172 J. GUIDANCE AND CONTROL VOL. 3, NO.

2
ARTICLE NO. 78-1307R

Optimizing the Gains of the

Baro-Inertial Vertical Channel
William S. Widnall*
Massachusetts Institute of Technology, Cambridge, Mass.
and
Prasun K. Sinhat
Intermetrics, Inc., Cambridge, Mass.

The selection of the three gains in the baro-inertial vertical channel has been formulated as a stochastic op-
timal control problem, where the objective is to minimize the mean-square error of the indicated vertical
velocity. The optimal set of gains is surprisingly different from a conventional set of gains, and it provides a
Downloaded by BEIHANG UNIVERSITY on November 15, 2019 | http://arc.aiaa.org | DOI: 10.2514/3.55966

significant performance improvement. Sensitivity of the results to the statistical assumptions is explored. Ap-
proximate analytical formulas are presented giving the optimal gains and pole locations as a function of the
assumed statistics of the sources of error. A time domain simulation also exhibits the performance improvement.

Introduction where h is^the indicated altitude, vz is the indicated vertical

velocity, 6cr is the computed vertical acceleration error, fz is
T HE first aircraft inertial navigation systems were
primarily two-channel systems that provided horizontal
navigation data.J-3 Inertial navigators instrumenting three
the measured vertical specific force, g is the magnitude of
gravity computed as a function of indicated altitude and
channels were introduced for missile navigation and guidance. latitude, and klt k2, k3 are the loop gains. This third-order
In addition, the value of inertially derived vertical velocity vertical channel mechanization is superior to a second-order
was recognized in aircraft applications involving flight path mechanization, which omits the £a equation, because it has
angle determination and precision weapon delivery. It is well zero vertical velocity error due to any bias vertical ac-
known that the altitude channel of a pure inertial navigation celeration error such as accelerometer bias or gravity com-
mechanization, in which gravity magnitude is computed as a putation error.
function of the indicated altitude, is unstable.4-6 Near the The Litton CAINS (Carrier Aircraft Inertial Navigation
surface of the Earth, the time constant of this exponential System) implements such a third-order baro-inertial vertical
instability is about 10 min. Hence, for typical cruise channel. In the CAINS, the three gains have been chosen so
navigation durations, the vertical channel of a terrestrial that the characteristic equation of the errors has a triple pole
inertial navigator must be stabilized by some external altitude at the complex frequency s= - 1/r, where T is the desired time
reference. constant. For such a triple pole placement, it can be shown
The most commonly used external altitude reference in one chooses the gains to be
aircraft is a barometric altimeter. The optimal time-varying
combination of the inertial and barometric data may be A:, =3/7 k2=3/T2+2g/R k3=l/r3 (2)
obtained using a Kalman filter.7 However, in applications not
demanding the minimum navigation errors or in which the where R is the geocentric radius. In the CAINS, the time
computer capacity is severely limited, a simple mechanization constant has been chosen to be T = 100 s. We have no literature
is commonly used in which the difference between the in- explaining the designer's choice of the triple pole and its
dicated and barometric altitude is fed back through constant location. Perhaps the triple pole configuration was arbitrarily
gains or simple transfer functions to stabilize the altitude selected to reduce the gain-setting problem from three
navigation variables.8 A typical set of constant-gain baro- parameters (kl9 k2> k3) to one parameter (T). Speculating
inertial mechanization equations, which is analyzed in detail further, perhaps the time constant of 100 s was an order-of-
in this paper, is magnitude choice, selected so as to be faster than the 571 s
time constant of the pure-inertial vertical-channel instability
yet slower than the typical barometric error fluctuations
associated with short-term aircraft maneuvers. This choice
vz=fz-g(h,L) + Coriolis terms-fc2 (/*-/*,,) -da (Ib) would be expected to both stabilize the vertical channel and
provide some smoothing of the barometric altimeter errors.
ba=k3(h-hb) (Ic) Regardless of the reasoning, the CAINS has peformed well in
its intended applications. We shall refer to the CAINS set of
Received June 26, 1978; presented as Paper 78-1307 at the AIAA gains, given by Eq. (2) with r = 100 s, as the baseline set.
Guidance and Control Conference, Palo Alto, Calif., Aug. 7-9, 1978; Some applications have more demanding vertical velocity
revision received July 3, 1979. Copyright © 1979 by W.S. Widnall requirements than the CAINS requirements. In such ap-
and P.K. Sinha. Published by the American Institute of Aeronautics plications it may be necessary to optimize the vertical channel
and Astronautics with permission. Reprints of this article may be gains to reduce the vertical velocity errors. One such ap-
ordered from AIAA Special Publications, 1290 Avenue of the plication was the use of the Magnavox X-set GPS navigator in
Americas, New York, N.Y. 10019. Order by Article No. at top of the demonstration of pinpoint bombing on a target whose
page. Member price $2.00 each, nonmember, $3.00 each. Remittance
must accompany order. absolute coordinates were known. The X-set GPS navigator
Index category: Guidance and Control. includes a barometric-inertial navigation subsystem and a
*Associate Professor, Dept. of Aeronautics and Astronautics. GPS X-set receiver whose outputs are combined by a Kalman
Associate Fellow AIAA. filter. An error analysis by Ausman9 predicted that the two
tSenior Engineer, Navigation and Analysis Dept. largest contributions to bomb miss distance would be due to
 MARCH-APRIL 1980 OPTIMUM GAINS FOR A BARO-INERTIAL VERTICAL CHANNEL 173

the altitude navigation error and the vertical velocity

navigation error at bomb release. Flying at 450 knots, at 5000- w
al
ft altitude above ground level, with a level release, and using
low-drag bombs, Ausman indicated that the anticipated
altitude and vertical velocity navigation errors of 19 m and
0.75 m/s would contribute 24 m and 17 m, respectively, to
downrange bomb miss distance. The absolute altitude error of
the integrated navigation system is caused by the bias-like
errors in the GPS measurements. The choice of baro-inertial
gains has little or no effect on the absolute altitude accuracy.
However, the absolute vertical velocity errors are a noticeable
function of both the GPS measurement errors and the baro-
inertial errors. The integrated system vertical velocity errors
can be reduced if first the baro-inertial vertical velocity errors
are minimized.
There is no reason to assume that the triple pole gain set w
provides the best performance. Other pole placements might bl
provide superior performance. To obtain new insights into the Fig. 1 Baro-inertial vertical channel error model.
effect of the gains, we have formulated the vertical channel
gain setting problem as a parameter optimization problem for
Downloaded by BEIHANG UNIVERSITY on November 15, 2019 | http://arc.aiaa.org | DOI: 10.2514/3.55966

the control of a stationary stochastic process. This paper factor error and input axis misalignment. The error state db is
presents the optimization problem formulation, com- a random walk modeling any bias or slowly varying error in
putational results, analysis of the results, and a time-domain the altitude indicated by the barometric altimeter. Physical
simulation of the performance with the recommended gains. error sources include zero setting error, static pressure
Before proceeding, we comment that any constant gain set measurement error, variation in the height of a constant
(including our optimized set) will be less optimal than ap- pressure surface, and scale factor error due to nonstandard
propriate time-varying gains that take into account the day temperature. The white noise wb2 into the integrator
nonstationary nature of the inertial and barometric altimeter provides the random walk. The white noise ww models short
errors. The optimal time-varying combination of the inertial correlation time altimeter error, such as due to changes in the
and barometric data may be obtained using a Kalman filter. angle of attack or sideslip angle during a maneuver, or due to
Of course it is also possible to select effective time-varying altimeter quantization or other noise.
gains with non-Kalman approaches. One example of the use It is difficult to suggest appropriate values for the spectral
of time-varying gains is provided by Whalley.10 Whalley densities of the four independent white noises in this
points out the large error in barometric altitude when a stochastic model. Nevertheless, it must be done for the
supersonic aircraft passes through Mach 1 due to shock waves analysis to proceed. Table 1 shows the nominal values of the
moving past the static pressure port. Whalley suggests noise spectral densities that have been selected. These
eliminating this source of error by switching out the air data somewhat arbitrary numerical values have been arrived at by
automatically from say Mach 0.95 to Mach 1.1. This could be the following considerations .
mechanized by programming klt k2, k3 to be zero in this For the short correlation time acceleration error, a typical
Mach interval. amplitude could be 200 fig. This error could be caused by a
Another example of the use of time-varying gains is vertical accelerometer input axis misalignment of 200 ^ rad
provided by Ausman and associates.11'12 They note that in (41 arc-sec) together with a horizontal maneuver acceleration
-subsonic flight the largest source of barometric altimeter error of one g. The typical duration of a horizontal maneuver is
is often the scale factor error due to the atmosphere not assumed to be of the order of 60 s. For a repeated series of
having the standard-day temperature-vs-altitude profile. In random aircraft maneuvers, the autocorrelation function of
climbs and dives the scale factor error induces significant the acceleration error would have area approximately
vertical velocity error into a constant gain baro-inertial
vertical channel. Ausman and associates designed and im- / = (200xlO- 6 xlOms- 2 ) 2 x(60s) = - 3 (3)
plemented vertical channel mechanizations that reduce the
gains during climbs and dives, while observing the baro-bias The area of the autocorrelation function equals the low-
shift due to scale factor error. The estimated baro-bias shift is frequency value of the spectral density. For the white noise,
automatically subtracted from the baro-inertial error feed- whose autocorrelation function is a Dirac delta function with
back so that loop transients due to the scale factor error are area Qah the spectral density Qal applies at all frequencies. Of
minimized. interest is the response of the vertical channel at frequencies
Also before proceeding, we comment that additional ex- lower than the higher frequencies of the short-correlation
ternal data may be useful in reducing the effect of non- acceleration error. So the low-frequency density of Eq. (3) is
standard-day temperature. Blanchard13 proposes using in- used for the spectral density of the white noise.
flight measured temperature data, in place of the standard-
day lapse-rate assumption, to relate more accurately pressure
changes to altitude changes.
Table 1 Nominal values of noise spectral densities
Formulation of the Gain-Optimization Problem
The error model for the vertical channel is shown in Fig. 1. Noise density Noise density
White noise for symbol value
The positive feedback with gain 2g/R is the destabilizing
effect of normal gravity being calculated at the closed-loop Short correlation time 0.1 2 . 4 x l O - 4 m 2 s -3
altitude, which is in error by dh. The error state da is a random acceleration error
walk modeling any bias or slowly varying error in the vertical Acceleration error Qa2 1 . 0 x l O - 9 m 2 s -5
acceleration due to accelerometer bias, gravity anomaly, or random walk
error in the Coriolis terms. The white noise wa2 into the in- Short correlation time Qb, 100m2 s
tegration provides the random walk. The white noise wal altimeter error
models short correlation time acceleration error, such as the Altimeter error random walk Qu 100m2 s' 1
effect during a short maneuver of vertical accelerometer scale
 174 W.S. WIDNALL AND P.K. SINHA J. GUIDANCE AND CONTROL

The acceleration error random walk models the slowly where near the surface of the Earth c=2g/R = 3.07 x
varying error due to accelerometer bias shifts, changes in the 10~6 s ~ 2 . For the steady-state solution to exist and be equal to
gravity anomaly, and changes in the error in the Coriolis Eq. (9), the set of loop gains must yield a stable system.
terms. If over a period of 1000 s, the accelerometer bias were Therefore, the permissible values of the gains are in the
expected to shift 100 fig, the appropriate noise spectral density regions defined by
for the random walk is
kj>0, k2-c>0, k3>0, k1(k2-c)-k3>0 (10)
Qa2 = (lOOx 10~ 6 x 10ms -2)2/(1000s) = l.Ox 10~ 9 m 2 s ~5
(4) The explicit computation of the mean-squared vertical
velocity error is used in a computer program that seeks a set of
For the short correlation time altimeter error, it is assumed gain values that minimizes the mean-squared error. The
that repeated random fluctuations of the order of 10 m may pattern search algorithm of Hooke and Jeeves15 has been
be present in the baro-indicated altitude, and that these errors utilized. The algorithm does not require explicit gradient
persist for correlation times of the order of 1 s. To match the information.
low-frequency spectral density of this error, the white noise The natural frequencies (poles) of the closed-loop portion
error model should have density of the baro-inertial vertical channel are the three roots of the
characteristic equation
e w =(10m) 2 x(ls)=100m 2 s (5)
s3 + kjs2 + (k2 -c)s+k3 = 0 (11)
The altimeter error random walk models the slowly varying
Downloaded by BEIHANG UNIVERSITY on November 15, 2019 | http://arc.aiaa.org | DOI: 10.2514/3.55966

error due to: changes in the static pressure measurement error For a candidate set of gains, it is often of interest to inspect
(due to speed changes), variations in the height of a constant the locations of the three poles plt p2, p3. In such a case, the
pressure surface (the weather pattern of highs and lows), and roots of the cubic Eq. (11) are computed according to the
scale factor error (related to nonstandard-day temperature known formulas for those roots. When the time constant of a
and nonzero aircraft climb or descent rate). For an at- pole is mentioned, it is defined to be the inverse of the real
mospheric scale factor error of 3% and an aircraft climb or part of the complex frequency of the pole.
descent rate of 33 m/s (6500 ft/min), the error rate of the
baro-indicated altitude is 1.0 m/s. If the climb or descent Optimization Results
continues for 100 s, the change in altimeter error will be 100 To provide a baseline design and performance against
m. Assume that the aircraft trajectory is characterized by a which to compare the optimized performance, the mean-
random sequence of such climbs and descents. The ap- squared velocity error is evaluated for the set of gains, Eq. (2),
propriate noise spectral density for the random walk is then which place a triple pole at T = 100 s
e w =(100m) 2 /(100s)=100m 2 s- 1 (6) A:7=3.0xlO~2s k2 = 3.0307 x 10 - 4 s

The mean-squared error of the indicated vertical velocity

has been selected as an appropriate performance index. Note (12)
that with the random walk error models for acceleration error
and for altimeter error, only the vertical velocity error has a The mean-squared velocity error, with the nominal (Table
stationary and finite mean-square value. All other error states 1) values of noise spectral densities, is found to be
have mean-square values that grow unbounded with time.
Referring to Fig 1, 60 tracks 5a+ (2g/R)5h as this sum (6v)2 =0.818 m 2 s ~ 2 = (0.904 m/s)2 (13)
wanders off, and dh tracks db as it wanders off.
The mean-squared vertical velocity error may be computed Using the Hooke and Jeeves pattern search procedure, the
as an explicit function of the input noise spectral densities and gains that minimize the mean-square velocity error, with the
of the loop gains. One first calculates four transfer functions nominal noise densities , are found to be
Hf(s) relating the vertical velocity error response to each
independent white noise input. The power spectrum of the 7 = 1.003s-1 A: 2 =4.17xlO- 3 s- 2
vertical velocity error is then
(14)

(7) The corresponding value of the mean-square velocity error is

(dv)2 =0.418 m 2 s ~ 2 = (0.647 m/s)2 (15)
where Q, is the spectral density of the rth white noise. The
mean-square value of the vertical velocity error is the integral This is a significant performance improvement relative to the
of the power spectrum baseline case. The rms velocity error is 30% lower.
The three poles associated with the gain set, Eq. (14), are

wr2=E^-. (8)
located at
Pj = -0.998s-1 p2,p3= -2.082xl0^ 3 ±/2.34x 10~4 S"1
The four integrals are evaluated using an appropriate table of (16)
integrals.14 The result is
They have time constants of

2[kI(k2-c)-k}\ T7 = 1.002s 480.3s (17)

2[k1(k2-c)-k3] The optimized gains and resulting pole placements (and time
+ (k3+ck1)2k1]Qb2 constants) are radically different from the baseline triple pole
(9) set. One time constant is a factor of 100 faster; the other two
2k3[k1(k2-c)-k3] time constants are a factor of 5 slower.
 MARCH-APRIL 1980 OPTIMUM GAINS FOR A BARO-INERTIAL VERTICAL CHANNEL 175

Table 2 Contributions to mean-square velocity error of order vertical channel are that of a fast first-order loop nested
nominal noise densities inside a slower second-order system. From Table 3 it is clear
that the fast pole frequency is simply related to the first gain
Mean-square velocity error , (m/s)2
Noise Triple pole, Pi~-kt (18)
density r=100s Optimized set
With such a fast real pole, the characteristic equation of the
Qa, 0.018 0.0291 third-order vertical channel can be factored as
0.00175 0.0275
c£ 0.00018
0.798
0.00087
0.361 (19)
Kj
2 2
Total 0.818 = (0.904) 0.418 = (0.647)
where

The individual contributions of the various white noises to (20)

the mean-square velocity error are shown in Table 2, for the
nominal values of the white noise spectral densities. The data This can be shown by multiplying the two factors in Eq. (19)
of this table show that the mean-square velocity error is to obtain
dominated by the altimeter error random walk (Qb2), while
Downloaded by BEIHANG UNIVERSITY on November 15, 2019 | http://arc.aiaa.org | DOI: 10.2514/3.55966

the contribution of short correlation time altimeter error

(Cw) is least.
To obtain further insight into the nature of the optimal
solution and to exhibit the sensitivity of the optimal solution The correct characteristic equation of the third-order vertical
to the noise density assumptions, the optimal solution has channel, Eq. (1 1), in terms of k'2 is
been computed for various values of the noise spectral
densities. Table 3 shows the results for four cases in which one (22)
of the noise densities is increased while holding the other three
densities at their nominal (Table 1) values. Comparing Eqs. (21) and (22), it is clear that sufficient
With the optimized gains being so different from the conditions for the factorization of Eq. (19) to be true are
baseline gains, it is interesting to ask: for what set of input
noise densities are the "triple pole" gains optimal? From k'2<k] (23)
Table 2, one notes that in the "triple pole" case, the con-
tributions of the altimeter error noises to the mean-square (24)
velocity error would be more nearly equal if Qbl were 10 times
larger and Qb2 were 100 times smaller. The optimization Inspecting the second-order factor of Eq. (19), one sees that
program has been rerun with these altered values for noise the two slower poles are a functipn only of the two gain ratios
density. The results are presented as the last case in Table 3. k2/kj and k3/kt. The pole locations are
These results demonstrate that the baseline triple pole set is
close to being an optimal set if the random walk component
of the altimeter error is significantly smaller and if the short
correlation time altimeter error is somewhat larger than the
nominal assumed values. Equation (25) gives good agreement with the values of poles
and gains in the computed results (Table 3).
Analysis of Results The optimized gain kl in the nominal noise density case has
In all cases presented in the previous section (except the a very curious numerical value of unity. This provides a clue
greatly reduced Qb2 case) the dynamics of the optimal third- that there exists a very simple relationship between this gain

Table 3 Changes in optimal solution with altered noise densities

Optimal gains: kl, s ~ *, k2, s ~ 2 , k3, s -3 Pole locations, s" Time constants, s
Nominal case 1.003 -0.998 1.002
(Table 1 values) 4.17X10- 3 -2.082x10-3 480.3
4.39xlO~ 6 ±j 2.34 x 10 ~ 4 480.3
Increased Qbl 0.449 -0.445 2.2
(5 nominal) 1.86X10"3 -2.084x10-3 479.9
1.96X10-6 ±j 2.6 X 10 ~ 4 479.9
Increased Qb2 2.21 -2.20 0.45
(5 nominal) 8.07X10- 3 -1.83x10-3 546
7.46 X 10 -6 ±7 2.6 x 10 ~ 4 546
Increased Qal 1.018 -1.008 0.99
(33 nominal) 9.84x10-3 -9.26x10-3 10&
4.44 x 10 ~ 6 -4.91 xlO- 4 2037
Increased Qa2 1.010 -1.003 0.997
(33 nominal) 6.74x10-3 -3.33x10-3 300
1.82X10-5 ±j 2.63 xlO -3 300
Reduced Qb2 4.66X10-2 -2.90X10-2 34
(0.01 nominal) 5.86 x 10 ~ 4 -1.05x10-2 95
Increased Qbl 1.004xlO~ 6 -6.78x10-3 147
(10 nominal)
 176 W.S. WIDNALL AND P.K. SINHA J. GUIDANCE AND CONTROL

and the assumed noise densities, two of which have equal Similarly, using Eq. (27) in Eq. (28) and applying Eq. (30) one
values. A dimensionally correct expression that also gives the obtains
right numerical value is discovered to be
k2/kt (31)
(26)
The preceding formulas for the gain ratios are in excellent
This formula is in excellent agreement with the numerical agreement with the numerical results in Table 3, for all cases
results in Table 3 (except for the greatly reduced Qb2 case). that have the nested fast loop.
The formula appears valid under the same conditions that An approximate formula for the location of the slower
give rise to the nested fast loop. A remarkable conclusion is poles as a function of the noise densities may be derived by
that the first gain is optimized based only on the relative assuming k2 « k2 and by using Eqs. (30) and (31) in Eq. (25).
strengths of the two noise densities in the assumed altimeter
error model. a2/Qb2] *
One may derive additional useful formulas for the other
gains as a function of the noise densities as follows. The (32)
explicit formula for the mean-square velocity error / as a
function of the noise densities was given in Eq. (9). Assume This formula also is in excellent agreement with the numerical
that the nested fast loop conditions Eqs. (23) and (24) apply. results.
Also assume that k2>c. Note that the numerical results in- An interesting observation supported by the computer
Downloaded by BEIHANG UNIVERSITY on November 15, 2019 | http://arc.aiaa.org | DOI: 10.2514/3.55966

dicated that the contribution of the term proportional to Qbl results is that when the nested fast loop is optimal, the op-
was negligible. Assume that the gradient of this term with timal second and third gain ratios (k2/kj and k 3 / k j ) as well
respect to k2 and k3 is also negligible. Delete this term from as the optimal second and third pole locations are not a
the analysis. An approximate formula for the cost (mean function of the assumed density of the short-correlation-time
square velocity error) is then altimeter error. The computer results showed a factor of 5
increase in Qbl producing a shift in the optimal time constant
— Lr *za* ' K
K 2
I, K
I, ^a*
2 3 <27) of less than 0.1%.
When the strength of the altimeter error random walk Qb2
is sufficiently large relative to the strength of the other sources
Necessary conditions for a set of gains to be optimal include of error, that is when
Qai/Qb2<2c (33)
82J/dk2=0 (28)
Qa2/Qb2<C2 (34)
82J/dk3=0 (29)
then the optimal gain ratios and their associated pole
Using Eq. (27) in Eq. (29) one obtains locations and time constants are simply
~fQb~2 (30) A: 2 /A: 7 «2v / c=3.5xlO- 3 s- 1 (35)

Table 4 Error source initial values and statistics

Random walks x =w
Error source Initial value Noise spectral density N
X, Y (level) gyro drift rates 0.003 deg/h (0.003 deg/h)2 /h .
Z (azimuth) gyro drift rate -0.295 deg/h (0.005 deg/h)2 /h
2
X, Y (horizontal) accelerometer 50 fig (10/i£) /h
biases
Z (altitude) accelerometer bias 100^
First-order Markov processes
Error source Initial and 1 a value Inverse correlation time, /3
Barometric altimeter 500ft y/(250n.mi.)
time-varying error
East deflection of gravity y/(10n. mi.)
North deflection of gravity 17 ^ i;/(10n.mi.)
Gravity anomaly 35 ^ y/(60n.mi.)
Random constants
Error source Initial value
G-sensitive gyro drift coefficients 0.3 deg/h/£
G2 -sensitive gyro drift coefficients 0.04deg/h/g2
X, ygyro scale factor errors 300 ppm
Zgyro scale factor error 1000 ppm
Gyro input axis misalignments ±40 arc-sec
Accelerometer scale factor errors 150 ppm
X, Y accelerometer input axis misalignment
about Z ±180 arc-sec
Other accelerometer input axis misalignments ± 30 arc-sec
Barometric altimeter scale factor error 0.03
Coefficient of static pressure measurement error 1.54xlQ- 4 ft/(ft/sec) 2
Barometric time delay 0.25s
 MARCH-APRIL 1980 OPTIMUM GAINS FOR A BARO-INERTIAL VERTICAL CHANNEL 177

35

30

10 n————i————i————i—
300 800 1000 1800
TIMECSEZCS)
5

\
200 400 600 800 1000 1200
TIME (SECS)
Fig. 2 Trajectory altitude history.
Downloaded by BEIHANG UNIVERSITY on November 15, 2019 | http://arc.aiaa.org | DOI: 10.2514/3.55966

1
0 800 600 800 100®. 1800
TIMECSECS)
Fig. 4 Performance with baseline gains.

200 400 600 800 1000 1200

TIME (SECS)
Fig. 3 Trajectory heading history.

(36)

P2,p3 « -Vc= - 1.75 x 10 (37)

r 2 ,T J «//Vc=571s (38)

These limiting results are a function only of the destabilizing

gravity gradient c. Note that the nominal case, the increased
Qbl case, and especially the increased Qb2 case have computed
results (Table 3) approaching this limiting case. An important
conclusion is that even if the measured specific force is perfect
(zero Qal and Qa2), the feedback gain ratios k2lkl and k$lkl
must be maintained at certain nonzero values to stabilize the
vertical channel and to minimize the effect of the gravity 800 1000 1£00
computation error. These required values correspond to an '
upper limit on the optimal double pole time constant of 571 s. Fig. 5 Performance with optimized gains.
When the strength of the acceleration short correlation
error Qa] is important in the sense that
These formulas approximate computed results obtained in the
al /Qb2 - Qal /Qb2 >2-Jc2 + Qa2 (39) increased Qal case (Table 3).
When the strength of the acceleration error random walk
then the optimal gain ratio k2lkl simplifies to Qa2 is important in the sense that
(40) al /Qb2 <2^2 + Qa2/Qb2

and Eq. (32) yields two real poles at then the optimal gain ratios are
(41) (44)
(42) (45)
 178 W.S. WIDNALL AND P.K. SINHA J. GUIDANCE AND CONTROL

and the associated pole locations are The optimal gains produce a significant performance
improvement compared with the baseline case. The rms
(Qa2/Qb2) * (46) vertical velocity error is reduced 30%.
The recommended value for kt perhaps should be accepted
These formulas approximate the computed residuals obtained with some degree of skepticism. However, the recommended
in the increased Qa2 case (Table 3). values for the gain ratios k2/k1 and k3/kj can be adopted
with some confidence, because of their low sensitivity to the
Simulated Performance assumed noise values. The low sensitivity is a result of the
The vertical channel gains obtained by the optimization gain ratios approaching a fundamental limit imposed by the
procedure have been evaluated using a time-domain three- destabilizing feedback of the gravity computation error.
channel simulation that includes detailed models for the If the short-correlation time acceleration error is more
sources of acceleration and altitude errors and that exhibits important than assumed in the nominal case, the optimal gain
the dependence of these sources of error on the aircraft ratio k2lkl increases and the optimal double pole splits into
trajectory. Error sources included in the simulation are listed two real poles. On the other hand, if the acceleration error
in Table 4. The aircraft trajectory simulated represents a F4 bias (such as due to accelerometer bias and gravity anomaly)
is shifting more than assumed in the nominal case, both
Downloaded by BEIHANG UNIVERSITY on November 15, 2019 | http://arc.aiaa.org | DOI: 10.2514/3.55966

tactical mission profile. Figure 2 shows the altitude vs time.

Figure 3 shows the heading vs time. This is a high dynamic optimal gain ratios are increased and the optimal double pole
trajectory. During the first descent the aircraft executes a splits into a complex conjugate pole pair.
"figure-eight" maneuver. During the remainder of the flight A detailed baro-inertial error simulation has exhibited the
the aircraft is performing rapid zigzag evasive maneuvers reduced vertical velocity errors that can be obtained with the
while climbing and diving. In some of these maneuvers, optimized gains. It provides confidence that the fundamental
maximum bank angles of 70 deg are used, with associated assumptions of the stochastic analysis are sound.
load factors of 3g. One pull-up has a load factor of 5g. Because of the very long settling time associated with the
The vertical channel performance results with the baseline recommended optimized gains, one should also implement a
gains, Eq. (12), and with the optimized gains, Eq. (14), are faster set of gains for use in the ground alignment mode.
exhibited in Figs. 4 and 5. The optimized gains do provide a
lower level of vertical velocity error. A noticeable con-
sequence of the fast (r =1 s) loop around the indicated
altitude is the increased noise content in the indicated altitude References
due to the baro-error fluctuations. This may be a disad- Draper, C.S., Wrigley, W., and Hovorka, J., Inertial Guidance,
vantage of the optimized gain set in some applications. Pergamon Press, New York, 1960.
2
McClure, C.L., Theory of Inertial Guidance, Prentice-Hall,
Englewood Cliffs, N. J., 1960.
Conclusions 3
Markey, W.R., The Mechanics of Inertial Position and Heading
The fundamental assumptions underlying the results of this Indication, John Wiley & Sons, New York, 1961.
analysis are that the most important sources of error in the 4
Pitman, G.R., Jr. (ed), Inertial Guidance, John Wiley & Sons,
third-order vertical channel may be adequately modeled by New York, 1962, p. 41.
5
random walks and white noises as presented earlier. If these O'Donnel, C.F., Inertial Navigation, Analysis and Design,
assumptions are correct, then the following conclusions are McGraw-Hill, New York, 1964, pp. 35-36.
6
obtained. Broxmeyer, C., Inertial Navigation Systems, McGraw-Hill, New
The most significant source of error in the vertical channel York,1964, p.148.
7
is the fluctuation in the altimeter bias (such as due to altimeter Farrell, J.L., Integrated Aircraft Navigation, Academic Press,
New York, 1976, pp. 21-22.
scale factor error and nonzero vertical velocity). The second 8
Kayton, M. and Fried, W.R., Avionics Navigation Systems, John
most significant source of error is the short-correlation time Wiley & Sons, New York, 1969, pp. 317-319.
9
acceleration error (such as due to specific force measurement Ausman, J.S., "GPS Bombing Demonstration Error Analysis,"
error during a maneuver). Altimeter noise at the assumed Litton Guidance & Control Systems Division presentation to General
level has negligible effect. Dynamics Electronics Division, May 1976.
10
The optimal choice of the gain set is radically different Whalley, R., "Inertial Techniques in Height Measurement,"
from the baseline gain set, which provides a triple pole at Journal of the Institute of Navigation, Vol. 19, No. 1, 1966, pp. 61-
T = 100 s. The optimal gains include a kl that is two orders- 67.
11
Ausman, J.S. and Kouba, J., "Airborne Ran^e Instrumentation
of-magnitude larger than that in the baseline set. The optimal System (ARIS)," Vol. II (Mechanization Equations), Litton Guidance
value of kl is a function only of the relative strengths of the and Control Systems Document No. 402029, Woodland Hills, Calif.,
altimeter bias fluctuation and the altimeter noise. For the Nov. 1972, pp. 25-28.
assumed relative strengths in the nominal case, the optimal 12
Ausman, J.S., et al., "Close Air Support System (CLASS) F-4D
gain value recommended is kl = 1.0 s ~ *. The optimal values Flight Test," Vol. I (Summary), Air Force Avionics Laboratory,
for k2 and k3 are sensitive to the assumed numerical values Wright-Patterson Air Force Base, Ohio, TR-73-363, Sept. 1973, pp.
for the noise densities. However, the optimal gain ratios 96-102.
13
k2/k1 and k3/kt are relatively //isensitive to the assumed Blanchard, R.L., "A New Algorithm for Computing Inertial
noises. The gain ratios recommended by the nominal optimal Altitude and Vertical Velocity," IEEE Transactions on Aerospace
and Electronic Systems, Vol. AES-7, Nov. 1971, pp. 1143-1146.
solution are k2/kt =4.2x 10~ 3 s ~ l and k3/kj =4.4x10~ 6 14
Newton, G.C., Gould, L.A., and Kaiser, J.N., Analytic Design
s ~ 2 . This choice of gains will place a fast pole at r = 1 s and a of Feedback Controls, John Wiley & Sons, New York, 1961.
double pole at r =480 s. This time constant of the optimal 15
Hooke, R. and Jeeves, T.A., "Direct Search Solution of
double poles is slower than the time constant of the baseline Numerical and Statistical Problems," Journal of the Association for
triple poles. Computing Machinery, Vol. 8, April 1961, pp. 212-229.