 2.2. Let
n-lw(t) = E (a -c )e(i )+ a -a i -7ip ) (3.46)
i=0 ilp ilp top Lip zP
By using this input, the error system becomes:
e (n) = n-l a1 1 pe(i) + a + a ZPX p.
(Xm Xmip)] (3.47)
by setting
atop -0 (3.48)
at, = a11p i c [0,n-l]
bilp = bil p i E [O,m-l]
and uM() 0 1 C El'm-l)
Thsxm()(0) = 0 1 e [O,n-lJ.
Thus, (3.47) becomes
1 x(n) = Z 1 apx(i) + azop + a.l p(x(t) 'x (3.49)
• , Ijot
S 42
I which must be asymptotically stable in order to insure the minimal stabil-
ity of (3.13)-(3.16). The final condition of Theorem 3.1,
X = xmip (3.50)
requires substantive a priori knowledge of the plant. This type of know-
j ledge is not readily available and condition (3.50) limits the piecewise
linear HNREI.
I Power Series HNREI
Consider the following nonlinear plant:
n = n E j= aix (1 iJ] + - (i)j -- . )J] (3.51)i=0 j1l 1=0 O
1 where k, Zi is a finite integer and (.)(l)j indicates d L-m raised to the1 dt
jth power. This plant has a corresponding model:
I (n) = n-I k i -I ziXm (n [ E ajmij + ME [ Z b )J( ] (3.52)
i=0 j=o i_10 j=l
Define the tracking error as:
1 • - x - xm (3.53)
I and the error dynamics become
n-l ki n-I kia e (n) = n E a i j (x( i ) J ' x m(i ) ) ] + niO E (aij-aij)xm()]
* 1=0 j=l i=0 Jul ij1 mI
1 b (b- )u(i)J (3.54)1-0 j l
I Let the response error be defined,
n-I kv(t) z e(n) - E [ cij(x(i)i-x (i)j (3.55)
iJ=O Jul
I mand define
W(t) v(t) (3.56)
II 43
I Thus, the error system to be made hyperstable is (3.54)-(3.56).
i Theorem 3.3
If the system (3.54)-(3.56) is minimally stable, and the input, u,
Iinto the plant (3.51) and model (3.52) is sufficiently frequency rich,
then the following parameter adaptation equations insure the hyperstability
I of system (3.54)-(3.56) and the parameter convergence
a.. - a1j, i .-" b.. as t- :I A bM3
ct. iJx v le[0,n-1], jcel,ki), a1j)>O (3.57)
b. z 81.U~)V le[0m-l], jE:[l,2. 1) 8 >0 (3.58)
=i ''~x _xM (I)Jv ic[0,n-l1, je(l,k1], Yij>0 (3.59)
Proof
Since (3.56) holds the upper limit of ri(0,t) defined as
n(0,t) - ft W(Tr) V(T) dT < 0. (3.60)0
Thus, the other limit for (3.60) must be obtained to prove hypersLabillty.
Substituting (3.55) and (3.56) into (3.60) yields:
nE~)-f [ E - (a jcjj)x _x )d0 i=O J-1
k
'0=0- i li )xm ('JvT+ ft
.2(b 1i'-bij)u(1 )Jvd-r 0 - (3.61)
By using a sufficiency argument similar to that used in Theorem 2.1 and
I Theorem 3.1, (3.61) can be written:
1 44-t (ij ()j) d >_2
f (a. (-c))vdt > y2 iE[O,n-l],
Saij -ij m Yij
I je[l,k i] (3.62)^ - i >"y 1
t - (ai -aij)x m vdT > CL iE[O,n-l], jE[lk i] (3.63)
101J - 8ij'
(bi -bi )u()JvdT > -Y 2 ic[Om-l], jecl,k iJ (3.64)
Also, following an argument similar to Theorem 2.1 and Theorem 3.1, and
defining
f. (t) = -(a ij-c.ij) (3.65)
gij(t) = -(aij-aij) (3.66)
hij(t) = -(bij-b.i) (3.67)
the inequalities (3.62)-(3.64) are found to be, respectively,
2 ft(x(i)JXm ()j)vd + fij(O) 2 f2 i(O)2 0 Yij 2yif2 i .(0)
=-2
=2Y y2 (3.68)
2_ Yl j yl j 2
jLioXm JvdT +-a.. . 2al > 2ai..02 13j --
"Y 2 (3.69)
[ft () h(0)2 h2i(0) h21 (0)vd[ + u(1Jvd0 +2] U > . j =
2 0 d i+ 2 8ij - 28ij
"Y 2 (3.70)
which bounds (3.61) by -y0 where11'. . ...I. . . .. . . I Il II i . . l I I I IlI l .. .
I, 45
0)k2 (0) 2n-li Cc. -a' a a.) rn-i* j -2 + ] +
0 i=0 jzl 3 1 2i0
| i (bi (O )-b .)2 ]
1[ 2 (3.71)
j=l 2aii
where the initial conditions
fij(0) = c i(0) - aij (3.72)
gij(0) = aij(0) - aij (3.73)
and h. (0) = bi(0) - b.. (3.74)13 ii 1
have been substituted in (3.68)-(3.70) to obtain (3.71), and hyperstability
is proved.
The conditions on the plant (3.51) to insure minimal stability of the
error system (3.54)-(3.56) are obtained in the following theorem:
Theorem 3.4
If the plant (3.51) is asymptotically stable, then the error system
(3.54)-(3.56) is minimally stable.
Proof
The proof of this theorem parallels that of Theorems 2.2 and 3.2 and
consists of finding a w(t) such that
1) n(O,t) < 0 and
2) llm e(t) = 0
for any initial condition e(0) = e0. Letn-l ki
S w(t) = [ z (ajj-cij)(x ''' -X j)) (375)iO J=l I
by letting
aij , aij (3.76)
bij = bij (3.77)
1 46
I The error equation (3.54) then becomes:
k.e(n) . n-i i ij (1)
a. (x x ] (3.78)i=0 j=l 1
i
which has lim e(t) = 0 only if the plant (3.51) is asymptotically stable.t-cn
Thus, minimal stability is proved.
Thus, two HNREI algorithms have been developed which yield parameter
adaptation. This parameter convergence is guaranteed if the plant is asymp-
U totically stable and the input, u, to the plant and model is sufficiently
i frequency rich. The next section discusses methods of implementing these
algorithms and design criteria associated with them.
l
l
l
l
4 'SI
i I
1447jIV. DESIGN REQUIREMENTS
In the design of nonlinear, parametric identifiers, several design
considerations must be made. These considerations can be roughly grouped
into two catagories: 1) what type of identifier is to be chosen, and 2)
specific nonlinearity considerations. The answer to the first depends
on the a priori knowledge of the nonlinear system and the maount of noise
in the system. If the functional form of the nonlinearities is known,
then a linear, time-invariant identifier might be advised. Similarly, if
the nonlinearity forms are unknown and only small (<1%) amounts of noise
exist, then the series parallel (equation error) MRAS approach might be
I best. The criteria for choosing between identifiers is largely a func-
tion of experience. However, with unknown structure nonlinearities in
the presence of noise, parallel MRAS yields best results.
I e Fig. IV-I shows a general block diagram of the nonlinear response
error parametric identification algorithm. The algorithm may be broken
I into two main loops. The inner loop consists of state variable filters
(SVF) adjustable model, dynamic compensator (N(s)), and the parameter
I adaptation. This inner loop presents several considerations for the de-
£ signer:
1 1) SVF design
i 2) Parameter gains93) PRF and the dynamic compensator design
Ii The SVF's function is to provide "pseudo-derivatives" for use in develop-d
ing the response error. If p d then L(p) denotes the SVF and (1) be-
comes
L• l(pnx P) + ... + L[F i(P ix P) + ... + L[f°O(p k xp)x p ]
U L[go(p u p)u p] + ... + L[gm(pWu p)u p] (4.1)
!I
44
( -i
LU .
2: LU_
CD)
tnr
k44
49
where xp and up are the plant output and input respectively. Sprague and
Kohr [7] have shown that if L(.) is a transport operator then L(.) commutes
4with the nonlinearity to giveSL(pnx p) + ... + [Fi L(pJxp) + k+ fo[L(pkxp)] Xp
g0[L(pu p)]up + ... + gm[L(pwu p)]up (4.2)
allowing the SVF to be moved from the output of the model to its input.
This gives compatability between the model states and the corresponding
plant "pseudo-states".
Parameter gains should be chosen in such a manner as to give rapid
convergence and decrease asymptotically to zero to prevent noise bias.
Several algorithms exist [11,12] for the decreasing adaptive gains'
The PRF requirement was addressed earlier. If a priori knowledge of
the range of plant parameters exists then a fixed ci compensator can be
developed, saving much computation time and implementation hardware. How-
ever, in general, the nonlinear adaptive compensator must be used to ob-
tain the PRF requirement.
The outer loop of the nonlinear identification algorithm consists of
the pattern recognition, model structure, and nonlinear interval blocks.
Important design consideration occurring in the outer loop are:
1) Pattern recognition block
2) Model structure blocka) Model order (numerator, denominator degree)b) Nonlinearity location
c) State information
3) Nonlinearity interval blocka) Interval sizingb) Number of intervals
4) Identification time frame
I1
50
The pattern recognition block must determine if a "possible" nonlin-
earity has been identified. This block decides when a nonlinearity des-
1cribed by a set of points is Class C , i > o, to qualify as an allowable
4 nonlinearity, Fig. IV-2. Define:
AIfl = P2 - Pl
A1 f2 = P3 " P2 (4.3)
Alfn-l= Pm - Pn-l
the "first difference" fo the set of points.
Then A2fi A l AI
and AfA fi+l f Ai-I f (4.5)and Aif = Ai-lj+l f.
represent higher differences of the set of points. Now the performance
criteria: iN A -Z.
< E e is greater than zero (4.6)Nj=l TZ1
i
defines the class of functions C i- where only the A ftj that differ in
sign from the previous, A i fgj-l are used to define the class. N is the
number of sign changes, and T.i is the width of the interval between the
points.
The results of the pattern recognition block feed into the model
structure block. In this block almost all of the decisions concerning the
model structiire and nonlinearity placement are made. This block is nearly
always "filled" by the designer himself.
The first piece of information necessary from the model structure
block is the model order and preliminary structure. Several model order
algorithms exist [13,14,15] to provide possible model structures and in-
formation on the number of poles and zeros for the model and any time de-
1 51
P ATT E RN R E COG NI TIO N
P4
N AI)PEJ
J=L TJ
FO0R CL A SS C
Figure IV-2.
52
lays.
The placing of any nonlinearities in the afore found model structure
is largely a matter of trial and error or intuition as to what states
4will be affected by the nonlinearity. By minimizing J in (4.6) of the
nonlinear pattern recognition block with respect to i an E can be picked
which fixes the class of functions allowable for that particular model
state. If the expected plant nonlinearity was not of this class then
another model state might be chosen in which to locate a model nonlinea-
rity.
Since both plant and model state information and used to identify
possible nonlinearities, it is important that the proper state information
is obtained to get the correct answer. A decision must be made whether to
use the noisy plant states, themodel states, or a combination of the two.
The problem is illustrated in Fig. (IV-3). As can be seen, at t1 the ac-
tual plant state (unmeasurable in general) is in interval 2 while the model
state is in interval 3 and the noisy plant state is in interval 4 of the
intervalized state space. This problem can be posed as an optimal com-
munications problem where a decision function:
S (X) + x) (4.7)
is defined and the functions f and q are chosen to minimize the probabil-
ity of selecting the incorrect interval. Q, then, is the independent var-
iable used to determine the interval number of the model's state space.
The nonlinearity interval block decides how the model's state spaces
will be broken up in order to identify the nonlinearities (if any) exis-
ting in the corresponding plant's state spaces. This problem can be bro-
ken into two interrelated parts: 1) interval size and 2) number of in-
tervals. Statistical measurements can be made for a particular input to
determine how the Q's of (4.7) range. The interval size can then be cho-
I!
53
I LU
CD,
3fnVA 31V-
54
I sen to allow for uniform (if any) convergence of the parameters.
The number of intervals presents other problems. As the number of
intervals increase, the amount of state space in these intervals decrease.
As teh SNR of the independent variables decreases, the probability of an
internal misselect increases. Let
Yu==o output with noise present and no input
Yuso = output with noise present and nonzero input drivingfunction
NZ YuMo(it)2
SNR = N- (4.8)
SYu=o(it)2
i=l
As the SNR decreases, the number of intervals the nonlinearity can be
broken into and still get an "allowable" (in the sense of (4.6)) fit de-
creases, (Number of Intervals) c(SNR (independent variable), where the pro-
portionality constant is designer-selected. As offline test for the proper
number of intervals is given using the Theil criterion [16].
Nl 2
K i(4.9)N N p
As shown in Fig. (IV-4), for a particular example, an optimum number of
intervals was found.
The identification time frame is also designer-selected. In general,
the longer the time for identification, the more accurate the final answer.
However, in practice the amount of time must be weighed against the speed
of parameter adaptation and measurement noise levels in order to give an
optimum adaptation time.i |
I
* 55
'-I
CD,
C) L-%
1 56
V. RESULTS
The response error MRAS approach was tested on numerous models,
I since no human pilot operating records were available from FDL at Wright-
Patterson AFB, Ohio. The examples illustrate the good parameter track-
I ing possible. It should be emphasized that since no optimal control
pilot model formulation was explicitly a part of this study, but instead
only the feasibility of MRAS identification for plants with correlated
noise, results and their interpretation should be limited to the scope
of MRAS nonlinear function and linear parameter estimation.
Example 1 (No Noise)
Plant: xp + f(x)x + x = u
where f(x)x is the saturation function shown in Figure V-i
Model: xm + ailqxm + xm a bou
bo = constant
allq = supposed intervalized nonlinearity
Using the interval select nonlinear MRAS approach, with ci's constant.
With SVF poles at -20 and -30, the results are shown in Figure V-i,
with noisy measurements
Xavail X +
E{n} - 0
a a -= .05xp
Example 2 (Noise)
Plant:
Spx -f(Xp) - xp + uP p
y(output) x p + nx'7
I!
'~1
57
IL
ILa
N
LLLLI
! --
!58
The available (measured) u is u' - u + nu
u Z Aisin witIi=l
I ( {.2, .4, .6, .8, 1, 1.2, 1.4, 1.6}
First, a linear model will be fit to the system and then a nonlinear
I plant, to compare the improvement in tracking accuracy of the nonlinear
version over that of the linear version. The linear version is
Ym blm(S) =u-u-(ss2 + a21s + a,,
and the nonlinear version
xm = u fl(xm) f2(Xm)
Ym = f3(Xm) + f4 (Xm)
Each case used ci values varying. The nonlinear case used 5 intervals
for each nonlinearity.
The results for the linear case are shown in Figures V-2 through
V-13. The nonlinear case is shown in Figures V-14 through V-16.
Results show that the linear model has a poor fit to the (actual)
nonlinear plant, because of the forced fit, or lack of degrees of freedom,
of the linear model. The nonlinear results, on the other hand, show
) good model-plant correlation, as can be seen in Figure V-16(b), where a
phase plane of yp vs ym is plotted for the test input. If the model were
a perfect replication, there would be a single line at 450; only small
variations about this line occur.
A complete computer program listing is given in the Appendix of the
identifier system for practical nonlinear identification by the interval
approach. A flowchart is shown in Figure V-17.
/
.I-... .....,I
.1 59
.1w
00
00
0
00
0
La w
o 0
0 1
w c0 -0
-1 c-
La Ir-jW-0 -MO'
o c0cou mm
0 11 0
oz :0 40 0> Nw C*I w
z o o ivw *
o n<U III...4 *00 w*La 0 ja <*WW J
co a0 Z
0 0 *0
z A - - - -0- -
I.* 60
II
N, 0
I- %.
-
/
II
& 4 4 I I - 3 I3 I3 S S * '
1 61
!
e
L 2i ;-
i *-
I ~, p p p
I 62
II
I7
_ _ _ _ _ _ 7
63
I4
pa
f fA
64
UNITS
SL -
0>
k 'I._4
- I
t
,_ _...
) -4
"I
65
0
* (I)
siiwn
fr
191
I~~Ca a - - 6
C 0(n0
simn
4
,. N
- w - w w - 0 S - - - - - - - - - -
67
It
00C.)I
0~: 0
_________________________ Cu
C.,
wa.w 'S
00
- 0
I
)0~6.
'A
0U) U) -
0 0 a4 4w Cal w w0 0 00 0 0 0o 0 0 0§ 0 0 0
'0 V Cu 0
shill
II
j~ * . . . a a a a a 4~ . a ~ ~ ) ) ) )-A
68
Iu
IcI 0 L
C!a C; C C; a
sIw
69
C;C
I-I
I 70
2 30 3
11 M
00
0
0
0
o 0
00
- 14 0(0 0 0
o -Jo 0 wo
0 0.J j 0I - - . fn
- 0 0 oo- 0
00*
0 0 5-5-0 55080 00 0 o
1 1 1 1 + 1 1 1 1 111-
0 0 0 00 0
*0-8 8 00 0 *it c .
o0 0-00000 05 -0.951O 01 1 11 14 W" " 1 1 1??0 -1-cOW nono 0 ow M00 000 00 0 0 I 1 l- W 08 81
b~~~~~~ a- 00 0 0 0 0 00 0 0 C 10 - 00 0 --1+ --- 11 101 1 -+++
... .. J 0-.- 0so 000000 _000 -0, I * T1. 46 t 4*44** MINION *
- 1-81108.110
N ~ ~ .0 , n n- 6-0- ....r4 W1 inW inC 1-- o C J-
w 0~C1II j I.-44 0 00 -
ll l + 11111 I4 4 4 44 4 10
-OZ A 00 * OZ- aaaaa11111*4 1- 4* 1--0 - *'- -0 0*0
z 0 :;~-.4 0: .
I.-
72
IO
C
I In
C
MF I-
73
CL
C;
C0 1
f tmn) r0
'Np
I NOML.weft 1"ir-N ro) 74
Il
Se *ILwevvol s
UjVW-'/o* \Y ju
UJT~LLC.
Par ~ *t~ i &w&E)e
FE 4i*
'1v
L--j
75
IVI. SUMMARY AND CONCLUSIONS
The response error MRAS identification can result in meaningful
nonlinear identification if set correctly. Chief problems include ci
varying by interval, noise biasing effects. This is shown in Figure VI-l
and 2. If one uses fixed ci terms, the nonlinear intervalized response
acts like time varying ai(t) terms for the plant. Even if one uses varying
ci terms, ci , there is the sharp time-varying tracking problem as different
intervals are entered (Figure VI-2). As shown, in Figure VI-3, there must
be vector arrays of ci terms, the vector entries being for different in-
tervals. This means that if there are two ci terms, c1 and c2 , whereAT A lI Cl2
T, . ., c ] where there are p intervals to the nonlin-
earity.
The effect of the number of intervals in the presence of noise has
been shown to be a parabola, Figure VI-4. Too few intervals yields an
"almost linear" fit, and too many, given noise, yields nonsense due to the
interval search problem discussed previously.
To perform total human operator identification, a complete Optimal
Control Model (0CM) computer package, plus operator remnant construction,
would be needed. Such work was beyond the scope of this project, but
could be extended from the present results.
The methods developed here do provide for good nonlinear model iden-
tification, if the noise levels are reasonable. Work is underway to mini-
mize the effects of the large correlated noise levels.
elA
t.
76
PRF PROBLEM
CI F I XED
IoIK
L
L-0.
F
S I
77
C1 AD JU ST I NG
AA
HIA
MO
A L
78
SOLUTION
x
FXo AA
A1 C1
* S
2 XO AA2
1
00
b N0XA
79
I.- c=
U- J4u
Ij2F
&
1 80
I
I
)
.1
I~~6
* 1 81
-0 -0 10.04 A
2! 1I ca (W (V000;g
000 0008
U. .L U. U. U.A
zz Zz z
zzz zz~zfz000 000o
ZZ22zzZzz
Ia~~0 4- 4a14 .
- - -z - z- - - - - -
000 .0 030Q
In M00I
-- c C -- i 44 44
-- ~-- J44 .
-------- f 44 4>4-- -- -- -- -- 44 44 '04i
-- - 4 04 0400-- -- 44 '0 4
III I44444 fi
TJ. cc w I.- co (a CD--
~j~.- in in 0----
* It*
m m ")IVrdrdE
000 zzga, (P (P U. -.0 &
I& )U. Om
HIHmi i LL Wzz .
000 00fiz II-
82
0
c-icc
A4
0 U.
C- 0 0z - c-s I.w
w ca
0.0 0 ' W LEL
o *> w~ VC,-0 . =08 :4
4 W -- W -W -W -W2WZ O 0. I- 1. f I.- m 0.I-m -mW0
*.00 ,- 40 0 2 Z 210 aIF i- 7 -
-:m mal *0 10 m- M 10-4 .Il- 0*- C w0 !-. 4
-a 39d -i Ma - QW "I 4 f w-.1 I - -- 4
A -" w- A) -L MJL M M CUL <MONO$.- AA 0LS AM AM. 0
LL0 WJ a a 4b W0. 0W rCa.Z-a 0 0al 0 40 0 0 -- 0
U v -w w w w - - -w
83
0
M
az
U)
0
0
w 0 -C
z 4-
a. 4O .4 w> z 1 1 C I- I t
f4 w - -V O 0- U) U,z U,~...W ' W I £
-C 4 -C - 0 . - O
Z:- -. )C :t -- . 4. P z - 0f,- Qw wp- V- 0 & 1 0 I
> :- -- .- .
c~ ---- ---- - 0-
N CoQ a-0NOI- ,
CC CL .' M . -tnn..;
la >-a W > W3,.<t tt t0 1- 3.
X W - 0 6u'2
0 0 0 0 0
--b ah -b -l m -b A A a * A
84
w m.
00 <
2.0.4 1. 0 *- ; -.
w - W z
ciU - 93w o 0 a- -0w n Zw
I - * - w olooo~ 2 -. mo
C a w0 - 0. X Co 1. 0. 0*-Q .a 8 'C 0 *-'C: C - Ia.j -- at 1 6 z-- w
- 4 I-,.;gco§ 2 Z. 12NWu 3
CWC2~L0W00Wt U. CLI' ~ f I~~~~~- D~r z 00. L. ~a
*~WZJOCw 0E -a s.. W ~ 0Z i
1~ * * w * w w w w w w w w w w w w qw w w w w
85
U.-
I- z I
z 4
w cv w r
0w
S.r
z~~ ~ ~ ~ I- IW -0-> "0- 8. 1. 0c
6 a i !-! E0 -++w w Z-* x m W .
L w~ l c11 W- (le 42 0 m0 ++ 24 0-z+
O*;n a w4-? I -s s 00-: ;;g
W oA 66 z ;T-m l ... "c I- 1w -2V -II- ;- 2 -4-- 2 i 2 -.
rv Ww o-. -o-D4 >+ N+ M + E W +I." z . .W I
a0 LLtjM . >n CDM
fm 2i v jc i r
As4 In a0 1 -a goZ ~ l~ a ! ft a I b4. h A
86
$w
w1
wwC w r
w w 4c
aa zCLa
0c I-
wI. 9- w u
La 0 zti au+ Z -.
w Z- OC C LNC 0. 0 rr~~' Z X - Z4 3- - a
t 70 Z4 wq 0 +W CL zAwCL . LZ-w ot - - 2 -
ZZ- - - - Z . 9II - a * Uj "i Z a z , <
W* -C 4l IL E U.L LW. ou4 IL w>> -La - 0- 0 CL
4~ In9- a inZ a- , - La90 In9. - Q. .5 u. uu v v uuu ni 1
4 ~ ~~~ -- 0- 9 -LO040 -~
87
IrIw
w m(
- 0 0
0 -
4 W -Cr~ 0 00 C
IL 0 w 0. >.0cc 0 0 I-4U. -3 -. 0, 0 -.
-. a; in 0 aw - - zOz
0. -z w -C 0 w- - 0a i- 0 -00 Z 00- .w . a. Co 00 I- 5. w- 1-g a -
IA .I .0 a --. 0owl a- .9 @. -4 -w W- 4j x~ x a
CL V: C 0 - .0 0-Ia 000
K I..WB- - FK -::x . 1 W-a-' -0 !21-
*~0cc 0. 1. ". . 0- a- . 084n - m0 o- 04O
N4B-ww - M C c--ZB-017 "o 0 ZB000 +04 um ~BO.-1OOWO O K .0 W 00
w- N-i j -1f Ci. 4 -j -ZZ 3c 00 0 - - -- 4-u.inw -j .-IW B-ixLL
Ocnm-0w 0a. 1'- 0 -1 0 4W Z W 00 - 010 p1- 0 .0 -u 00.lflO- u.- U.J- - - II 0-.0 4 -"UQ0~WC-
0P 000I C,''-4
Ia a A a.w- 4 A. aO AWaOaOa U
88
,I
z 44 z L
~i z
.a - z 4 0.
Zi'J~ 1-I- I I rd .v Mm'.2 A17 9- CV 49 -1. - . - - o W- C
-v -, -- 1 * 0 0.L J 99 W. I-Cf.0 W- . -a-. . re a - e #.. W
41 0 -a ~ - .- AW -. = 4- x 0C -. --- X- Ca. CL Z 0j-M02- c.4 0- '. x ' . a. M2; aJ.40- 0 W 4 4 '
4M< - C -4 40 0-x f CL 0 a-' 0 M 4P-. W x44 0
x-' W.ru jX M - - - - f Q - ... - - - -I.- * 00- 0- -I. M-44 CL
- CL 0 1 - a- -- 110000000 - 0 (s-0 - 0 s-00 00000000 0wZ-0 . . O 2uO000000- 0 wKo- 0 wEoo 00000000 OW 00
NM*e~q- -r0w ;4-- - - S - .0 - - * ------------ -W *-ZZZZZZZ*- d -4 -. oi' 30a- 4i~ -u............. - ~. - -i - --. .. .. -- o ----------
'o 490*-2 S1. - - - - -re-: 4- ---------- -- .4 I - -
3. .au aj-L u2 'L4 w Z- j -z - -- -- -- - -- --- 4 -
N ,I--1--40A. 9- -C au u 33 33 30. u-.303U-a3 iu33 3 O3u1i U...LLLLU...
0. 0 0000CrdI9 0 0 0 0 0 2000 0
w -o- -w w q- -p w w
89
* * * *
I- .2 0 0 *r c 0Z I 5< Q 0-0 c u- .1 4~ * --; % CL I1-- .4 CCW Wac
- z Z W~-1 I.- )' .5 *c0 (V 4) * .JW W'- 4 W - M -0
uj ow L WWI - - <
w W z 0 2 w0 0 4c ** 2-i L( c4 ; -- W 4 o2 z -~ I - Id I W
4c D 4W Z Z~ * 0-i .>. .- W * c r
£0-~wW m m = * 44 .' A * .r-j - La. 92X . W 10 I- 0 .- W
4 W W 0. I- * W I -a2 12 - * 0 3-< z to )4c u Q Ix . - . -.. * W * *
- 0.4- i,- Z * W~4W 0 0* I
*o 5. ZW! 0-- l * L w j CL CL CL - Ir) -C- ..OC.MA. W -W -J CL M (,4
:rI00.-44CW - * W~ -ZO I. a4 44 D Ci*a* .. J-44 -I-w o* *Z) -4 -4-..0. 0... - * It 9
.00.- *..JQa30 .* C3W ciIZ~l * .:-w-4 *fl* LJW O *WI-4CW44 . a > . $. * - 0 I *
CC a ) a-.0 1 1. 1*-l0 ZW.-.I -I . - - * le z E2 - En
. .- .Z I . W s W I I I I- OZ L U I . .iaI d c8r
............. 9*x5 - * W 4- * z wo
*-WBWU o- ------ ~ 0 - - -4 S -40 * 2- - -- - -- - -- --- II- --................. I-- - - - -- ' -. --- o 0
a * -d 0"'I t--x*00 0 : a a zZ
.1 NNNN~ edevf 00000 *
jOOOOOOOOOOO50N O0 0000 000 00000 * 0~~~~ciccveqeve NNN ciicc Lwvq e *- 0000000 u - 00000000 .Q
*b a Ab b *i *f *h da Am ab Ab -h a a b A a aa
90
Iw
11
.00
0'
0 0*0
w -D- - :Zo - 0 ! ;0-o 1000
.- ,
- > w 'a 1 0 K I4. K - VM 1 C oT a XC>0 '-az .- a
N 0 0- z' Dio'liom in WO- 53 10 0 06 1 w x0 M0 w
U 91
REFERENCES
[1] Astrom, Eykhoff, "System Identification - A Survey," Automatica,vol. 12, 1971, pp. 123-162.
[2] Rajbman, N.S., "The Application of Identification Methods in theU.S.S.R. - A Survey," Automatica, vol. 12, 1976, pp. 73-95.
[3] Sage, Melsa, System Identification, Academic Press, 1971.
[4] Bekey, Beneken, "Identification of Biological Systems: A Survey,"Automatica, vol. 14, no.1, 1978, pp. 41-47.
[5] Landau, "A Survey of Model-Reference Adaptive Techniques - Theoryand Applications," Automatica, vol. 10, 1974, pp. 353-379.
[6] Kudva, Narendra, "An Identification Procedure for Discrete MultivariableSystems," IEEE Trans. Auto. Control, vol. AC-19, no. 5, Oct. 1974
[7] Sprague, Kohr, "The Use of Piecewise Continuous Expansions in theIdentification of Nonlinear Systems," ASME J. Basic Engineering,June 1969, pp. 179-184.
[8] Gallman P., "A Comparison of Two Hammerstein Model IdentificationAlgorithms," IEEE Trans. Auto. Control, vol. AC-21, February 1974,pp. 124-126.
[9] Landau, M. and Leondes, C.T., "Volterra Series Synthesis of Non-linear Stochastic Tracking System," IEEE Trans. Aero ElectronicSystems, vol. AES-lI, no. 2, March 1975, pp. 245-265.
[10] Gallman, P., "An Interactive Method for the Identification of Non-linear Systems Using aUryson Model," IEEE Trans. Auto. Control,vol. AC-20, no. 6, December 1976, pp. 771-775.
[11] Landau, I.D. "Elimination of the Real Positive Condition in theDesign of Parallel MRAS," IEEE Trans. Auto Control, vol. AC-23no. 6, December 1978, pp. 1015-1020.
[12] Johnson, C.R., "SCHARF", Submitted July 1979 to Proc. IEEE.
[13] Wellstead, P., "Model Order Identification Using an AuxillarySystem," Proc. IEEE, vol. 123, No. 12, December 1976, pp. 1373-1379.
[14] Van den Boom, A. and Van den Eden, A., "The Determination of theOrders of Process and Noise Dynamics," Automatica, vol. 10, 1974,PP. 245-256.
[15] Unbehauen, H. and Gohring, B., "Tests for Determinating Model Orderin Parameter Estimation," Automatica, vol. 10, 1974, pp. 233-244.
I
7 A-A094 411 TEXAS A AND M UNIV COLLEGE STATION TELECOMMUNICATION--ETC F/6 12/2NONLINEAR MODEL IDENTIFICATION FROM OPERATING RECOROS. (U)NOV A0 B K COLGURN, A R SCHATTE F'49620-79-C-0091
UNCLASSIFIED TCSL-80-1 AFOSR-TR-81-0076 NL2EF
I [16) Kheir, N.A. and Holmes, W.M., "On Validating Simulation Modelsof Missile Systems," Simulation, April 1978, pp. 117-128.
m [17] Lion, "Rapid Identification of Linear and Nonlinear Systems" AJournal, October 1967, pp. 1835-1842.
[18] Hang, "The Design of Model Reference Parameter Estimation Syst,Using Hyperstability Theories," Proc. 3rd IFAC Sym. Sys. IdentJune 1973, pp. 741-744.
[19] Sehitoglu, Klein, "Identification of Multitalued and Meibory Noilinearities in Dynamic Processes," Simulation, Sept. 1975, pp.
[20] Tomizuka, "Series-Parallel and Parallel Identification Schemesa Class of Continuous Nonlinear Systems," ASME J. Dyn. Sys. MeCon., June 1977.
l [21) Colburn, B.K. and Schatte, A.R., "Nonlinear Dynamic System Idefication Using the MRAS Approach," Proc. 13th Asilomar Conf. Cand Systems, Nov. 1979.
II
~I
lt(
-~ I
-l.. .... .. . ... . .. . . , _. . - -"