6 Ü3
1975CRANFIELD
INSTITUTE OF TECHNOLOGY
Bibliotheek T U Delft
Faculteit der Luchtvaart- en Ruimtevaarttechniek Kluyverv^/eg 1
2629 HS Delft
ON THE ADEQUATE MODEL FOR
AIRCRAFT PARAMETER ESTIMATION
by
Cranfield Institute of Technology
ON THE ADEQUATE MODEL FOR AIRCRAFT PARAMETER ESTIMATION
by
SUMMARY
The problem of the selection from measured data of an aircraft of
an adequate model which v.ould he the simplest and sufficient
approximation to the correct model and which would facilitate the
successful determination of the unknown parameters is discussed.
Two v/ays for the proper model structure verification are recommended,
namely sensitivity analysis and/or testing of a hypothesis as to the
significance of unknovm parameters in the model proposed, and the
analysis of residuals. Finally some approaches towards the assessment
of parameter and adequate model accuracies are proposed. The procedures
mentioned are demonstrated in an example.
CONTENTS
Page
1 . INTRODUCTION "•
2. CORRECT AND ADEQUATE iMODEL 3
3. SIGNIFICANT PARAMETERS 5
4. ANALYSIS OF RESIDUALS 9
5. ACCURACY OF THE MODEL AND PARAMETER ESTIMATES 11
6. EXAMPLE 12
7 . CONCLUSION ^^
8 . REFERENCES 17
TABLES
LIST OF TABLES
Table
1. Estimates, Sensitivities and t- Statistics for Different
Models Used («^ = 20.5 deg).
2. Comparison of Parameter Determined from Repeated Measurements.
(«g = 7.2 deg).
LIST OF FIGURES
Fig.
1. Comparison of Time Histories Measured with those Computed.
(a = 20.5 deg).
2. Time Histories of Normalized Residuals (a = 20.5 deg).
3. Sample Autocovariance Functions of Residuals (a = 20.5 deg).
4. Cumulative Frequency of Residuals. Adequate Model (a = 20.5 deg).
5. Input Forms Used in the Excitation of the Aircraft Motion
(ag = 7 . 2 d e g ) .
6. Comparison of Time Histories Measured with those Computed. Run 1
(og = 7.2 deg).
7. Comparison of Time Histories Measured with those Predicted. Run 1
(cg = 7.2 deg).
8. Time Histories of Differences betv/een Measured and Predicted
Output Variables. Run 1 (« = 7.2 deg).
NOTATION
g acceleration due to gravity, (m/s^)
H sensitivity matrix
M a) pitching moment, (Nm)
b) information matrix
m number of outputs
H number of data points
n reading of the vertical accelerometer, (g units)
* g
q a) pitching velocity, (rad/s)
b) number of unknown parameters
R measurement noise covariance matrix
R autocovariance function
r correlation lag number
s(.) standard error
2
s (•) variance estimate
t a) time, (s)
b) Student's t - v a r i a b l e
u a) longitudinal airspeed component, (m/s)
b) input vector
Z vertical force, (M)
X state vector
y output vector
z measurement vector
o a) angle of attack
b) level of confidence
0 vector of unknown parameters
n eleven d e f l e c t i o n , ' ( r a d )
0 pitch angle, (rad)
V residual
o standard deviation ^ ,
a^-
standard error
Matrix exponents:
T indicates transpose matrix operation
-1 indicates inverse natrix operation
Subscripts:
E measured quantity
P predicted quantity
e steady-state value
Additional notation:
cov( •) covariance
|.| determinant of a matrix enclosed
over symbol indicates an estimate
INTRODUCTION
The e s t i m a t i o n o f s t a b i l i t y and c o n t r o l parameters from f l i g h t data** has Lecome the standard procedure f o r a i r c r a f t and f l i g h t c o n d i t i o n s where the: aerodynamic c h a r a c t e r i s t i c s can be described i n l i n e a r terms o n l y , and where no s i g n i f i c a n t e x t e r n a l disturbances are presented. New t h e r e i s a tendency t o widen the parameter e s t i m a t i o n technique i n t o f l i g h t regimes w i t h n o n - l i n e a r aerodynamic e f f e c t s and i n t o f l i g h t manoeuvres f o r which the non-linear functions o f the a i r c r a f t s t a t e s expressing the i n e r t i a forces and niotrents, and g r a v i t y terms must i n general be used.
The problem of modelling a complicated system raises the fundamental question o f how complex the model should bo. Although a more complex model can be j u s t i f i e d f o r proper d e s c r i p t i o n o f a i r c r a f t m o t i o n , i t has not been c l e a r i n the case of parameter e s t i m a t i o n what would be the best r e l a t i o n s h i p between model complexity and measurement
i n f o r m a t i o n . I f too many unknown parameters are sought f o r a l i m i t e d amount o f d a t a , then a reduced confidence i n evaluated parameters can be expected ( l a r g e covariance and/or non-physical values of some parariie"".ers), or attempts t o i d e n t i f y a l l parameters niiglit f a i l .
I n the f i e l d of system i d e n t i f i c a t i o n w i t h general a p p l i c a t i o n a number of d i f f e r e n t methods f o r determining an adequate model have' been developed. Most o f these nietnods introduced i n R e f . l are
connected w i t h the determination of model order in parameter e s t i m a t i o n f o r the s i n g l e i n p u t - s i n g l e output syste.-ii.
One of the f i r s t attempts t o t e s t the correctness of the model r e p r e s e n t i n g an a i r c r a f t was introduced in K e f . 2 , The appropriate
s t a t i s t i c was formed by the r a t i o between variance estiirates of the neasureJ frequency response c u r v e . Gne o f these cstiiiiates was outdined from repeated measurements under the same c o n d i t i o n s , the other from the r e s i d u a l s .
In Ref ,3 the analysis of r e s i d u a l s v/as recommended f o r checking
for finding the parameters in the model proposed whose effect on the
measured responses v/as neglinible. Neither approach, however, has
been brought to a conclusion as to the adequate structure of the
model used.
More comprehensive treatment of the model structure determination is
proposed and used by Hall and others in Ref.5. It is based on the
testing of model parameters in the regression analysis. A criterion
for retaining significant parameters is proposed. An adequate model
is developed either from simulated data using a priori knowledge from
wind-tunnel measurements or other sources, or from measured flight
data which are then analysed by the more advanced technique, e.g. the
maximum likelihood method. In the first case the results can be used
for the proper design of an experiment, mainly an optimal input. In the
second case the expected adequate form of the model enters the final part
of the analysis providing even better estimates of unknov/n parameters.
The procedure mentioned assumes, without any explicit statement, the
same adequate model for both the regression analysis and the maximum
likelihood method and no effect of bias errors in the least-squares
para-meter estimates on the F-statisties used for testing of the parapara-meter
significance.
Stephner and Mehra in Ref.6 proposed a new criterion for fit error.
This criterion does not improve monotonically with the increased number
of unknown parameters, but has some minimum value which could determine
the optimal number of parameters included in a m.odel.
Finally Taylor in Ref.7 developed another criterion for finding the
optimal number of unknown parameters which is based on the expected model
response error. The criterion selects the most promising model from
various candidate ciodels.
In the report presented an attempt is made in the generalisation of
experiences fron: Rof.3 and Ref;.4 to achieve an objective approach towards
a decision as to an adequate model for parameter estimation from a given
set of flight data.
CORRECT AND ADEQUATE MODEL
To represent any f l y i n g v e h i c l e completely v/ould be a task of immense d i f f i c u l t y . The c o r r e c t model of an a i r c r a f t i s i n general unknown and unknowable. In system i d e n t i f i c a t i o n the problem i s t h e r e f o r e the s e l e c t i o n from m.easured data o f an adequate model which would be the • simplest and s u f f i c i e n t approximation t o the c o r r e c t model and v/hich v/ould f a c i l i t a t e the successful determination of the unknov/n parameters. An adequate model should include only these terms which have physical meaning and which s i g n i f i c a n t l y i n f l u e n c e the output o f the system.
The i d e n t i f i c a t i o n procedure w i t h an adequate model w i l l wery l i k e l y r e s u l t i n greater f i t e r r o r than t h a t w i t h the over-parameterized one. On the other hand the use of an adequate model should provide more accurate parameter e s t i m a t e s .
The determination of an adequate model involves tv/o s t e p s , namely c h a r a c t e r i z a t i o n and v e r i f i c a t i o n . For an a i r c r a f t of orthodox design i n a s m a l l - d i s t u r b a n c e l o n g i t u d i n a l or l a t e r a l motion around some steady-state e q u i l i b r i u m c o n d i t i o n s , the equations of motion are y/ery w e l l known and no problems are expected in the f o r m u l a t i o n o f an adequate model. In g e n e r a l , hov/ever, the c h a r a c t e r i z a t i o n could be more d i f f i c u l t and must take i n t o account the known physics of the i n v e s t i g a t e d manoeuvre and the a p r i o r i knowledge of the aerodynamic c h a r a c t e r i s t i c s of the a i r c r a f t , mainly obtained from wind-tunnel m.easurements. These considerations w i l l then be r e f l e c t e d i n model c o m p l e x i t y , i . e . the number o f s t a t e variables and the form o f expressions f o r the aerodynamic forces and moments.
The models f o r l a r g o - d i s t u r b a n c e manoeuvres or coupled motions of an
a i r c r a f t can i n some cases be s i m p l i f i e d by i n t r o d u c i n g the less s i g n i f i c a n t s t a t e v a r i a b l e s as the a d d i t i o n a l i n p u t to the system. As examples t h e r e are the s h o r t - p e r i o d l o n g i t u d i n a l motion w i t h small airspeed f l u c t u a t i o n s and the r a p i d - r o l l i n g r.anocuvro w i t h induced l o n g i t u d i n a l motion.
A m.ore d i f f i c u l t procedure, however, w i l l be i n deciding which aero-dynamic parameters adequately describe the i n p u t - o u t p u t data f o r a given
mianoeuvre. I f the l i n e a r approximation seems to be i n s u f f i c i e n t , then the quadratic f u n c t i o n or the e n t i r e polynomial must be used. Both i n t e r p r e t a t i o n s r e s u l t from s e r i e s expansions around some t r i m c o n d i t i o n s assuming continuous d e r i v a t i v e s .
The n o n - l i n e a r r e p r e s e n t a t i o n of the aerodynamic forces and moments has been s u c c e s s f u l l y used i n some cases, as reported in R e f . 4 , by Wells i n Ref.8 (quadratic approximation) and by Hall and others i n Ref.5 (polynomial, f o r m ) . However, even i n those cases where non-linear aerodynamic terms are s u b s t a n t i a t e d , the l i n e a r model can s t i l l be good approximation and the estimated parameters from i t could provide a d d i t i o n a l a p r i o r i
i n f o r m a t i o n .
When the parameter e s t i m a t i o n i s completed, the accuracy of the model s t r u c t u r e should be v e r i f i e d . In t h i s r e p o r t tv/o ways are recommended. The f i r s t i s connected w i t h the e f f e c t o f a l l unknown parameters included i n the model proposed. S e n s i t i v i t y analysis ( s u b j e c t i v e approach) and the t e s t i n g o f a c e r t a i n hypothesis ( o b j e c t i v e approach) can be employed.
The second assessment i s based on the analysis of r e s i d u a l s , and i s considered as the i n v e s t i g a t i o n of the model's o v e r a l l adequacy. When both checks are p o s i t i v e , the f i n a l accuracy of the model and parameters can be e s t a b l i s h e d . I f the opposite r e s u l t s are o b t a i n e d , then the whole i d e n t i f i c a t i o n must be repeated on the c h a r a c t e r i z a t i o n l e v e l .
SIGNIFICANT PARAMETERS
The state and output equations of a system with no process noise can
be v/ritten as
f(x, u, 6 ) , x(0)
(3.1)
g(x, u, e)
(3.2)
where x i s an (n x 1) s t a t e v e c t o r , u i s an (P x 1) i n p u t v e c t o r , B i s a (q X 1) vector o f unknown parameters and y i s an (m x 1) output v e c t o r . I t i s assumed v/ithout any f u r t h e r loss o f g e n e r a l i t y t h a t a l l the i n i t i a l c o n d i t i o n s are equal t o zero.
The concept of s e n s i t i v i t y can be t r e a t e d as the s e n s i t i v i t y o f the f u n c t i o n y = g ( x , u , 8 ) w i t h respect to the parameters. Therefore the parameter s e n s i t i v i t y f u n c t i o n s are
'^w
ae,
/ k = 1 , 2 , . . . ,m
' ' ' ^ j - 1 . 2 . . . . , q
At each time i n t e r v a l t - the values of the parameter s e n s i t i v i t y f u n c t i o n s form the elements of the (m x q) s e n s i t i v i t y matrix H . , i = 1 , 2 , . . . , N , where N i s the number of data p o i n t s . I f the modified Newton-Raphson computing technique f o r parameter e s t i m a t i o n i s used, then the s e n s i t i v i t y matrix enters the i n f o r m a t i o n matrix v/hich f o r the maximum l i k e l i h o o d (ML) e s t i m a t i o n i s given as
N M = z i = l ( — - ) ^-^ f • ' \ N i = l r T 1 1 H. R H . ( 3 . 3 )
v/here R i s the measurement noise covariance m a t r i x ,
For the comparison o f a l l s e n s i t i v i t y functions r e l a t e d t o d i f f e r e n t parameters, the r e l a t i v e s e n s i t i v i t y f u n c t i o n
1 9 y k ( t ) B- 9 y k ( t )
Oy a i n p . a^ aPj
of the measureP'.ent noise.
Nevertheless, the assessment of the sensitivity based on (3.4) is not
very practical. It is therefore preferable to replace the time functions in (3.4) by constant terms defining the sensitivity for the time interval considered.
One o f the possibilities v/ould be the formulation of the term for the sensitivity as
0,^
1 = 1 I
\
J "•''
taking into account each output variable separately, and the formulation of the term for the sensitivity of the v/hole system as
k J
Comparing (3.G) with the form of tht information matrix in (3.3) yields the relationship between the relative sensitivity and the information matrix in the form
2 m , N ray^(ti) ,2
J k = l OyJ- 1=1 93j J "J JJ
= B ' M , . (3.7)
where M.- i s tfie main diagonal element i n M corresponding t o &• JJ J •
The o v e r a l l s e n s i t i v i t y ( 3 . 6 ) w i l l depend on the c h a r a c t e r i s t i c s o f the system, tiie design o f the experiment, and the accuracy of measurement. The nuhierical values o f ( 3 . 6 ) can detect c r i t i c a l parai„eters, i . e . parameters -which have a n e g l i g i b l e e f f e c t on the response o f the system and -which
t h e r e f o r e w i l l yery l i k e l y be estimated w i t h an unacceptable accuracy.
For the o b j e c t i v e d e c i s i o n as to which terms should be deleted from the
model proposed f o r paranieter e s t i i n a t i o n , a t e s t s t a t i s t i c i.-ay be used, This s t a t i s t i c has the fonr.
4. _ ! i - ( 3 . 8 )
'' ~ s - .
BJ
where B i s the parameter estimate and S j - i s the standard e r r o r o f the
J p j
e s t i i r a t e , see e . g . Tcef.S. Using the ML estimation the s t a t i s t i c i s changed as
t = —i- (3.9)
JJ
wliere M.' is the main diagonal element in the error covariance matrix for the estimated parameters.
The null hypothesis thcit 3- = ü is accepted if
'^' ^ ''^a, N-q (3.1Ü)
where a is the level of confidence (usually a = Ü.35) and N-q is the
number of statistical degrees of freedom. The value of t , ., ,. ^ . ^ a/2, N-q IS found
from the tables of Student's distribution. For a = 0.95 and N-q-x», t=1.96.
It| < t^/2,M-q, tliG alternative hypothesis that B- ={= 0 is then valid.
J
The a p p l i c a t i o n of the t e s t mentioned i s as f u n o \ / s : suppose a confidence l e v e l of 0.95 i s cliosen. The t - s t u t i s t i c is then computed f o r each estimated paraiLC'ter and the parameter v;ith n.iniiiium | t | i s found. I f I t l . >t , , ^ ' ' ' 'min a/2 a l l terms are ccncluoed to be s i g n i f i c a n t at the 0,u5 l e v e l . I f
U L - < t , , the parameter corresponding t o the Itl,.,- i s dropped from the ' 'min a/2 liiin ' ^
model and the e s t i m a t i o n i s recomputed w i t h the new model. The process i s ' rcpeatea t i l l the adequate form of the model i s reached.
the t e s t i n g procedure, the a p r i o r i weighting can be used. With a decreasing number of parameters, the c o n s t r a i n t due t o a p r i o r i v/eighting can gradually be l i f t e d and possibly abandoned f o r an
adequate model,
The adequate miodel r e s u l t i n g from the t e s t i n g of paramieters i s not n e c e s s a r i l y unique. The proper form of t h i s model w i l l depend on the i n i t i a l m o d e l l i n g , where tfie a p r i o r i knowledge as to the a e r o -dynandc c h a r a c t e r i s t i c s of the a i r c r a f t and the physics of the motion are considered.
4. ANALYSIS OF RESIDUALS
For the ML estimation, the residuals v. = z. - y. should form a sequence
of uncorrelated random variables with gaussian distribution and zero mean,
see e.g.Ref.6.
The simplest way to check some of the assumptions mentioned is the time
sequence plot of residuals. From this time history trends can be apparent
and possible deterministic components discovered. If the normalized residuals
^t,(^-;)/s(yi,), k= 1.2 ...,m are plotted, then any values greater than three
can be considered as outliers.
The check of normality uses the calculated percentage cumulative distribution,
P-, for each of N samples of the residuals V|,(t.). Because of the large
number of data points, the grouped residuals are usually used in the
calculation. The probability of the residuals v. less than or equal to the
jth class 1imit is
P. = P , v . i « j . ^ , , . (4.1)
wherev^are the midpoints of the class intervals and Av is the length of the
Jclass intervals.
It is advantageous to plot v. as abcissas and the corresponding P. values
J J
as ordinates on the probability paper, on which the ordinate scale is
graduated according to the area under a normal distribution function. Then
the fitted line of all values of P,. m.ust, in the case of a normal distribution,
be a straight 1ine.
The mean value can also be checked and is zero if the condition
P{v. < v.. = 0} = 50%
is satisfied. The difference in abcissas for P^. = 50% and P. = 15.9% is
o Jan estimate of
a.
The established value of the standard error can be
therefore used for fitting the straight line to the plotted (P,v.) data.
-J -JThe assumption of uncorrelated residuals is checked by the autocorrelation
function of the residuals. Its estimate is found from the expression
T N-r
^ ( 0 = - ^ , ^ ^ - i . r ' ^ = ^ ' » 1 - - - ' W ('-^^
i - 1
I f the r e s i d u a l s are to be u n c o r r e l a t e d , and i n the normal case also independent, then the c o n d i t i o n K ( r f 0) = O must be s a t i s f i e d .
In p r a c t i c e even f o r an adequate model R ( r f 0) i s never e x a c t l y z e r o , but v a r i e s s l i g h t l y around t h i s value. Hov/ever, these non-zero values should be v/ell w i t h i n the 2a-bounds. The r e l a t i v e standard e r r o r f o r the a u t o c o r r e l a t i o n f u n c t i o n estimate i s approximated as
— = — ( 4 . 3 ) R ( r = 0) ^
ACCURACY OF THE MODEL AND PARAMETER ESTIMATES
When the parameter and adequate p-odel estimations are completed, the accuracy of the results obtained ought to be checked. In this process presumably the most important point is the comparison between the estimated parameters and those for which either a priori values or at least limits on their values are known. This comparison must also take into account the standard errors and correlation coefficients of the estimates.
The standard errors define the confidence limits for each parameter
regardless of the remaining parameters. The correlation coefficient is the measure of stochastic dependence. There has been no effort to develop a criterion for testing the significant correlation. Usually a value of the correlation coefficient greater than 0.85 is considered significant.
In the next step of the accuracy assessment the output time history match between the actual and estimated responses found from the identified model is checked. A small fit error is a necessary but not sufficient condition for accurate result.
Increased confidence in the estimated parameters and adequate form of a model can be obtained from repeated measurements under the same conditions. The problem of applying either similar or different inputs must be carefully considered. A decision should be taken separately for each group of
repeated measurements. In principle, the differences in input forms should not substantially change the sensitivities with respect to unknown parameters If these changes occur, then their effect must be taken into account in
comparing repeated parameter estimates.
From a large number of repeated measurements the ensemble mean values and ensemble standard errors of estimated parameters can be obtained. For accurate results the two quantities are expected to be close to the corresponding ones from each individual measurement.
The ability of the estimates from one test to predict the response of another flight can also be checked. For comparing the predicted and measured output time histories, the confidence limits for the predictions may be found. These limits follov/ from the covariance matrix of prediction
error, which has the form
/ • •
cov {y(t. B)} = H.M
H T+ R (5.1)
The prediction error is defined as
Spi = 2^(3) -
y^ih
(5.2)
EXAMPLE
As an example, the parameter e s t i m a t i o n and adequate model determination f o r a slender d e l t a - w i n g research a i r c r a f t are presented. The l o n g i t u d i n a l responses o f the a i r c r a f t were e x c i t e d from the h o r i z o n t a l s t e a d y - s t a t e f l i g h t s by elevon d e f l e c t i o n . Because o f the input form used and the a i r c r a f t c h a r a c t e r i s t i c s , the airspeed changes during the t r a n s i e n t motion v/ere negl i g i b l e .
Taking i n t o account the physics o f the motion, the v/ind-tunnel measurements and p r e l i m i n a r y f l i g h t t e s t r e s u l t s , the equation of motion were formulated as a = Z a + Z q + Z 2^ + ^„a Q" + Z „ na + a q^ a / q" ' n + Z n + Z'n + Z n n o q = M a + M q + M y'^^ + Ka^°^ + ''•'' na + ( 6 . 1 ) ' a q a'^ q ^ na M n ^ M-n + M n + n 0 0 = q
In these equations only the pitch rate, q, was measured. The second output
variable n is related to the state variables by the equation
n* = Z a + Zlq + Z 9 + Z 23^ + Z„ 1^ +
Z a q^ a n^ qa '
+ Z not + Z n + Z 'n + Z ^
(f, o\
The s t a t e and out[;'jt oouations are developed i n R e f . 4 , v/hore the parameters contained i n these equations are also d e f i n e d . The ML e s t i m a t i o n techni^'jc without process noise described i n Ref.3 vas used i n the data a n a l y s i s .
The r e s u l t s from one ter-t run at h-;nh angle of attack (a = 20,5 deg.) are summarized in Table 1 . These r e s u l t s include the estimated parameters,
standard e r r o r s of the estimates (lower bounds), variance estimates of measurement n o i s e , the logarithm of the determinant of measurement noise covariance m a t r i x ( o v e r a l l f i t ) , s e n s i t i v i t i e s and t - s t a t i s t i c s . The tabulated t - v a l u e f o r a = 0.05 v/as equal to 1.98. ^
The f i r s t paraireter ostimatt.'S v.'ero obtained from the l i n e a r model
a = Z CÏ + Z 'i + Z n + Z q = M a + M t, + I'! n + M ^ a q T) : 0 e = q (6.3) Z ni + Z / ' i + Z o + Z n + Z^ o; 'i 0 r, o
where Z„ was known and Z„, Z,, and 7. v/ere fixed on the wind-tunnel values. 0 T f] n
The r e s u l t e d estimates agreed reasonably w i t h those from wind-tunnel d a t a , but the parameters f1 and M v/ere s t r o n g l y c o r r e l a t e d and the f i t between measured and computed v e r t i c a l a c c e l e r a t i o n time h i s t o r i e s was r a t h e r poor.
The next e s t i m a t i o n was based on the .complete equations ( 6 . 1 ) and (6.2) w i t h Z , Z and Z. treated as knovn values. Because of inconsistency i n
the number of unknown parameters and measured o u t p u t s , v/eighting of the basic l i n e a r parameters had t o be used. The previous estimates of Z , M ,
' ' a a
M and M from the l i n e a r model and t h e i r standard e r r o r s formed the a p r i o r i q n
data.
The e s t i n ; 3 t i o n r e s u l t e d in s i g n i f i c a n t l y changed damping parameter M , b e t t e r accuracy of the estimates, and b e t t o r f i t i n both output v a r i a b l e s
The sensitivity analysis revealed small effects of all non-linear terms
in the lift-force equation on measured responses. This v/as also confirmed
by com.paring the computed and tabulated t-statistics.
In the following estimation the insignificant param.eters Z 2> Z a and Z
were successively dropped from the model. Then the remaining parameters
v/ere proved to be significant. The adequate model obtained had increased
sensitivities, lower standard errors in some estimates, and only slightly
higher fit errors. Hov/ever, the main advantage of the adequate model in
comparison with the complete one v/as a considerable improvement in the
convergence of the iterative procedure.
Even for the adequate model, it was still necessary to use a priori
v/eighting. Nevertheless the effect of the gradual decrease of v/eights v/as
investigated. As a result, no substantial changes in the estimates v/ere
observed, but the convergence v/as slowing down and the parameter covariances
were increasing.
The measured and computed outputs are plotted in Fig.l. The time histories
of normalized residuals for the linear and adequate model are compared in
Fig.2, the corresponding autocovariance functions in Fig.3. A substantial
improvement in the shape of the autocovariance functions for the adequate
model is apparent. However, their forms are still different from those
for v/hite noise. This might be due to poor quality of measured data, which
is demonstrated by the residuals for both inodels used and also by the
cumulative frequencies for the adequate m.odel plotted in Fig.4.
The comparison of results from two repeated measurements with some of the
pararrieters obtained from wind-tunnel and steady-state measurements
(parameter M ) is made in Table 2. The inputs used in both runs are
presented in Fig.5, the measured and computed outputs for Run 1 in Fig.6.
The repeatibility of results from the tv/o runs is
very
good, as follows
from Table 2 and also from Figs. 7 and 8. In Fig.7 the measured outputs
are plotted, together with the predicted ones based on the estimates from
Run 2. The resulting differences between the measured and predicted
responses are shown in Fig.8. These differences are within the 2a bounds
for the prediction error defined b^- equation (5.2).
CONCLUSION
Because in general the correct model of an aircraft is unknown and
unknowable the problem of identification also encompasses the selection
from measured data of an adequate model. This model should be the
simplest and sufficient approximation to the correct model and should
facilitate the successful determination of the unknown parameters. The
determination of an adequate model involves two steps, namely
character-ization and verification.
The characterization must take into account the known physics of the
investigated manoeuvre and the a priori knowledge of the aerodynamic
characteristics of the aircraft. These considerations will then be
reflected in model complexity, i.e. the number of state variables and
the form of expressions for the aerodynamic forces and moments.
For the verification of the model proposed two steps have been developed.
The first one is based on the sensitivity analysis and/or testing a certain
hypothesis. The sensitivity analysis is the subjective approach which can
detect critical parameters which have a negligible effect on the response
of the aircraft under test and which therefore will very likely be
estimated with an unacceptable accuracy.
For the objective decision as to which terms should be deleted from the
model proposed the t-test may be used. In this test the computed t-statisties
for each estimated parameter are compared with the tabulated value from
Student's distribution and the insignificant terms successively dropped
from the model till its adequate form is found.
Both approaches require a very limited amount of additional computing when
the parameter estimation is completed. Their disadvantage could be in
using an over-parameterized model in the first step of the procedure.
This drawback can be overcome, however, by applying a priori weighting
during the parameter estimation,
The second step in the model verification is based on the analysis of
residuals and is considered as the investigation of the model's overall
adequacy.
Where both steps
are
positive, the final accuracy of the model and
parameters can be established. If the opposite results are obtained,
then the whole identification must be repeated on the characterization
level.
With the tendency to widen the aircraft parameter estimation into flight
regimes with non-linear aerodynamic effects and into flight manoeuvres
with other non-linear functions the increasing importance.of the proper
model structure can be expected. For this reason more experience with
the procedures developed for the determination of an adequate model and
the further research in this area are highly recommended,
REFERENCES
1. UNBEHAUEN, H, GORING, B,
Test for Determining Model Order in Parameter Estimation.
Automatica, Vol. 10, pp.233-244, Perganon Press, 1974
2. KLEIN, V. TOSOVSKY, J,
General Theory of Complex Random Variable and its A|)plication to the Curve Fitting a Frequency Response
(Summary Report).
Zprava VZLU, 2-11 December 1967.
KLEIN, V, Parameter I d e n t i f i c a t i o n Applied t o A i r c r a f t .
Report CIT-FI-73-018, October 1973 or
Cranfield Report Aero No.26. December 1974.
KLEIN, V, Longitudinal Aerodynamic D e r i v a t i v e s o f a Slender Delta-Wing Research A i r c r a f t Extracted from F l i g h t Data. Report CIT-FI-74-023, J u l y 1974 or
C r a n f i e l d Report Aero No.27, January 1975,
5. HALL, W.E., J r . GUPTA, N.K. TYLER, J . S , , J r .
Model S t r u c t u r e Determination and
Parameter I d e n t i f i c a t i o n f o r Non-Linear Aerodynamic F l i g h t Regimes.
Paper presented at AGARD S p e c i a l i s t s Meeting on Methods f o r A i r c r a f t State and Parameter E s t i m a t i o n , NASA Langley Research Center, Hampton, V i r g i n i a , November 5 - 8 , 1974.
STEPHNER, D.E MEHRA, R.K.
Maximum L i k e l i h o o d I d e n t i f i c a t i o n
and Optimal Input Design f o r I d e n t i f y i n g A i r c r a f t S t a b i l i t y and Control D e r i v a t i v e s ,
NASA CR-22Ü0 March 1973.
TAYLOR L,W., Jr. Application of a New Criterion for Model1 ing Systems.
Paper presented at AGARD Specialists Meeting on Methods for Aircraft State and Parameter Estimation, NASA Langley Research Center, Hampton, Virginia, November 5-8, 1974.
RAMACHANDRAN, S, WELLS, W.R.
Estimation of Non-linear Aerodynamic Derivatives of a Variable Geometry Fighter Aircraft from Flight Data. AlAA Paper No. 74-790, AlAA.
Mechanics and Control of Flight
Conference, Anaheim, Cal., August 5-9, 1974.
HIMMELBLAU, D.M. Process Analysis by Statistical Methods John Wiley & Sons., Inc., 1970.
Za Z„2 \ a \ . \ Zn Ma M q Ma^
V
Mria ^ >^fl1 s^q)
s 2 ( n * ) I H I R Ih
- 1 . 4 3 5 -- 0 , 2 7 -- 1 . 7 6 - 0 . 7 8 -- 3 . 5 4 s ( 6 j ) 0 . 0 8 0 -0 . 1 5 0 . 1 0 -0 . 1 1 -4 . 5 8 X 1 0 - 5 1 6 . 5 X 1 0 " ^ - 2 1 , 0 0 <%'i
- 1 . 4 3 0 - 1 . 7 1.4 3 , 2 - 0 . 2 0 4 - 0 , 0 0 5 6 - 1 . 5 8 - 0 . 3 8 7 - 5 5 18 164 - 3 . 4 0 6 - 0 . 3 3 7 s ( 0 j ) 0 . 0 5 1 5 . 9 2 . 7 2 . 6 0 . 0 1 1 0 . 0 0 1 6 0 . 1 3 0 . 0 5 6 13 9 . 2 44 0 . 0 9 3 0 . 0 2 5 B?M.. J J J 7286 2 2 ^ 646 29 3636 1153 1005 184 312 5 4 9 5 6 1374 1 . 0 9 X 1 0 - 5 5 . 2 0 X 1 0 - 6 - 2 3 . 5 9 x)11|
2 8 . 2 0 . 3 0 . 5 1.2 1 9 . 2 3 , 4 1 1 . 9 6 . 9 4 . 2 2 . 0 3 . 7 3 6 . 7 1 3 . 3 6 . J - 1 . 4 3 2 -- 0 . 2 0 4 - 0 . 0 0 4 9 - 1 . 5 4 - 0 . 4 7 - 7 0 23 179 - 3 . 2 2 2 - 0 . 3 0 7 s ( 6 . ) J 0 . 0 4 0 -0 . -0 -0 8 9 0 . 0 0 1 0 0 . 1 3 0 . 0 5 8 11 6 . 9 48 0 . 0 0 8 7 0 . 0 2 5 8 ? M . . J J J 13063 -1726 48 4485 1819 2536 436 363 79594 2973 1 . 2 1 X 1 0 - 5 4 . 8 7 X 1 0 - 6 - 2 3 . 5 5 x )Itl
3 5 . 9 -2 -2 . 9 4 . 9 1 1 . 9 8 . 0 6 . 2 3 . 4 3 . 7 3 7 . 112.5
1
TABLE 1. ESTIMATES, SENSITIVITIES AND T - STATISTICS FOR DIFFERENT MODELS USED (ttg = 20,5 deg).
ITEM Z a Z na
z
n
z.
n
M a M q M qa Mn
M.n
s2(q) s2(n*) In R| PRED - 1,7 - 0.38 - 6.8 - 1.6 -13h
- 1.525 - 0.38 - 5,366 - 1,993 -14,08s(êj)
0.031 0.091 0.062 0,23 2.28 X 10~^ _6 4,80 X 10 - 20.68 B. J - 1.516 7.8 - 0.292 - 0.0115 - 5.450 - 2.197 -17.6 -13.8 - 0.083s(ê.)
0.021 2.5 0.019 0,0012 0.061 0.056 2.9 0.15 0.032 2.08 X 10 ^ _6 2.57 X 10 - 23.65 6. J - 1.544 - 0.38 - 5,96 - 1,743 -13.31s(ê.)
0.041 0.12 0.077 0,29 -5 3,96 X 10 _6 7,70 X 10 - 21.91 B. J - 1,505 3.3 - 0.402 - 0.0096 - 5.678 - 1.754 -18.9 -12,32 - 0.495 s(B^) 0.024 3.0 0.030 0,0018 0.077 0,070 3.8 0.16 0.031 -5 1.87 X 10 _6 4.45 X 10 - 23.21TABLE 2. COMPARISON OF PARAMETER DETERMINED FROM REPEATED MEASUREMENTS, (a- = 7,2deg).
q [deg/s]
O
- 2 - 3 - 4 - 5 ^1
r
V'
/ /f
y^
\/WVv Measured ouni|Jutt;u,uueL|uuLt: model z [g units] • 2• 15
• I
•05
O - • 0 5 - .1f\
\' ' j
-V
V
t\^
r\
\ \vJ
V
\ ^/VW-WV
4 5
'[O
FIG. 1 COMPARISON OF TIME HISTORIES MEASURED WITH THOSE COMPUTED. («g = 20.5 deg).
n 6
O
- 2
O
4
t[s]
FIG. 1 COMPARISON OF TIME HISTORIES MEASURED WITH THOSE COMPUTED. (o;g= 20.5 deg). _ C o n c l u d e d .
s(q) O - 4 4
. r
\Y
'^
hh-J
Vyy*"^
- V W M A / V " (f YM^
\ ^ A A / ^ ^ • ' ' ^ .' nz O - 4 sCq) •4 4II
H/^A^
v^
/^yvM»
huH
\ *«„
WV|/VV
Adequate ModelrtJA.
^ V
# • nz - 4R
qO
Adequate Model 0 - • 5 1 0 nzO - . 5K
\.__ riN_
r s •^ ^ —^\ 1
^ ^ ^ ^ : r ^ ^ > ^
10 15
r
J
m
80
50
20
y
yo / "r
/o/°
o /p
y"
- • 6- 3
• 3[deg/s]
6
99
P. J{i\
80
50
20
n ^ O O ^ o ^ ^ . ^ ' o ^ ^ o -{•50-75
•75 1 5 0 10x1)^^ [q units]FIG, 4 CUMULATIVE FREQUENCY OF RESIDUALS, ADEQUATE MODEL («g = 20,5 deg).
[deg]
•'^•V^v^.^s/^ v v v » . . » mtm,"VV^
/fc^«w>r>rV*^8
[deg]
O
- 2
A
1
V
- \ /\ r^-^—
Run 2i 6
FIG. 5 INPUT FORMS USED IN THE EXCITATION OF THE AIRCRAFT MOTION (ag,= 7.2deg).
q
[deg/s]
O
- 2
- 4
8
10
ir^
/ /V
[ V
. . y ^ " ^'^^^^^^
/ /' . - > . - / i >^A
/vr^^rs^/ruv\
S-yVÖi / \ i ^ 1•v^v'v r V r
Measured
computea,aaequate
model
i.[g units]
• 5
4
• 3
•2
-•2
- • 3
ft
n
f
\ \ \ \ \" V s ^
•** —""^ 'VWü^A-t^^g^VAv*-^--^J
5 6
t [s]
FIG, 6 COMPARISON OF TIME HISTORIES MEASURED WITH THOSE COMPUTED. RUN 1 («g = 7.2 deg).
j
/\j
f
Measured,Run 1 Predict edjparametcrs from Run 2 0r
1
, A . .1
V
0
\ \ ; ^ . ) 1 • • • • " \ A \ ..:^.A'V^'VVA/
> 6
t[s]
G.7 COMPARISON OF TIME HISTORIES MEASURED WITH THOSE PREDICTED. RUN 1 (ag=7.2deg).
n ^ n
-[g units]
• 0804
O
- . 0 408
» 1
• v11 ^
f
yI l i
UKyn
Sivjw
\ .X
VAMKAT
Al
AA
%f
5 6
t[0
FIG. 8 TIME HISTORIES OF DIFFERENCES BETWEEN MEASURED AND PREDICTED OUTPUT VARIABLES. Run 1 (0^ = 7.2 deg).