RCH1EF
I.
INTRODUCTTAn algorit
was presented in (1]
for determining
unknown parameters
in an
otherwise known dynamical system from
measurements of the time
history of one
state variable.
The system
considered
could be nonlinear and of any order
and
the unknown
parameters could appear
non-linearly.
The method determines
the
parameters by seeking
the least
square
error between the calculated and
Observed time history.
A Somewhat
different technique
that bears a
strong
rese1lance to this has
recently been
presented by Chapman
and Kirk (2].
A computer
program based on the
algorjt
of [1] was
developed in which
only the dynamical equatio
(and the
equations of
differential
corrections)
need to be
supol.je
In applying the
program to identify systems with a large
ntmther of unkø
paraxfleters
itwas-found that
sometimes the
parameters were
not determined uniquely and
sometimes
convergence was imoossiblo
to achieve.
It was felt
that
ifficultjes
could be eliminated by basing the
identi-fication on
measurements of more than
one state variable, and so the algoritJ
and computer program have been
generajjz.
ed to include
any number of
measured
records.
Subsequent application
of this
program to several
Systems has
sub-stantiated the capability of the
genera'-ized algorjt
to identify
Inany unknown
parameters in large
Systems.
Further-more, because the
method is of the
hoOtj
t
e, memory
are
relat.ve y small so that the
identi-fication can be
accompijse on a
computer of moderate
size.
The generalized
algorithm is
presented below together
with examples
work was
supported by the Naval
Ship Research
and Development
Center,
ava1 Air Systems
Command and Oceazijcs.
internal research.
1
?IL4U LZt5'UC_AL KINETICS*
Theodore R. Goodman
and Theodore P.
Sargent
OCEANIcs, Inc., Plainview, New York
ABSTRACT
.-r
Lab. v.
ScheepsbouwkuncJe
Technische Hogesthool
De1IL
A METHOD FORIDENTIFYING NONLINEAP. SYSTEMS WITH APPLICATIONS TO VEHIcr.1
A mathematical
method is presented for determjnjn
unknown parameters in
an
otherwise known dynamical system from
measurements of the time
history of one or
more
of the state
variables.
The system considered
can be nonlinear and of any order and
the unknown
parameters can appear
nonlinearly.
A computer
program was developed in
which only the dynamical equations
and the equations
of differential
corrections need
to be supplied in order to carry
out the identification
The results of applying
the
program to five distinct
syStems are presented.
The fiVe systems
are:
a surface
ship, a hydrofoil
craft, a V/STQi. craft, a conventional airplane
and a chemical
kinetic process.
of its application to five different
Systems.
The examples presented
demonstrate the
versatility of both the
algorjt
and the computer program in a
variety of different situations.
II.
DERIV7 OF THE BASIc ALGORITHMThe dynaxnicaj.
equations describing
the system
are assumed to be given in the
form
- V ? k.
= g
(yj,
cic);
y(0)
C(1)
where the dot denotes
differentiation
with respect
to tine t, a denotes
the
unknown parameter
vector and c denotes
the initial value
of the Solution
vector
y and, may or may not
be totally known.
SUppoSe measuremants
b1mg b2,
have been taken
of the state
var.ablas
y1,
21 .. at times t
It is required
to find an initial
vecor c together with
a Parameter
torawiicl minimize
the
s
of the squares of the deviations:
+ r
best agreem
with the
measuremen
in
a least square
sense.
Solutjo
of (1) is sought
which is in
where the weighting
factors w are chosen
the same order of magnitude.
to make each sum nondimensional and
Thus, the
of
m=l
V
[Y2(tm)_b2ij2
+ ...
(2)
The parameter
vector, a, will be
Suppressed in (1) by
considering its
components to be additional
state
variables hto
the equation
1 VThe number n is
thereby increased to
include the
addjtic)
state variables
j
determined in the following
way; The'" ,initial vector is estimated and (1)
Iii"ntegrated.
is
The esIiiid initial vector
is denoted by
and the resulting
solution of (1) by3. The deviation
can
then
e calculated 'hd its value denoted
by c*
Suppose the initial vector to be
changed by an increment óc; this would
cause the solution vector to be changed
by an increment 6y and the deviation
by
an increment 6c.
From (2) it is seen
that
S
S
= v1
[Yl(tm)_bn]6Yi(tm)
+ ...
(4)The equations which the incremental
solution vector satisfies are called the
equations of differential corrections
and are obtained by expanding
(1) in a
Taylor series and retaining only linear
terms:
n
*d;(t)
6Y(5)
j=1.
The asterisk means that the coefficients
are calculated using solution
y.
Eguatiàn (5) is now integrated
thejth tine the integration
n times;
is performed.
the initial conditions are
that óy.(0)=l
and all the other dy. (0 's vanjsh.
This
s ecia]. solution is enoted by
and
the genera
so ution can then
rtten,
by superposition:
n
=
j
(Y
ii
Ct))(6)
This incremental solution
vector is used
to express 6y1, 5y2,
... in terms of óc,
and upon substituting into (4) and
inter-changing the order of summation
the
variation of the deviation
becomes
mation on j from 1 to
n.
The variation
where the repeated suffix ianlies
sum-of the deviation has thus been
expressed
directly in terrs of the variation
of
I.'.
=
6c{w1
[Y1(tm)_bl]45Yl(tm)
+...
}
(7)
each of the initial conditions.
In
order for s to be minimum 6c
must vanish
for an arbitrary variation in
the initial
conditions.
This meais that if
C
and the
xt'nded c vectorjnr!1,s the
U.unknown parameter vector
(additi
the state variable initial conditions.
The parameters of. the system will be
is defined to be
=
1
:
[Yltm_bjjYijtm+
then the error will be minimized with
respect to the cIs whenever
0, j = 1,
..
In general, using the estimated
vector c
and the resulting solution
vector y* the values of U. will not
vanish.
Denote the valueof U. as
calculated in this way by U.
Theobjective is to make the U.3vanish by
an iteration procedure.
Cnsider the
increment in U1 caused by the increment
my.. From () there is obtained
6U v1
6y1(t)
m=l
6Yij(tmH
:"
(10)
In order for each U. and hence 6c to
vanish the conditjo2
Uj =-.0
must be imposed.
Upon substituting (6)
into (10) and interchanging the order of
summations there is finally obtained
=
{wi
y13 (tm) 45Y1i(t)+.
(12)
Equations (8) with y = y, together
with
Equations (11) and (12), constitute
n
simultaneous linear algebraic equations
for the n unknowns
Upon adding the
incremental values to the estimated
values of c, improved estimates of the
c4 are obtaj.ned, and the procedure is
then repeated until
convergence is
achieved.
A modification of the above algorithm
which at tines is found to be
useful is
to introduce some or all of the bs
into
the right hand sides of Equations
(1) and
(6) in place of the respective
ye's.
III.
GUIDELINES ON USING THE PROGRJMThe algorithm presented in Section
II is in the form of an iteration
procedure, and as with most iteration
procedures convergence is not assured.
The principle
purpose of this section
will be to show how convergence can be
achieved even with a
poor initial guess
for the unknown parameters (c*Is)
and how
confidence in the final results
can be
established.
(9)
2
'-
-I,',
[
---/
Convergence depends upon a number of
factors, the most evident being the compatibility between the mathematical model and the physical system. Assuming such compatibility to exist, the next most evident factor is the quality of
the initial guess of the unknown para-meters. Short of saying that they should
represent the best available guess, no further prescription is offered.
Fre-guently, however, even the best-initial guess is not good enough to achieve
convergence with a given set-of records. -When this situation occurs, recourse may
be made to a number of different
strat-egies. The first of these has already been mentioned in the introduction,
namely, using the measured records of
additional state variables if available.
The second strategy is to use a
gain' on the óc vector. This can be
explained by considering that the 6c vector when added to the c* vector should move the updated c vector in the direction in c-space such that c/tends towards its minimum. Sometimes, however,
6c may be so large that the updated value
of c lies outside the domain of convergence even though convergence might have been achieved by preserving
the di1rectjori in c-space indicated by óc but reducing its size. This can be
accomplished simply by multiplying the 6c vector by a scalar less than unity (the gain) before adding it to c. Thus,
the gain can be considered as a safety
valve to inhibit an overcorrectjon.
The most powerful strategy for inducing convergence is by varying the
length of record. If the record is long
and
the guess for the cs is poor, thenthe trajectories calculated from Equatior. t Cl) may bear no resemblance to the
measured records and the ensuing
iter-ations will diverge. Difficulties can also occur in the incremental solution vector, Equation (5), because of a
resonance phenomena. Resonance can be
demonstrated explicitly when the
dynamical equations are linear, for then the equations of differential correction
-. are of the same form as the dynamical
equations except that they have
resonating forcing functions. It might
be expected that even for nonlinear systems a kind of
resonance
may also occur, in which case the incrementalsolutions will become large over a
sufficiently long time interval. This
behavior is clearly opposite to the ultimate behavior required for con-vergence, which is that the incremental
solutions vanish identically. Thus, if
too long a record is used then such a record, when combined with a poor guess
3
OCEANICS
for the c*ls, may cause the iteration
procedure to diverge. The remedy for this is to shorten the record so that
such divergence is inhibited. Hence, a
practical strategy for achieving con-vergence with a poor guess for the cs is to start with a short record and use the converged output as the guess for a
record that is somewhat longer. This
procedure is repeated until the complete
record is processed. This strategy is similar to and, indeed, motivated by the one used by Roberts and Shipman (3) in applying shooting methods to the numerical integration of two-point
boundary value problems.
-Assuming that sufficient measured data is used and that the calculated and measured trajectories are well matched, it might be presumed that the values obtained for the unknown parameters are a
good approximation to the true values.
However, it can only be said that this set of values minimizes the mean square error between measured and calculated trajectories, while it is still
con-ceivable that another set of values, grossly- different from the converged set,
and which produces a somewhat larger mean sua1e' error, still results in an accept-able fit between the measured and
-calculated trajectories. Indeed, it might be the latter Set that best
represents the physical system. This
could happen, for example, when a
parameter is sought that, plays a minor role in the data that is being analyzed,
in which case virtually no information abtit that parameter is contained in the dta and so the value determined for it. by system identification will be spurious.
The spurious value could even contaminate the values determined for the other un-known parameters.
Procedures will now be presented
which will aid in establishing
confidence-in the results and which use the method
of system. identification itself. These procedures, of course, should be
supplemented by available engineering estimates of the parameters obtained
from other sources.
The first factor that can affect confidence is the number of measured variables used as the basis of the
identification. Generally, the more,
records that are fitted the more
confidence
there is in the identifiedparameters. Thus, the record of every state variable that has been measured
is normally used.
It has already been -indicated that
gradual lengthening of the record is an
aid in promoting convergence. At the caine time, by monitoring the converged
c*Is after each lengthening, this procedure can also be.used to indicate
confidence in the identified parameters.
It would be expected that a short record would not contain sufficient information
to properly define the ce's. However,
as more information is introduced into the trajectory by lengthening the record, the c1s should stabilize, after which no further lengthening necessary.
Failure to stabilize is an indication of
low confidence.
Another means of establishing confidence is to vary the weighting factors from the values prescribed in
Section II. Clearly, if changing the weighting factors has no apparent effect
on the results, confidence is increased. If confidence cannot be established using the techniques just presented, this may be due to one of several causes. One possibility is that the data may be inadequate in one or more of the
follow-ing ways: too short a record, too large a sampling time, measurement of map-'propriate state variables, measurement
of too few state variables, lack of accuray in the data. All of these possibilities should be examined before-hand by inputting computer generated
data to the progrith isumcd values
parameters. To simulate reality 'Ei11data should be contaminated with
noise. By attempting to retrieve the assumed parameters from this data the measuring requirements for the actual
experiment can be established.
IV. APPLICATIONS 1) Surface Ship
The nonlinear three degree of
freedom system of 'equations describing
the steering and maneuvering of a surface ship are
ia -f(u) + C1v2 + C2r2 + C3u26.
= C4uv + C5ur + C6 £. + C7u26 (13)
r - C8UV
+ C9ur + c10 + C11u26where u = surge veloOity,
V = sway velocity,
r =
yaw rate,6 rudder deflection,
f Cu) known function of u that represents the difference between thrust and resistance,
C1 coefficients to be determined.
.The,umeasuredM input to the system identification program is provided by a maneuver generated on the computer with parameters typical of a conventional
.4
OCEAMCS
cargo
shiv
[4J. The maneuver chosen is a turning circle initiated by commanding35 degree rde.r angle. -
-Difficulty in converging to the known set of eleven coefficients was encounter-ed with the algorithm as formulatencounter-ed even
though the predicted time histories of u,
v and thcxrselves became indistinguish-able from their respective input records.
Efforts to improve the results by
sampling at a higher rate or by taking a longer record proved fruitless and only by retaining an inordinate number of significant digits in the input data did
the converged coefficients agree with
their known values. Unfortunately, the
accuracy of data measured by real sensors is limited. The difficulty in obtaining good coefficients was
elim-inated by applying the modification to the algorithm suggested at the end of
Section II. Using the modified algorithm
the effect of varying the sampling
period, length of record, number of records, and accuracy of the data was
investigated to indicate the measuring requirements for the identification of a
real ship or model from a maneuver. It
was found that a 'sampling period of from
one to two seconds and a record length of from one-half to one minute was adequate
for, successful identification. However,
it was found necessary to have measured all three variables, namely u, v and r,
to identify the eleven coefficients in' Equation (13). In order to assess the
accuracy requirements, the input was
contaminated by adding Gaussian noise of
2 and 5% error magnitudes, where this
error is based on the indicated per-centage of the maximum value of the
variable. The results of identification with these inputs are shown in Table 1.
Table 1. Results of System
Identif
i-cation of Surface Ship
4 Dig.
True Digit 2% 5%
Values Data Noise Noise C1 x102 -.203 -.202 -.071 .127 C2 x10 .113 .109 -.589 -1.64 C3 x103 -.214 -.213 -.164 -.091 C4 xl0 -.146 -.149 -.185: -.231 C5 -.334 -.337 -.370 -.415 C6 x1D .958 .942 .796' .611 C7 x103 .323 .322 .321 .318 x10 -.122 -.114 -.072 -.012 C9 x102. -.397 -.391 -.360 -.315 C10 -.120. -.116 -.097 -.070 C11x105 -.584 -.583 -.586 -.591
It is seen that as the noise level is
increased some coefficients remain close to their true values while others drift away and still others lose all
signif-icance. The coefficients C1 and C are
seen to deteriorate the most rapidy.
These coefficients have been shown in (4] to be of minor importance.
2) Hydrofoil Craft
The nonlinear longitudinal
equa-tioris of motion for a typical hydrofoil
craft under autopilot control are
Normal Force Equation1
h=3217
L LRPitching Moment Equation,
ei(ciwp)
r2(+d2'R)
The lift on the forward foil is
y5+C5 ci m 3
[
+ C3 6+T4] [Y6+C6 CpJ.where
2j CF0 + h - zeSimilar expressions are given for the lift LR on the rearward foil, where the unknown parameters C31C5,C61C9 are
replaced by C4I,C71C8 and C10.
Autopilot Transfer Functions,
4
I
(ra[_ts+1 + - l3 rs+lj
Flap Actuation System Transfer Functions,
4 ls
6fs+')4f
e (r2e+1, ecosunand command (14) where h heave, 0 pitch angle,v
.craft velocity (assumed constant),.5
1JCEANICS
LF,XR,&HS = lengths from c.g. to forward foil, roal foil and height sensor, = given constants,
= vertical distance from
height sensor to un-disturbed water surface at design foilborne condition,
s Laplace transform C's = unknown parameters.
To demonstrate the feasibility of using system identification to estimate
the C's from full-scale tests1 comPuter trajectories were first used as trial inputs. The
sponsoring
agency suppliedthe computer generated trajectories but withheld the values of the C's used in
the generation until the results of system identification became known.
Three sets of trajectories for each of
the three different configurations, termed A, B, and C (i.e., for three different sets of C's) were suDplied. A
run in each case consisted of a step
change in the commanded height after the craft had reached equilibrium.
Altogeth-er, there were 9 sets of trajectories
supplied.
Even
though there was no priorinformation given as to the proper values for the C 's Other than the crudest order
of magnitude estimates, no difficulty was
encountered by the system identification
program in converging to a set of C's for
each configuration and run. Confidence
in the converged results was investigated
by varying the record length as described
in Section III, and also by comparing the
results from different runs of the same
configuration. In all cases high con-fidence in the results were indicated.
The estimates from system identification
are given in Table 2.
Table 2. Hydrofoil System Identif-ication Summary Configuration True (Estimated) A B C .0450(.0439) .0450(.0439) .131 (.131) .131 (.129) C3 .17941.179) .17941.179) C4 .32761.327) .4446(.445) C5 1.912(1.97) 1.912(1.97) C6 .5097L53) .50971.53) C7 2.668(2.67) 2.053(2.06) C8 .44271.45) .4427(.45) C9 0 (0(.2)) 10.0 (10.3) C 0 (01.1)) 0. (01.2)) .09501.0942) .131 (.129) .1906(.l90) .4446 (.445) 2.114(2.18) .55791.58) 1.806(1.87) .51261.51) 0 (0(.2)) 0 (0(.l)) eco +
(Tis+l)
Wcommand - CHS) connand l5 'i + '16 CBS CR$O - h + HS0The parameters C and C0 represent
coefficients of nonlinear lift. Non-linear lift appeared to play an insig-nificant role in almost all the trajec-tories making it impossible to obtain firm estimates of these parameters. It
was only possible to estimate the order of magnitude of these parameters. This has been indicated in the table by use of an order symbol. As an example1 the entry DC. 2) is to be interpreted to mean that the value of the parameter is no greater than ±.2. Th the case of C9 for configuration B, the nonlinear lift term had been "beefed up so that it played a more significant role in the trajectory, and was therefore detectable.
Also presented in Table 2 are the true values of the C's disclosed-after
the results of identification became known. The comparison shows a remark-able agreement between the values
esti-mated by system identification and their
respective true values. 3) V/STOI, Model Tests
The next case is the determination of longitudinal dynamic stability
derivatives from model tests of a quad
configuration, ducted-propel],er V/STOL
aircrft flying at low speeds and high
duct incidence (5]. The model closely resembles a 0.145 scale dynamic model of the Bell X-22A V/STOL aircraft. The tests were performed at the Princeton Dynamic Model Track. A character.stic of this facility is its ability to
re-strict the degrees of freedom of model motion. The data presented in (5] are plotted time histories of motion in various longitudinal degrees of freedom when the model is disturbed from trimmed
level flight. This data was analyzed
in
(6] on the assumption that linearized small perturbation equations are valid.
The procedure used in (63 to determine stability derivatives was first to
analyze the one-degree-of-freedom (pitch
angle only) tests to give values for pitch damping (Me) and angle-of-attack stability (M) and then, knowing these quantities, to analyze the
two-degree-of-freedom tests (pitch angle and horizontal velocity) to give the speed stability
(Mu) and the rate of change of horizontal
force with velocity (X). Finally, the
three-degree-of-freedom (pitch angle. horizontal velocity and vertical velocity)
tests were analyzed to give the lift curve slope (Z) and the rate of
change
of lift
with velocity (Zn). Al]. of theseanalyses were carried out using the root
locus method (see e.g. [73).
Analysis of this model is complicated
6
OCEANICS
by the fact that it is. dynamically un-stable for all but the lowest duct angle
tested. The resulting divergent node dominates the motion, thereby making it difficult to measure the characteristics
of the othermodes. To make the data more tractable for analysis, known
amounts of augmented stability were introduced by adding a mechanical spring
(ke
)in
the one-degree-of-freedom testsrn
and
various amounts of damping feedback (sMê) in the two- and three-degree-of-freedom tests. This caused the time histories to be longerbefore
the modelhit the stops and, in addition, produced
oscillatory behavior.
It wa5 originally thought that
system
identification
could estimate all. the stability derivatives simultaneouslyfrom the complete three-degree-of-freedom
tests. However, the amount of pItch rate feedback introduced into the model to permit sufficiently long test runs for
analysis purposes also produced a highly
damped mode so that the characteristic roots of the response modes of the model were
widely
separated. Even though the modes are separated, they are not suE f i-ciently far apart to permit thereplace-ment of the equations of motion by an
approximate set of reduced equations. Thus,' in order to estimate all of the stability derivatives from measured data, it is necessary first to determine both
and
M from the one-degree-of-freedomtests an
then
to use the known values ofand M in two- or three-degree-of-freedom tests to
find
the other para-meters.To simplify, the presentation, certain known additional terms in the equations of motion have not been shown. These include -those due to the
displace-ment of
the model center of gravity fromthe pivot axis and the mounting linkage mass of the model. On the other hand, these terms have been accounted for in
determining the aerodynamic
character-istics from the analyzed results. The same nomenclature used in (5] and (6] has been adopted. In addition, the downwash lag (M) and rate of change of horizontal force with velocity (X) are assumed to be negligible, which is consistent with
The following is a
smunary of the equations used in (6] for analysis of themodel test data;
One degree of freedom
V
B f decrees of freedom- - Xu
+ g8 0 Uf + B eree decrees of freedom
- X, u
+ - gB
0Uf+WfZWf-Z U eO
W- M
Wf +B -Mj
- M U09 = 0
here
pitch angle,
model horizontal perturbation
velocity,
model vertical perturbation velocity,
ratio of vertLcal to horizontal
masses,
model initial velocity along flight
f path,
g = acceleration due to gravity,
= + ,
mechanical spring constant,
model pitch moment of inertia about
'
pivot axis.
The quantities -, k , I and th
t In
y
are known for each experiment and have uen accounted for in present.thg the inal results. The two unknown deny-tives, N and iç,,
ifl
the one-degree-of-reedom tests can be determined by measur-ng the oscillation period and dampimeasur-ng,ut for completeness, they are detemined ing the system identification program. e remaining stablitv derivatives are
hen found from the higher degree of
reedom tests with Me and assumed
own. Since there .s no inherent re-tniction on the number of parameters
at can be obtained by the system iden-ification program, it is possible to
etermine all four remaining derivatives
'
X, Z
and Z, directly from thethree-egree-of-frea -tests using Iguation
17), without tho intermediate step of étermining Mu and X from the two-degree-f-freedom tests using Equation (16).
th approaches have been taken here and
he results for both are presented in
Yable 3. Typical matches between bserved time histories and
those
pre-acted by the system identificationpro-ram are shown in Figure 1.
Well-defined values for Z, and
ould not be found even thoug the pre-icted time histories matched well with
he observed time histories, and it is (16)
believed that any attempt to do so
amounts to an overrefined analysis of the-available data. The reason that and
are indeterminate is probably because
a they play an insignificant role in the
measured three-degree-of-freedom
trajec-tories. There is a
discrepancy
in the values of and between the resultsobtained from the two-degree-of-freedom
tests and the results obtained from the (17) three-degree-of-freedom tests. More
confidence is placed on the two-degree-of-freedom results because the deriv-atives and have been inhibited in
these tests.
Table 3. esu1ta of V/STOL Model
Identification - Duct Angle of
.70C
xu
Zn Zw 7OCEANICS
Present Method Reference (6]
_955a
.297a _,700b 484b -.lgc - 14cAlso
presented in Table 3 forcompar-ison purposes aestimates of the
stability derivatives presented
in (61
as obtained by the root locus method. Theinput to this
method
is the damping and period measured from the data records.For real data these inputs constitute
only
a portion of the informationavail-able in the cornolete records and only define overall features of the data. As a consequence the precision with which
the unknown parameters are determined using the method -may be illusory and
no
criterion is provided to evaluate the
accuracy of the results. By contrast, the method presented here uses complete records of all measured state variables and not merely some of their overall
features.
4) T-33 Flight Test
The fourth ease is the determi-nation of the lateral stabilityand control derivatives of a T-33 aircraft
from flight test data. The equations of motien* appropriate to this maneuver are
*Supplied by the sponsoring agency.
aobined
from analyzingof-freedom test.
a
One-degree-b
obtained from analyzing a
two-degree-of-freedom test.
C
obtained from analyzing. a three-degree-of-freedom test.
0
100
0
LpL8Lr
.05467 a0 Y8 -l
e. 0
NpNNr
I
The data consisted of digitized time histories of p. 8 and
r
sampled everytwentieth of a second for 4.7 seconds.
It can be seen from 5,ation (18)
that there are eleven unknown parameters to be estimated from the data. thile it
is not expected that all eleven
para-meters
play
equallysignificant
roles inthe trajectory, no attempt was made
beforehand
to eliminate any from the equations, but rather the systemidenti-fication
program itself was used toindicate the confidence to be placed on
the estimated results by varying the length of record analyzed.
Table
4.
Results of T-33
.Xdentifjcatjon.8
10EA'IICS
estimaths of some of the parameters.
This disagreement can probably be
attrib-uted to a difference in weighting factors and not to the method used. The
para-meters with the greater discrepancy are
those with the lesser confidence, and so the parameters near the bottom of both lists should, at best, be considered approximate.
5) Chemical AdsOrptIon
The last case is a problem from chemical kinetics chosen to show the application of the present methdd to identify the parameters in a
differ-ential equation of a very
different
formfrom those previously considered. The
problem is to fit chemisorption rate data
to the Elovich equation:
ae (19)
where q is
the volume of material adsorbedand a and a are the parameters sought.
The Elovich equation was chosen
be-cause it is typical of the kind
encounter-ed in chcmisorption processes and also because measured data (9] and estimates
of a and
a using this
data are bothavail-in
theopen
literature. In fact, there are two independent estimates of these parameters from this same set of data,one obtained by graphical means (10] and
the other by analog matching (11]. The values of a and determined by
the present method are given
in
Table 5along with the values presented in (10] and [11]. Time histories of q calculated on the basis of the present values of a and a are shom in Figure 3 together with
measured data points taken from (9].
Although not
shown, time histories of 'qusing values for a and a taken. frem (10]
and (111 also agree well with the measur-ed data points even though their
respec-tive mean-square errors are not minimal.
V.
CONCLUSIONSA technique has been presented for
identifying many unknown parameters
appearing in large dynamical systems from
measurements
of the time history of one or more of the state variab1s. As a demonstration of the versatilityand
validity of the
technique,
it has been successfully appliedto
five distinct systems whose unknown parameters wereTable 5. Results of Identifying
Chemisorption Parameters Present Temp. Method [10] (11] 100°C a .0146 .0164 .0167 s 7.79 8.50 8.37 132°C a .0388 .036 .0364 a 7.09 7.00 6.92 Present Quasi-Method Linearization -,991 -21,1 -20.0 a -2,67 -2.51 -10.4 -9.40 Lr 2.07 1.59 N3 6.87 6.61 -.299 - .377 .974 .694 .776
.0904
a0 .102.0148
N8 -.860-1.37
T a est.mated resu ts are presented
in. Table 4 in order of their confidence,
and the trajectories predicted on the basis of these results are shown in
Figure 2. Also shown in Table 4 and Figure 2 are
a
determination of the para-meters and the corresponding trajectoriesas obtained using the method of
quasi-linearization (81*. Both methods
dupli-cate the measured trajectories very well, even though they do not agree on their
tAnalyaie of the date using quasi-linearization was carried Out at the Cornell Aeronautical Laboratory and Supplied by the 5pOnsoring agency.
where
bank
angle, roll rate,8=
sideslip angle, yaw rate,,ither givcn
ifl
advance or had beendeter-med
by an earlier Lavestigator using are established method. A feature of
.c present
technique,believed to be
iquc, is the capability of determining level of confidence -for each of the
verged pararetere while in the curae
f performing the identification. urthermOrc, the system identification omputer program is structured in euch a
ay that
the user need only supply state-ts describing the system: then, witho requirement for any apptoximatioa to
e system solution and, in fact,
no
re-irement for any additional mathematicalnalysis whatsoever, the user can expect o obtain estimates of the
parameter;,
e auality of which will be as high as be sustained by the data.
FERENCS
GOODMAN, T.R., "S::stiem Zdentiuication and Prediction An
Algorithm
tsing a Newtonian Iteration Procedure,' Quart.Applied Math, XXIV, L-.!! £49-255. CHAPMAN, C.?. and
RIF!c, D.B., 'A
Method for Extractina Aerodynamic Coefficients from Free-Flight Data," AIAA Journal, 8, 4, 1970, 753758.ROBERTo, S.M. and S!IPMAN, J.6., "Continuation in Shooting Method; for
Two-Point Boundary Value Problem;,"
J. Math. Anal. and Appi. 18, 1, 1967, 45-58.
STRQM-EJSEN, J., "A DLi.tal Computer Technique for Prediction of Standard
Maneuvers of Surface Ships," DTIIB Report 2130, Dcc. 1965.
PUTNAM, F., TRAYBAR, J.,, CURTISS, H.C.
JR., and <ukon, J.P., "An Investiga-tion of the Dynamic Stability
Char-aoteri;tic; of a Quad Configuration,
Ducted Propeller V/S?OL Model, Volume
II, Phase II - Longitudinal Dynanics
at High Duct Inc1enc Data Report,"
USAAVLABS T.R. 68-493, 1968.
CUR'PISS, H.C., JR., "An Investigation
of the Dynamic Stability
Charnctor-iStica
of a Quad Configuration
Ductad Propeller V/STO Model,' Vol. IV,USMVLABS T.R. 68-49D, May 1969.
SAV7T, C.J., 'Central System Design," McGra-Hi11, New York, 1958.
BELLMAN P.., AC-IWAi)A, ., KAIJ.DA, R.,
"Quasi-Linearisation, System
Identi-fication, and Prediction," p.and Corp. Memo P.M 3812-PR, Aug. 1963.
TAyLOR, H.S. and WILLIAMSON, A.T., "Molecular and Activated Adsorption of Hydrogen on Mangancus Oxide Surface;," J. Mier. Chem. Soc., 53,
1931, 2166-2180.
0. TAYLOR, H.A. and THOM, N., "Kinctic of Chcaisorption," J. Anar. Chem.
9
OCEANICS
Soc., 74, 1952, 4169-4173. 11. SOt4EP.FIELD, 3.?., "Solving the
Elovich Equation by Analog Computer,"
Instrument and Control Systems, 41, 1968, 137-138.
..
10
2.0
Figure 1
V/STOL MODEL TEST
TII4E HISTORIES
OCEAMCS
time-sec.
Data,Ref. (5),Run 980
Fit using present method
\
/
time - sec.
3
C4
5
S
I
-25
C)
-30
no
-4
-6
-8
-10
time - Sec.
2
\
timo - sec.
Figure
T-33 FLIGHT TEST TIME HISTORIES
Fit using qu1-arizatñ
/
/
1 \2
.3
time - soc
.2
-.2
S-.4
a)V
-.6
S S U S
.
S S S S40
20
80
time-mm.
40
time
-mm.
OCEANICS
120
60
160
80
Figure 3
TIME HISTORIES OF VOJUNE ADSORBED
O