Professional Documents
Culture Documents
Introduction
Currently, a large amount of effort is being devoted to
the development of commercial aircraft avionic systems,
Optimal Aircraft and, in particular, to solving the all-weather landing
problem. A few systems exist today that have demonstrated
Go-Around and Flare a fully automatic landing capability, and others have shown
a similar capability through laboratory simulation; yet, they
Maneuvers have received only limited acceptance from airline pilots
and from the Federal Aviation Agency. It also appears that
the more sophisticated and redundant systems presently
GLENN BUELL under development may receive the same lack of
Rockwell International acceptance.
Anaheim, Calif. 92803 One approach to instilling faith in a fully automatic
C,T. LEONDES, Fellow, IEEE landing system is to monitor the total vehicle performance
University of California during the approach and landing phase and to furnish the
Los Angeles, Calif. pilot with a decision making display advising him of the
current status of the approach. This represents a significant
departure from present day failure monitoring, which is
limited to examining a particular system or subsystem for
Abstract failures only. Although failure monitoring is necessary, it is
not sufficient. A hostile environment (wind shears, bent
This paper analyzes in detail two of the critical aircraft maneuvers ILS beams, etc.) can be just as detrimental to a successful
associated with approach and landing: the go-around maneuver and landing as a system failure and, conversely, some system
the flare maneuver. Optimal solutions that include state and control failures may not imply an aborted approach.
variable constraints are obtained for both problems. Two algorithms Basic to the development of a vehicle performance
are given for computation of the minimum and maximum altitude monitor is the need to continuously compute the height
loss associated with the pilot-controlled go-around maneuver. A above the terrain at which the landing is committed. A
matrix operator is obtained that can be used for in-flight second requirement is the capability of continuously
computation of the altitude loss on a small general-purpose digital predicting the aircraft longitudinal impact point under the
computer.
influence of known closed-loop guidance laws. Presented
The flare optimization presented is for a cost functional that
here are solutions to these two requirements that are
includes both the longitudinal touchdown dispersion and the normal
amenable to a small general- purpose digital computer.
By converting the classical aircraft longitudinal
acceleration. A closed-loop mechanization is given that
perturbation equations of motion into state space notation,
approximates the optimal trajectory. A second matrix operator and assuming a time polynomial forcing function, the
which can be used for prediction of the longitudinal touchdown convolution integral is used to obtain the go-around
point is obtained. Uncertainties are also obtained for the purpose of altitude loss for optimal and suboptimal control policies.
establishing a prediction confidence level. The optimal control policy is obtained using the maximum
It is proposed that these prediction techniques should be principle with state and control variable constraints, while
incorporated into a decision making performance monitor. This the suboptimal policy is representative of pilot control
monitor could provide the pilot with a continuous assessment of the actions in a noncritical environment. These solutions are
approach and could generate a preflare decision on whether or not then approximated by a Taylor series expansion, and the
to commit the aircraft to the flare maneuver. resulting elements are averaged to form a matrix operator
that can be used for computing the expected altitude loss.
Because the flare maneuver is a closed-loop control
function, it is first necessary to define a specific system
before the desired matrix operator can be obtained.
Although system optimization is somewhat arbitrary in that
a variety of choices exist for the cost functional, it still
presents a valid rationale for selecting a system. The cost
functional chosen here is the weighted sum of the 'miss-
distance' from a desired touchdown point and the squared
normal acceleration during the maneuver. The maximum
principle is used to demonstrate optimality, and the
Newton-Raphson method is used to secure convergence. It
Manuscript received May 8, 1972; revised June 28, 1972 is shown that the resulting optimal trajectories are closely
This work was supported, in part, by Air Force Office of Scientific approximated by a linear system that employs an
Research Grant 699-67. exponential flare control law summed with the normal
280 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. AES 9, NO.2 MARCH 1973
By using equation set (2), expanding and retaining only
first-order terms in the perturbed variables, (3) can be
reduced to
h(t) = ao + a1x1 (t) + a2X2(t) + a3x3(t) (4)
with the coefficients ai obtained from trim conditions.
Fig. 1. Longitudinal velocity profile.
Thus, if time solutions for X(t) can be obtained in closed
form, then h(t) follows immediately and h(t) is obtained
through a straightforward analytical integration of h(t).
acceleration. This system is then used to obtain the flare Now consider the first element of the control vector. In
matrix operator in a manner similar to that used for the the case of a go-around maneuver, it is valid to assume that
go-around. the throttle action of the pilot will be similar to a step
input to maximum allowable thrust. Because of the delay
11. Equations of motion
due to engine dynamics, this may be modeled as
1 1 0 0
g11 -g21
g21 gll I
10
1~~ 0
0 °
1 -i i 0O P= [W] ----- .
L-1 =-
21 31 -941
0 0 :1
0 0 1 1 0 0 1 g41 g31
o o I -i i
1l2 -922 1 ° °
i=-f.
g22 g12 1 0 0
Then, by combining (11) and (13), the equation of sta[te in Q=-[W]
the Z space is
0 g32 -g42
Z= [L-1 ALI Z + [L-1 T1] BU. (14) ° |0° g42 g321
If the eigenvalues of A are X=a ± ib, c ± id, then
and introducing the state transition vector
a -b j0
b a I O ol 1 (t)
L-1AL = ['4
;-_
02 1(t)
O O c -d t)
0° ° d c 03 3(t)
which is a block diagonal matrix of real num bers. 0( 3[
4 J
Furthermore, the matrix W,
W=TL# ZW L=K' T-, (15) (16) reduces to (for to = 0)
282 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS MARCH 1973
~~~- 8eMAX
X(t) = R.p(t) + P $(t - r)u (T) dt bOo+ SeMINt 0Ot Lt,
Jo 8(t)=
co
I '
I tM1N , tj4t 6tf
+Q At
0t- ) U2(r) dr (19)
Iof
-- 8eMIN
Fig. 2. Optimal go-around elevator control policy, no state
where ul and u2 are scalars and Ro is the initial value variable constraints.
matrix given by
Zl -1-Z2 I
I
0 0
6 TANGENT
_
2
where Kis a 4 X 4 matrix of constants obtained by evaluating
the time functions of Equation (26) at ti. Now, if final
values of any three functions of X(tj) are desired,' then
the necessary control coefficients may be computed. As an U
16
a
example, let the elements of K be k1,,i, I = 1,2,3,4, and let
the desired final value functions be xi(t, ) = xi1, i = 1,2,3.
Then
I
604
INITIATE
xi0 -kl,1
1.
kl,3 kl,4 -1 N
"PI
GO-AROUND
Et, 50 *
lo 2,1 N
k2,3 k2,4 x -k N
I
(28)
rA 40
30 _
0--
and the remaining value of x4(t,) is computed from (27). TANGENT
Because h(t) given in (4) is a linear function of X(t), it 20 t I a~ ~ ~ ~ ~ ~ ~ ~ ~-
-3 -2 -1 0 1
may also be expressed as a combination of simple time TIME - SEC
2 3
functions and the control coefficients. This, in turn, allows Fig. 3. Correlation with flight test data.
h(t) to be easily integrated, yielding h(t) and, for any
desired time t1,
time until the constraint relationships of (22) are satisfied
[hz(tsij [ j (29)
over the closed interval 10,
is then
tt
. The resulting altitude loss
hIFi
Ah = h(tf) -h(to) (31)
where L is a 2 X 4 matrix of constants similar to K. with h(tf) computed from (30).
Combining (27) and (29) gives Obtaining an optimum solution is valid only if the pilot
can be expected to perform in an optimum manner every
Xi (t1) time. Unfortunately, such is not the case, and some type of
upper bound on the expected altitude loss is required.
x2 (t,) Consider, as a second algorithm, using the same polynomial
X3(t1) = M , (30) elevator control law, but selecting the control coefficients
X4(ti) such that
X2 (tf) = mx - k3
h(tl)
h(t,) L X2 (tf) = 0 (32)
where h(tf) = O.
K This algorithm, herein called the tangent method, places the
[M]=EH
L
trajectory tangent to the state variable boundary
simultaneous with achieving h = 0. This method was
selected primarily because it is similar to the way a pilot
Now that the method for computing the desired control performs a go-around maneuver if the situation is not
coefficients has been developed, it only remains to select critical. He desires to keep the maneuver as smooth as
the proper value for tl. This is accomplished by an iteration possible while arresting the sink rate. Obviously, some
procedure that starts with a small value for t, and steps initial conditions do not require aT(tf) = amax for a
solution. This may be easily accommodated by eliminating
1 This is the more general case. If one control coefficient is
the first relationship from (32), setting Vm 0, and solving
known, then only two functions of X(t, ) are required to yield the for a/ and X. The solutions are called degenerate tangent.
two unknown coefficients. Fig. 3 presents a comparison of the two algorithms with
284 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS MARCH 1973
one set of DC-9 flight test data. This same agreement was
obtained on all DC-9 flight test data evaluated, provided
pilot pitch and thrust commands were simultaneous.
The two algorithms presented require the use of a large
general-purpose digital computer to obtain solutions.
Obviously, this is unacceptable for airborne use, and a X(t f) .
simplified means of accurately approximating the desired ,ft, -f
O O O
0
2 4
°
aX4|
f
la= °
If
VT(t, cos TOrt dt -VX [I + xi (t)] dt. O
If
(37)
with all elements of N evaluated at XO = 0. The algorithms Also, the vehicle normal acceleration may be approximated
are used to generate two values for each element of N. By by
averaging these values, the matrix N is developed that
predicts the best estimate for Ah, rather than a minimum or an (t) -i2 (t) X4 (t)- (38)
a maximum. To investigate the accuracy of the Taylor
series expansion, 300 computer runs were made with
As before, define an xn+l member of the state vector as
uniformly distributed random initial conditions. In all
cases, the second-order series expansion had less than a 5
percent error with a near zero mean. As a comparison, a
n =f =
a d' J an()] t
(39)
first-order series expansion was also considered. The errors
and define
I
encountered in this case were as large as 50 percent, and
thus unacceptable. Vx
It is therefore concluded that by precomputing the 01 = - [1 +x1(t)] dt- 1
elements of N as described, (34) presents an acceptable
sgn
J-d
solution to the on-board estimation of the go-around
altitude loss for any arbitrary value of X(to). V 2
92 =P
(40)
32.2
V. Computation of the Longitudinal Touchdown Point
Then xn becomes
As previously stated, it is necessary to define a specific
closed-loop flare mechanization before the flare prediction xn+ 1 (t)= 31(t)[1l+XI(t)J +02 [X2 (t)-X4 (t)] 2. (41)
BUELL ET AL.: AIRCRAFT GO-AROUND & FLARE MANEUVERS
285
DESIRED XTD = 9001
2
u s 5
, v = 100
Fig. 5. Op timal flare plots of altitude versus ground distance. Fig. 6. Closed-loop flare mechanization.
4 30
25 <AOPTIMAL v
U2 (t) - s(t) 25
2)
202b2 2 2
b222
i=1
E (a2ixi(t)) C 20 - CLOSED LOOP SYSTEM
E 15
10
.X4(t) + ktb2l t(l e-ot) + uI 0 b2l
N
(42)
5 _ <
with a2i and b2i the elements of the description matrixes A 0 200 400 600 800 1000 1200 1400
and B, and s(t) given by HORIZONTAL DISTANCE FROM FLARE INITIATION FEET
Fig. 7. Comparison of optimal and closed-loop trajectories.
4
s(t) = b2iPi(t) (43)
i=1
with pi(t) the elements of the adjoint vector p(t). h(t) + 4.5 h(t) + 9.0 = 0. (45)
Because the optimum control law is a combination of
the function s(t), the state vector, and time, an iterative
A closed-loop system could be developed that uses (45) as
numerical integration of the optimality equations is the elevator command, but this would not be sufficient to
required. A digital program was written to solve thisapproximate an optimal trajectory for a given Consider a v.
problem using the Newton-Raphson method described in second feed-forward loop that uses normal acceleration. If
[6]. Fig. 5 presents the resulting optimal flare trajectories
the loop gain is adjusted to correspond to the desired value
for various values of v. of 1v, then the sum of the two loops provides the control
However, such an optimal solution has a limited value
law for approximating the desired optimal trajectory. Fig. 6
unless it can be easily applied to a physical system.presents a block diagram for such a system. The altitude
On-board implementation of this optimization process for a
rate h, is obtained from complimentary filtering of the
commercial aircraft would require a large, high-speed digital
altitude and the measured normal acceleration, while the
computer, and thus be cost prohibitive. By examining the
normal acceleration loop uses the measured normal
trajectories of Fig. 5, it is seen that they are of an
acceleration fed through a bandpass filter. The integrator in
exponential form, that is the forward loop is used to eliminate steady-state errors,
and k converts from feet to degrees elevator command 6e .
h(t) + ah(t) + b = 0. (44)
Fig. 7 presents a comparison between the closed-loop
However, with boundary conditions of system and the optimal trajectory for v = 3.5. As may be
seen, the closed-loop mechanization presents a good
h(to) = 36 ft, h(to) = -10 ft/s approximation to the desired trajectory.
Because the system input variables are members of a
= o, A(rt) = -2 ft/s,
h(tf) state vector controlled by an explicit guidance scheme (and
assuming no random disturbances), it is proposed that the
the coefficients a and b may be solved for, and (44) air distance traveled during the flare maneuver is a function
becomes of only the initial conditions X(to). Further, the flare
286 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS MARCH 1973
3
Xl
\+sK 2
-DEG -
REGION I 1 REGION II
l 1. 36
xi-DEG -2 -1 1 2 3
-1 AY-DEG
/ 3.0 \\ REGION SWITCHING
-2 FUNCTION
\ LINEAR
2.5
-3 L V\
APPROXIMATION
-1. 0
.-1.5
-2.0
-2.5
-3.0
I1 -150
200'
/ ..
4C
.~~~~~~~~~~ -.
. . I. .
I.1 -3 -2 -1
a _-- I
1
I
2
r
L
3
-200: *.. 1 200
.
-400 . 400
i Xi-DEG
I :
1
I*I-_ - I
/ PREDICTION ERROR
- FEET -
I
XSHORT/1 . 2- \/V XLONG I -2 L
l____
I ___
-2 ; : .I
.3
XLONG
3
AY
Ay-DrEG -DEG-
I 3_ -_100'1 .__ -__
.1 jI 2 - -150'.------
- -200t__-_-
_
illl r--250'- - -
-3 *-I
-2 '. l L -InfI.
illi t?-,O V0--1 i
II 1 2
350' 3 3
t _ , I . ..I 1-I -I
x -DEG
~~ ~ ~ ~ ~
-400 / -200 * : .- 200 400
I'1 L _J
<- . . - . PREDICTION ERROR lII I L--J
.. III2
1
- FEET-
.
.
N- 11-
11-
.
-2 . . . X ____
..
IL _________
L_3
XSHORT N *
_3
- .. ; 'I--XLONG XSHORT
Fig. 10. Summary of flare prediction errors; xi Fig. 11. Flare uncertainty profiles;
variation, A'y variation. Xl ;Xshort variation.
represent long and short tolerances; that is, the prediction VI. Summary and Conclusions
XTD could be longer by as much as Xiong or shorter by This paper has presented a detailed analysis of two
xshort critical aircraft maneuvers. Cost functionals have been
To generate this touchdown "window," 300 flare defined and the associated optimal trajectories obtained.
trajectories were generated using random inputs as initial For the go-around maneuver, the optimal solution is shown
conditions. As in the case of the go-around maneuver, the to involve bounded control and state variables. A
IBM random number generator was used with X(tO) limited polynomial forcing function is proposed that satisfies all
to ± 3 degrees. Fig. 10 presents a summary of these data. constraints. Computational algorithmo are presented for
The dashed lines represent uncertainty limits as a function solving both optimal and suboptimal trajectories. These
of xl(to) or A'y(to). These uncertainties may then be solutions are approximated by a series expansion, and
plotted as shown in Fig. 11. Because the uncertainty formulated as a matrix operator for on-board prediction of
profiles are composed of straight lines, they are easily the go-around altitude loss.
programmed. Thus, the total flare prediction technique is as The optimal flare maneuver is shown to be free of
follows: bounded constraints. The nature of the optimal trajectories
1) Compute the constant matrixes N, and N2. leads to a closed-loop mechanization that uses an
2) Select the region from the switching function. exponential flare law and the normal acceleration. This
3) Compute XTD from (48). closed-loop system is the basis for a flare matrix operator
4) Compute the uncertainties Xiong and Xshort which can be used to obtain on-board predictions of the
aircraft longitudinal touchdown point. Associated
This is the proposed flare prediction technique suggested uncertainties for these predictions are presented.
for on-board implementation. Note that step 1 would be It is concluded that a digital logic network using the
precomputed and stored, as would the switching and go-around altitude loss, the predicted touchdown point, the
uncertainty functions of steps 2 and 4. touchdown uncertainties, and other relevant performance
288 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS MARCH 1973
criteria can provide the pilot with a valuable preflare [3] G.D. Buell, Jr., "On the prediction and optimality of aircraft
assessment before he commits the aircraft to the flare maneuvers associated with approach and landing," Ph.D.
maneuver. dissertation, University of California at Los Angeles, 1969.
[4] L.S. Pontryagin, "Some mathematical problems arising in
connection with the theory of optimal systems of automatic
control," Proc. Acad Sci. (USSR), October 1956.
References [5] L.I. Rozoneor, "L.S. Pontryagin's maximum principles in the
[11 J.H. Blackelock, Automatic Control of Aircraft and Missiles. theory of optimal systems," Automat. Remote Control (USSR),
vol. 22, October, November, and December 1959.
New York: Wiley, 1965. [6] E. Volganau, "Applications of variational methods to missile
[21 C.D. Perkins and R.E. Hage, Airplane Performance, Stability, and space vehicle guidance," University of California at Los
and Control. New York: Wiley, 1949. Angeles, Rept. 66-11, March 1966.
Glenn D. Buell, Jr. received the B.S. degree from Texas A & M University in 1956, the
M.S. degree from the University of Southern California in 1965, and the Ph.D. degree
from the University of California at Los Angeles in 1969.
He spent many years in the commercial avionics field with Douglas Aircraft Company,
Long Beach, Calif. He is currently the Program Development Engineer for Ship Dynamics
and Control at the Autonetics Division of Rockwell International, Anaheim, Calif. He is
also a former U.S. Air Force pilot with almost 2000 hours in a variety of single- and
multi-engine prop and jet aircraft. He has presented several technical papers.