Professional Documents
Culture Documents
1 s2.0 S1474667017631537 Main
1 s2.0 S1474667017631537 Main
1 s2.0 S1474667017631537 Main
R. E. Bach, Jr.
NASA Ames Research Cent er, Moffett Field, California 94035, USA
1155
1156 R. E . Bach , Jr .
computational advantage possible with th e When motion of the air mass must be consid -
coordinate- transformed plant . e r ed , the model is augmented with wind-
velocity s tat es (w x , wv , wz) and acce l eration
The paper is organized as follows. First, forcing functions (gx' - gy , gz) as in Eq. (4).
the usual nonlinear formulation of the rigid-
body model and its equivalent linear-state
r ep resentation are described. Then the
problem of estimating aircraft motions along (4 )
a trajectory is defined, and an algorithm is
selected to comp~re the performance of th e
formulations for solution of such a problem.
Some comments are made regarding how algo -
rithm efficiency is model - dep e nd en t . Two Since an estimatio n prob l em is generally
estimation experiments are described and concerned with "fitting " t he response of a
their results in comparing the computational state model to a set of measurem e nts in some
burden associated with the Jacobian matrices suitable way, a measurement model mus t be
for each plant are reported . Finally, a specified. The measurements available in an
possible application of the linear-state aircraft state-estimation problem usually
concept for real - time estimation of pitch include body accelerati0ns, angular veloci-
and roll angles is presented and its practi- ties, Euler angles, and position coordinates.
cal significance is discussed . These make up the linear e l ements of the
measurement model corresponding to the state
model of Eqs . (1) - (4). Also in the measure-
EQUATIONS OF MOTION ment set are aerodynamic variables (airspeed
V, angle - of-attack a , and sideslip angle 3 )
For the trajectory analysis problems consid - as well as position variables (slant range R,
ered in this paper, aircraft motions are bearing angle B, and eleva tion angle E) _
assumed to be governed by a six- degree - of - The set is completed by allowing for the
freedom kinematic model, referred to a flat, measurement of winds (horizontal magnitude
nonrotating Earth (Etkin , 1972) . In the and direction Wh, Wd and vertical Wz ).
usual formulation, aircraft states include All of these are nonlin ear functions of th e
Eu l er angles (~ , e , w) and body angular state variables, and are given by
veloci~ies (p , q, r) which are related by
(qsinlO
-
+ pj [~J
; qrSl.n~
+ rcos~)/cos e ~
=
[ a R-l
am
an
, (1)
+ rv - qw - gcin S ]
[t}m[:}[IH: + pw - ru + gsin~cos 6 ,
+ qu - pv + gcos~cos e
(2) The matrix .£ in Eq . (5) transforms from an
inertial f rame into a body - fixed frame and is
equal to the transpose of the matrix given in
Eq . (3). We should note in pass ing tha t any
where g is the gravitational co nstant and bias or scale-factor errors associated with
7n is a matrix that transforms from a body- the measurement model can be appended to the
fixed reference frame into a vehicle-carried state vec tor and tr eated as cdnstants , that is,
vertical frame; it is given by
b = 0 s = Cl (7)
Forcing functions for this state model are The system of nonlinear differential equa -
the body angular accelerations (a R- , am, an) tions (1) and (2) is said to be block triang-
in Eq . (1) and the linear accelerations ular , a class t ha t r eadily permits coordinate
(a x , ay, a z ) in (2) . For a situation that transformation into a p lant with a linear
requires accelerations to be treated as state state model. The formal procedure for gener -
variables , the state model can be extended by ating such a trans fo rmation was outlined by
specifying new " forcing" functions (d £ , d m , Meyer (1981). More ge n e ral nonlinear systems
d n ) , (d x , d y ' d z ) as shown in Eq . (4) . are considered by Brockett (1978) and Hunt
A Ma t hema ti ca l Model fo r Eff i c ient Es timat i o n of Ai rcraft Mot i on s 11 5 7
o -sin e ]
cos~ sin~cose (8)
- sin~ cos~cose
I+----p
(9)
~4f---t1 >-+---......- - - { ~
MEASUREMENT
MODEL
STATE
where ~ is the time - derivative of t he MODEL
matrix IR in Eq . (8) . Hence , the state
variables (~ , e , ~) , (~ , e,
~) , and (;p , e , ~) Fig . 1 . A transformed plant for aircraft
make up part of the linear state model . state estimation .
+
N-l /\
+ L [wT(i)Q-1w(i) + v T (i+l)R- 1v(i+l)]/2 .
W
(13)
MEASUREMENT
In Eq. (11), xCi) is the state vector and RECORD
w(i) is the (unknown) forcing-function
vector; in Eq . (12), veil is assumed to be
the error in measuring the system output Fig. 2. Block diagram of compu ting
h[x(i)] . The performance measure, Eq. (13) , strategy for solving a
is a squared- error criterion, where Xo is state-estimation problem.
an a priori estimate of the state, and Po'
Q, and R are weighting matrices . A Bayes -
ian maximum-likelihood interpretation would The smoothing algorithms shown in Table 1
consider Po as the error-covariance matrix are representative of those in present use .
for the a priori estimate and R as the The first, at least in form, is similar to
error-covariance matrix for the measurement extended Kalman-filter methods (Hansen,
sequence. Although the forcing-function 1978; Klein and Schiess, 1977; Stalford,
sequence is "deterministic ," good results 1980) . The second has been implemented as
are obtained when each diagonal element of part of a program called S~~CK (SMoothing
Q is chosen to match the expected mean- for AirCraft Kinematics) for use in flight-
square value of the corresponding element test analysis and to assist the National
of w(i). Transportation Safety Board in its investi-
gation of aircraft accidents. To determine
Two equivalent algorithms for solution of the which of the two algorithms might be more
nonlinear smoothing problem are outlined in improved by incorporating the transformed
Table 1; details were published previously plant, we should observe the manner in which
(Bach , 1980). The first algorithm consists the Jacobian matrices fx and hx enter
of a forward covariance filter a n d backward into the computational process . Generally,
smoother. Notice that sequences [e(i)] , both are sparse : for the linear state
[K(i)] , computed during the filtering pass, model, however, not only is fx very sparse ,
are utilized during the smoothing pass . but its entries are constant. On the other
Hence, the storage required depends on the hand, the hx will be more complicated and
dimensions of z(i) and xCi) and , of less sparse than the hx corresponding to
course , on the length of the data record . the nonlinear state model . Scrutiny of the
The second algorithm , shown in Table 1 , algorithms of Table 1 indicates that the
consists of a backward information filter second algorithm should be more improved,
and a forward smoother . An advantage of this since computations involving fx appear to
formulation is that the covariance of the dominate over those involving h x ' Because
a priori estimate can be ignored (P o - 1 con- that algorithm was already programmed in
sidered to be zero) . Here the sequences SMACK , it was a relatively simple matter to
[dei»), [L(i)] are utilized during the produce a version with the transformed plant
smoothing pass , so that the storage required for the experimental evaluation described
depends on the dimensions of w(i), xCi) in the next section .
and on the record length.
1. With Xo and [w(i)] obtained from the 1 . With Xo and [w(i)] obtained from the
preceding iteration (or an initial guess) , preceding iteration (or an initial guess) ,
compute a nominal trajectory using (11) , compute a nominal trajectory using (11 ) ,
evaluate the residual sequence [veil] using evaluate the residual sequence [veil] using
(12) and the performance me asure using (13) . (12) and the performance measure using (13) .
2 . Solve the forward covariance filter , 2 . Solve the backwa r d information filter ,
consisting of a time update: consisting of a measurement update:
where where
3. Solve the backward smoother : 3. Perform the Newton- Raphson compu tation:
a
No . 1. Flight testa No. 2 Accident
Jacobian matrix NF=2l, NW=9, NX=39 , NH=2l NF=13, NW=6 , NX=19,NH=7
Nonlinear model Linear model Nonlinear model Linear model
~N\.J -
f
w
[ nFI.J 1JF
Jt
_ NX _
t
f
x
[ 1+NFX NF 61; 5.94 s 43 ; 2 . 20 s
- NX -
t
[ NHX NH
1~
100; 0.47 s 12; 0 . 10 s 32 ; 0 . 20 s
a The first number in each column is the number of nonz e ro elements in each of th e Jacobian
matrices (NFW, NFX, NHX); the times shown are computation times associated with the
particular Jacobian matrix (per iteration).
bAll unity.
2Meta l-foil flight recorders, used on many Fig . 3. Roll and pitch estimation with
airliners and some military aircraft , are on-board measurements and
designed to withstand severe accidents . aircraft performance data .
A Mathemati ca l Model f or Eff i c i ent Es t imation of Aircraft Mot i ons 1 161
(a l , am' an) and (d x , d y ' d z ) . Hence , the The rigid - body aircraft model described in
dimensions for this problem are NF = 13 , this paper has been shown to be a " natural"
m, = 6, ' and NH = 7 . The number of param- formulation for a linearization coordinate
eters was chosen to be NX = 19 . The results transformation . Although not all nonlinear
of this experiment, shown in Table 2 , again systems are transformable, there are probably
indicate that computational savings are many other areas yet to be discovered for
obtained by using the transformed plant; in which the method can be successfully used.
this case, the savings are greater than 30%.
As in the previous experiment, each version
of the algorithm was equally robust and REFERENCES
provided similar fits to the data.
Bach, R.E., Jr . (1980) . A Variational
Technique for Smoothing Flight-Test and
CONCLUDING REMARKS Accident Data . AlAA Atmospheric Flight
Mechanics Conference , Danvers , Mass .,
This paper has shown that the usual formu- Aug . (to appear in AlAA Journal of
lation of the rigid-body equations of motion Aircraft , Aug. 1982).
used for representing an aircraft can be Bach , R. E. , and R.C. Wing rove (1980) .
readily transformed to a plant with a linear Equations for Determining Aircraft
state model . The potential advantage of this Motions from Accident Data. NASA
formulation when used with an algorithm to TM- 78,609.
solve a state-estimation problem is computa- Brockett, E. W. (1978). Feedback Invariants
tional, and depends on the fact that the for Nonlinear Systems. IFAC Congress,
state-Jacobian matrix is constant along any Helsinki .
trajectory . A recently implemented smooth- Brunovsky , P . (1970). A classification of
ing algorithm was chosen as a test-bed for linear controllable systems . Kibernetika
a performance comparison of the two plant (Praka) , 6, 173- 187.
formulations . Two experiments were described Et kin , B. (1972) . Dynamics of Atmospheric
in which the smoothing algorithm was used to Flight. Chaps. 4 and 5 . Wiley , New
estimate aircraft motions along a simulated York .
trajectory. The first experiment , typical Hansen, R. S . (1978). Discrete Extended
of a flight-test data consistency a nalysis , Kalman Filter/Smoother Program for Air-
indicated a savings of more than 25% i n craft and Rotorcraft Data Consistency .
computation time with the linear model . DEFKIS User ' s Guide , NASI-14549 , Systems
Similar results were reported in a second Control, Inc., Palo Alto , Calif ., Jan .
experiment that was typical of an aircraft Hunt , L . R. , and R. Su (1982a) . Global trans -
accident analysis . In that case the compu- formations on nonlinear systems .
tational savings were better than 30% . Submitted to IEEE Trans . Autom . Control .
Hunt , L.R . , and R. Su (1982b). Linear
The design of efficient algorithms for air- Equivalents of Nonlinear Time Varying
craft state estimation is especially impor- Systems . Proceedings of 1981 Inter-
tant in a real- time environment because of national Symposium on Mathematical Theory
word-length and cycle~time limitations of of Networks and Systems (to appear).
flight computers. Choice of a postflight Klein, V., and J . R. Schiess (1977). Compat-
smoothing algorithm as a vehicle for ibility Check of Measured Aircraft
comparing the rigid-body models was one of Responses Using Kinematic Equations and
convenience and expediency: the algorithm Extended Kalman Filter. NASA TN D-3514 .
was available and its performance was well t1yer , G. (1981). The Design of Exact
understood . It should be emphasized , Nonlinear Model Followers . JACC ,
however, that the experimental results of Charlottesville, Va . , June .
that comparison apply directly to the real- Stalford, H. L. (1980). The EBM System
time filtering problem . In fact , implemen- Identification Technique and Its Applica-
tation of the linear-state concept with a tion to High a/B Modeling of Aircraft .
real-time filtering algorithm is currently AlAA Atmospheric Flight Mechanics
under study. For an example of such an Conference, Danvers , Mass ., Aug .
application, note that the plant shown in Wingrove, R. C., R. E. Bach , and E. K. Parks
Fig . 3 could also serve as a model for on- (1979) . Aircraft Motion Analysis Using
board, real-time estimation of roll and Limited Flight and Radar Data . Presented
pitch angles. Such estimates could replace at 10th Annual Symposium of the Society of
or supplement corresponding gyroscopic Flight Test Engineers, Las Vegas, Nev . ,
measurements . Alternatively , real-time Sept.
estimates of roll and pitch angles could
be made by adding accelerometer measurements
(ax, ay); the aerodynamic "pseudomeasure-
ments" (a , B) would then not be necessa r y .