MMFEhmhhhhhh2hl OF KJUNM FILTER PERNUHANCEW CU)NAft … · 2014. 9. 27. · rMP-M47 ~ ~ I EMMM.or...
Transcript of MMFEhmhhhhhh2hl OF KJUNM FILTER PERNUHANCEW CU)NAft … · 2014. 9. 27. · rMP-M47 ~ ~ I EMMM.or...
rMP-M47 ~ ~ I EMMM.or -TIEK DwI ELLIPM 0S~OF KJUNM FILTER PERNUHANCEW CU)NAft POSTUMRUATE
WSCESIIED SCHOOL MONTEREY CA J JAAOS NRM 64F' 2MMFEhmhhhhhh2hlmhhmhhmmhhhhlm
-mmhhhhhhhhl-III"""mmo
-I'll.'
lul ~ ~ 2812.5
w-I~' lml 1112.2
S140 1112.0
1 .
1.25 1111.4 11 .
MICROCOPY RESOLUTION TEST CHART
NATIONAL BUREAU OF STANDARDS1963-A
-
S
S
- - J - Ji I .. . . . . . . . . .. .. .. . . . ... ...
lI
NAVAL POSTGRADUATE SCHOOLMonterey, California
00
¢ID
o DTICSELECTE
THESISDEVELOPMENT OF REAL-TIME ERROR ELLIPSES
AS AN INDICATOR OF KAL"IANFILTER PERFORMANCE
by
CD2 Joseph JarosC.D)
March 1984
,__ Thesis Advisor: A. Gerba, Jr.
Approved for public release, distribution unlimited
8A " ', 0,12
SECURITY CLASSIFICATtON OF THIS PAGE (fhen Des Entered)
READ INStRUCTIONSREPORT DOCMENTATION PAGE BEFORE COMPLETING FORMI. REPORT NUME 2. GOVT ACCESSION NO: 3. RECIPIENT'S CATALOG NUMBER
4. TITLE (nd Subtitle) S. TYPE OF REPORT & PERIOD COVERED
Development of Real-Time Error Master's Thesis;Ellipses as an Indicator of Kalman March 1984Filter Performance 6. PERFORMING ORG. REPORT NUMBER
* AUTNOR(d) S. CONTRACT OR GRANT NUEBER(s)
Joseph Jaros
S. PER1FONMING ORGANIZATIOm NAME AND ADDRESS I*. PROGRAM ELEMENT. PROJECT, TASKAREA & WOAK UNIT NUMBERS
Naval Postgraduate SchoolMonterey, California 93943
IS. CONTROLLING OFFICE NAME AND ADDRESS 12. REPORT DATE
Naval Postgraduate School March 1984Monterey, California 93943 13. NUMBER OF PAGES
11914. MONITORING AGENCY NAME G AOOttESS(I* dillerent trai Controlling Office) IS. SECURITY CLASS. (of this report)
UNCLASSIFIED
D5E. DECLASSIFICATION, OOWNGRADING 0SCHEDULE
14. OISTRIGUTION STATIMENT (of this Repo")
Approved for public release, distribution unlimited
17. DISTMIOUTION STATEMEN (of te abstrato, o med in Sock 20. It dlliferen from Report)
IS. SUPP.LEETARY NOTES
19. KEY WORDS (Cmntine on reveo ole, It ndeeaey md Identil by block number)
Error Ellipsoids; Kalman Filter; Extended Kalman Filter
20. ABSTRACT (Codtwe an puweus aide 1 necesary mnd Id"tty, by block member)An error ellipse plotting routine was developed to provide real-
time indication of Kalman filter performance. The study includedan evaluation of the Hewlett-Packard HP-88 computer system'scapability for providing real-time tracking information and anevaluation of the computer's possible use on the three-dimensionalunderwater tracking range at the Naval Underwater WeaponsEngineering Station, Keyport, Washington. A series of tracking
DO I , ,, ,III 1473 coTON OF I NOV 66 IOBSOLETE 1S'N 0102. LF 0146601 SECURITY CLASSIFICATION Of THIS PAGE (When, Data Enhoroc
SECURITY CLASSIVlCATIOM OF THIS PAGE (Whm Dma EXeCOM0
runs were used to demonstrate both linear and extended Kalmanfiltering. Information obtained from the error ellipses wasused to modify filter parameters for improved filter performance.It was found that the error ellipse was useful as a tool forindicating filter performance and for making decisions regardingfilter parameter modification. The HP-86 provided accurate,reliable results and it could be used for on-line graphics.However, the computing speed of the HP-86 computer as used inthis study was too slow for on-line processing of the three-dimensional tracking problem.
Accession ForNTIS GI.&T
B y ...... ... .
A v"i i.1 .1 i ty C c, es
7 k, .! aj./or --
Di st
of
S N 0102-LF-014-6601 2
SECURITY CLASSIFICATION OF THIS PAGtU a'l Daeft Entere)
Approved for public release, distribution unlimited
Development of Real-Time Error Ellipsesas an Indicator of Kalman Filter Performance
by
Joseph JarosCommander, United States Navy
B.S., University of Texas, 1967
Submitted in partial fulfillment of therequirements for the degree of
MASTER OF SCIENCE IN ELECTRICAL ENGINEERING
from the
NAVAL POSTGRADUATE SCHOOLMarch 1984
Author:
Approved by:Thesis Advisor
'0'0 *t" a-"m... Second Reader
Chairman, Depa mentof Electrical and Computer Engineering
I# dan of Science and Engineering
ABSTRACT
An error ellipse plotting routine was developed to
provide real-time indication of Kalman filter performance.
The study included an evaluation of the Hewlett-Packard HP-86
computer system's capability for providing real-time
tracking information and an evaluation of the computer's
possible use on the three-dimensional underwater tracking
range at the Naval Underwater Weapons Engineering Station,
Keyport, Washington. A series of tracking runs were used
to demonstrate both linear and extended Kalman filtering.
Information obtained from the error ellipses was used to
modify filter parameters for improved filter performance.
It was found that the error ellipse was useful as a tool
for indicating filter performance and for making decisions
regarding filter parameter modification. The HP-86 provided
accurate, reliable results and it could be used for on-line
graphics. However, the computing speed fo the HP-86 computer
as used in this study was too slow for on-line processing
of the three-dimensional tracking problem.
I4
TABLE OF CONTENTS
I. INTRODUCTION- ------------------------------------- 7
II. KALMAN FILTER THEORY----------------------------- 9
A. LINEAR MATHEMATICAL MODEL -------------------- 9
1. The Plant -------------------------------- 9
2. Noise Processes -------------------------- -0
3. Initial State Description ----------------- 11
B. DISCRETE TIME ESTIMATION --------------------- 12
1. The Estimator Equations ------------------ 12
2. Gain and Covariance Equations ------------ 14
C. NONLINEAR ESTIMATION - EXTENDED KALMANFILTER --------------------------------------- 17
1. Nonlinear Model -------------------------- 17
2. Extended Kalman Filter Equations --------- 17
III. ERROR ELLIPSOIDS --------------------------------- 20
A. THEORY --------------------------------------- 20
B. ERROR ELLIPSOIDS AND FILTER DIVERGENCE ------- 25
IV. PROBLEM DEFINITION ------------------------------- 27
A. PROBLEM DESIGN ------------------------------- 27
1. Linear Tracking -------------------------- 27
2. Nonlinear Tracking ----------------------- 32
B. COMPUTER SIMULATION -------------------------- 34
1. Computer Hardware/Software --------------- 34
2. Track Generation ------------------------- 35
3. Noise Generation ------------------------- 35
4. Gating Scheme ---------------------------- 36
5
5. Collection of Statistics- -- -- -- -- -- -- -- -- -- ----
V. TARGET TRACKING AND ERROR ELLIPSE ANALYSIS ------- 38
A. LINEAR 77RACKING--------------------------------- 38
B. NONLINEAR TRACKING----------------------------- 42
VI. CONCLUSIONS AND RECOMMENDATIONS--------------------8al
A. ERROR ELLIPSE----------------------------------- 81
B. COMPUTER PERFORMANCE--------------------------- 82
APPENDIX A - TRACK GENERATION---------------------------- 85
APPENDIX B - COMPUTER PROGRAM EXPLANATION---------------- 94
APPENDIX C - PROGRAM LISTING------------------------------ 98
LIST OF REFERENCES--------------------------------------- 117
BIBLIOGRAPHY---------------------------------------------- 118
INITIAL DISTRIBUTION LIST-------------------------------- 119
A
6
I. INTRODUCTION
The Kalman filter's importance as an estimator and
predictor is well documented. Providing real-time informa-
tion concerning filter performance so that on-line adjust-
ments to filter parameters can be made continues to be an
area of high interest. This study investigates the
usefulness of the error ellipse as a tool for providing a
real-time indication of filter performance.
Part of the investigation involves an evaluation of the
Hewlett-Packard HP-86 computer system's capacity to operate
in a real time tracking environment, and its capabilities for
providing information concerning filter performance. The
rationale behind this investigation is based on the require-
ment for the Naval Underwater Weapons Engineering Station,
Keyport, Washington, to accurately acoustically track
torpedoes on a three-dimensional underwater tracking range.
A knowledge of the range operation is not within the scope
of this study. It is sufficient to know that presently
the range receives four time measurements every 1.31 seconds,
and these measurements are nonlinear functions of the
torpedo position.
To gain a better understanding of the error ellipse, a
4-state tracking scenario was chosen for this study.
7
Initially, the linear tracking problem is discussed,
followed by an investigation of the nonlinear probie,. Of
primary interest are on-line methods to improve filter
performance using information provided by the error ellipse
for filter parameter modification.
L
8
ii KAL'., FILTER THEORY
A. LINEAR MATHEMATICAL MODEL
1. The Plant
For this model the state and measurement equations
for the plant are linear. Hence the discrete form is used.
The assumed plant model is described by a linear, vector
difference equation:
x(k+l) = cx(k) + Au(k) + rw(k) (State Equation) (2-1)
and a linear, vector measurement equation:
z(k) = cx(k) + v(k) (2-2)
where: Sx(k) is an n-dimensional column vector, denoting the
state of the plant at "time" k.
u(k) is the deterministic control input, an m-vector,at time k.
w(k) is a p-dimensional vector representing any randomforcing inputs at time k.
z(k) is a q-dimensional vector representing measure-ments made at time k.
v(k) is a q-dimensional vector representing randommeasurement made at time k.
, &, r, and c are assumed constant coefficient matrices of
dimension nxn, nxm, nxp, and qxn respectively.
9
2. Noise Processes
In order to place probabilistic structure on the
noise processes v(k) and w(k) the following assumptions are
made:
(a) v(k) and w(k) are individually white processes,
that is, for any k and 1, with k 9 1, v(k) and S
v(l) are independent random variables, and w(k)
and w(l) are independent random variables.
(b) v(k) and w(k) are individually zero mean, 4
Gaussian processes with known covariances.
(c) v(k) and w(k) are independent processes.
Thus for the measurement noise: 0
MEAN: E[v(k)] = 0 (k=0,l,2,3,...) (2-3)
T TCOVARIANCE: E[v(k)v (1)] E[v(k)]E[v (1)1
0 k lA: k k = 1
or E[v(k)v T (1)] = Rk kl (k,l=0,1,2,...) (2-4)
where 6kl is the Kronecker delta function defined as:
: 0, k 1 1
: 1, k I 1
10
0
Likewise for the random forcing input:
< = (z l,2, 3,...)
E[w(k)w ( 1)] k 6 kl k,l=0,l,2,3,...) (2-6)
and R k are nonnegative definite symmetric for all
k. Also since v(k) and w(k) are zero mean and independent
then:
E[v(k)wT (1)] = 0 (2-7)
For the purposes of this study, unless otherwise
specified nk and R are considered to be known and constant,-k k
although both may be time varying.
3. Initial State DescriDtion
For the initial state of the difference equation (2-1)
it is unlikely that x will be available. Hence, it is
assumed that x is a Gaussian random variable of known mean
x and known covariance P., i.e.,
E[x(0)] x
-0 --0 - 0 --
11
This choice for the initial state has the advanz_:,
of causing the subsequent estimation scheme to be unbiased
for all z. [Ref. 1] Further it is assumed that the initial
state and the measurement noise are uncorrelated:
E[x(O)v T (k)] : E[v(k)x T(0)] : 0 (k:0,!,2,3,...)
Also the initial state and the random forcing input are
uncorrelated:
E[x(0)w T (k)] E[w(k)xT (0)] 0 (k:0,l,2,3,...)
B. DISCRETE-TIME ESTIMATION
1. The Estimator Equations
The estimation problem involves generating an
optimal estimate for x(j) for the system described by the
difference equation (2-1) from the noisy measurements
z(0),z(l),...,z(j). This estimate will be denoted by
x(j/j), which means the estimate of x at time j given
measurements at times up to and including time j. The
estimate must be optin.al in the sense that the expected
value of the sum of the squares of the error in the estimate
is a minimum, i.e.:
TAE{[(x(k/k) - x(k)] [x(k/k) - x(k)]} minimum
12
I_ __ ,,. . . . . .. ..__ .. .. .._. . .. . . . . . . . . . .
The estimator is characterized by the linear relationship:
x(k/k) = x(k/k-i) + G(k)[z(k) - cx(k/k-l)] (k C,,,..)
(2-8)
where
x(k/k) is the optimal (minimum variance) estimate ofx(k) given observations at times up to andincluding k.
x(k/k-1) is the optimal one-step prediction of x(k)given observations at times up to and -including k-l.
G(k) is the optimal estimation gain matrix whichwill minimize the variance of estimationerror.
For the initial estimate ;(0/0), the estimator
equation (2-8) is initialized with x(O/-i), which is not a
random variable. If x(O/-l) is selected such that: 0
;(0/-1) = E[x(O)] =
it can be shown that this choice of x(O/-l) makes the
estimator unbiased for all k. [Ref. 1] The estimators best
available information concerning x(k-l) is the estimate
x(k-1/k-l), therefore it is reasonable to assume that
;(k/k-l) = x(k-l/k-l) + Au(k-l) (2-9)
is the best prediction.
1
13
l r I .. .... ' . . . m . . .. . .
In summary, equations (2-8) and (2-9) are the estimator
equations, with x(O/-l) = x as the initial cond-:ir
2. Gain and Covariance E;uations
Without going into detailed derivations, tne optimal
estimator gains, G(k), used in the estimator equation (2-8),
are those which satisfy:
G(k) = P(k/k-I)CT [cP(k/k-)C T + R] - I (2-10)
P(k/k) = [I - G(k)C]P(k/k-1) (2-11)
P(k+l/k) = P(k/k) T + Q (2-12)
with the initial conditions:
P(0/-l) P 0 E{[x(0) - ][x(O) ---- o --- o - --<
where
P(k/k) = Ef[x(k/k) - x(k)][x(k/k) - x(k)]T }
is the covariance of estimation error matrix.
0L
P(k/k-1) E{[x(k/k-1) - x(k)][x(k/k-1) - x(k)]}
is the covariance of one-step prediction error matrix.
Q E[r(k)-w(k).wT (k)'rT (k)]
is the state excitation matrix.
P(k/k) and P(k/k-l) are symmetric, positive definite matrices.
Several observations can be made concerning the linear
Kalman gain (2-10), covariance (2-11, 2-12) and estimator
(2-8) equations.
(a) The estimator gains, G(k), do not depend on the
measurement data and hence can be precomputed, stored, and
used as the processing measurements become available.
(b) Although not obvious from the equation, the
time-varying gain, G(k), depends in time as:
1G1k) =k-+ 1) [Ref. 21 (2-13)
Thus the effect is to weight the correction term,
[z(k) - cx(k/k-l)], in the estimator equation (2-8) less
heavily as time progresses. The advantage of a greater
initial weight allows for possibly large differences
between z(k) and x(k/k-1) during the initial observations,
and a large gain will result in a significant change in the
- I
next estimate. This advantage is also borne out in that
there is less confidence in the quality of the estimates
during the early observations compared with the quality
after numerous observations. Hence the later an observation,
the less drastic an estimate will be altered or affected by
an isolated observation discrepancy.
(c) In general, the variance of estimation error
decreases in a manner analogous to the gain schedule (2-13),
i.e., it decreases as k grown larger, reflecting greater
confidence in the estimate as the number of observations
increases. Selection of the proper initial condition, P ,
is important when studying the effect of measurement errors
on the behavior of the estimate. So P should be assigned-0
pessimistic values which would correspond to a lack of
information about the initial state. In cases when the 5
initial state is completely unknown, then P --I. [Ref. 3]-- o
(d) The Q matrix serves to compensate for model
errors and prevents the covariance matrix from becoming too 0
small or optimistic. A small covariance matrix would result
in a small filter gain, and subsequent observations are
essentially ignored, which could result in filter divergence. 0
The Q matrix prevents G(k) from approaching zero by adding
uncertainty to the system which is reflected in a
degradation of certainty (increase in P(k+l/k)).
16
F-
0
C. NONLINEAR ESTIMATION - EXTENDED KALMAN FILTER
in many practical applications, the state equations
and/or measurement equations are nonlinear. Before the
Kalman filter equations can be used, the problem must be
linearized and the Kalman filter equations are applied with
some modification.
1. Nonlinear Model
Consider a nonlinear discrete system of state and
observation equations given by: 0
x(k+l) = f(x(k), u(k),k) + w(k) (2-14)
and
z(k) = h(x(k), k) + v(k) (2-15)
In these equations f and h are nonlinear functions
of the state variable x, w(k) is the plant excitation noise,
and v(k) is the measurement noise. The plant noise and
measurement noise are assumed to be uncorrelated, zero-mean,
and white. The same equations (2-3 thru 2-7) apply as for 0
the linear model.
2. Extended Kalman Filter Equations
In order to apply the linear filter equations,
equations (2-14) and (2-15) are expanded about the best
estimate of the state at that time and only the first-order
terms are kept.
17
That is, defining A(k) as:
A~k) afA(k) a (x(kix), u(k), k)
and
H(k) (x(k/k-1))
Ux
As can be seen from the above equations, the filter
estimates, x(k/k) and x(k/k-l) are used as the "best"
estimates about which the linearization is performed. The
matrices Ak) and H(k) must be used to generate G(k) so it
is available to process z(k) when it is obtained. The
modified extended Kalman filter equations are then:
Gain Equation:
'(k) = P(k/k-l)H T (k)[H(k) • P(k/k-l) • HT (k) + R] -1
(2-16)
Filter Update Equation:
x(k/k) = x(k/k-l) + G(k)[z(k) - h(x^(k/k-l))] (2-17)
18
. . . I . . . . . . . . . .
Prediction Equation:
x(k+i/k) f(x(k/k), u(k), k) (2-18) 5
Covariance of Estimation Error Equations:
P(k/k-1) : A(k-l)P(k-1/k-1)A (k-1) + Q(k-1) (2-19)
P(k/k) = [I - G(k)H(k)]P(k/k-l) (2-20) 4
For the initial estimate x(0/0), equation (2-17) is
initialized with x(0/-l) with
x(o/-1) = E[x(0)] x
x(O/-l) is also used to initially evaluate H(k).
As in the linear case:
19
III. ERROR ELLIPSOIDS
A. THEORY
Since the estimate x(k/k) is unbiased in the Kalman
filter equations, the P(k/k) matrix represents the covariance
of the error in the estimate. If the estimate were biased,
P(k/k) would represent the second-moment matrix rather than
the covariance matrix. Hence, P(k/k) provides significant
information about the accuracy of the estimate. If the
physical model is accurately described by the state and
measurement equations (2-1, 2-2), then P(k/k) can be used
to describe the manner in which the estimate converges
(or diverges) to the true state. Examination of the P(k/k)
matrix directly, element by element, is not a realistic
2approach, since the matrix contains n elements, where n is
the number of state variables. To simplify the situation
the concept of the error ellipsoid is used. [Ref. 4]
As discussed earlier, the assumptions are made that the
initial state of the plant x is Gaussian, as are the random-o
processes v(k) and w(k). Using these assumptions it follows
that x(k) and x(k/k) are also Gaussian since they are linear
combinations of Gaussian variables and deterministic
quantities. Using the same rationale, the estimation error,
defined as:
20
e(k/k) x x(k/k) - x(k)
is also Gaussian. Using the fact that the mean of the
estimation error is 0, the probability density function for
e(k/k) is:
pe [(k/k)] = [( 2 )n/2 L?(k/k)1i/2-1
exp[-1/2eT (k/k)P- (k/k)e(k/k)] (3-1)
The density function, p_ [e(k/k)] will have a constant
value whenever the exponent has a constant value. That is:
-1/2e T (k/k)P- 1 (k/k)e(k/k) = c
or
T -l2
e (k/k)p- (k/k)e(k/k) = c (3-2)
where c is an arbitrary constant.
As demonstrated by Sorenson [Ref. 1] and Kirk [Ref. 21,
it can be shown that the locus of points e(k/k) which
satisfy equation (3-2) are hyperellipsoids. For the
two-dimensional case which is of concern, equation (3-2)
describes an ellipse. This can be seen by fixing time and
rewriting (3-2) as:
21
e~we
e we c( -)
where
w = P- (k/k) (a 2 x 2 symmetric matrix)
S
Expanding the left side of (3-3) gives:
2 2 2w1 1 e + W1 2W 2 1 e1e 2 +w 2 e2 c
which because of symmetry gives:
2 2w e + 2w 2e e2 + w22 e2 = c (3-4)
Since wl>0, w22>0, and W w 2 2 >w 1 2 , equation (3-4)
describes an ellipse, in which the principal axes do not
coincide with the coordinate axes. The ellipse is rewritten
in terms of z(1) and y(2) as the coordinate axis, and it can
be shown that y(l) and y( 2 ) are the eigenvectors of w with
1 and X2 defined as the corresponding eigenvalues. [Ref. 2]
(See Figure 3-1). The equation for the ellipse can be
rewritten in terms of a coordinate system having unit vectors
in the directions of y(1) and y(2) as the basis vectors.
The ellipse equation becomes:
Xil2(1) + X2 2(2) = c 2 (3-5)
22
*
e 20
Y(2)1<1
0
Figure 3-1 Error Ellipse
Remembering that w =P- (k/k), it can be shown that the
-II
corresponding eigenvectors and eigenvalues for w- 1P(k/k)
Y (1( )
2, 2
y ( 1 z ( 2 z c ( 3 -6 )1 2
23
Fi g ur 3 -1 E rro Ell p se
2 23
Z (i) Z (2
For the purposes of this study, the term error ellipsoi:
refers to the specific case when c 1. So equation <-,9)
becomes:
~2y2 (1) y (2)
+ -a 2
In terms of the Cartesian coordinates x', yt, which use e1
and e2 as basis vectors:
xf = xcosO + ysine
y I =-xsine + ycose
where e is the angle of rotation of the axes and can be
computed from:
2cov(e e)= 1/2 tan-l[ xve x ] [Ref. 5]
var(e) - var(ex y
For a given value of c it is possible to integrate the
probability density over the surface of the error ellipse to
obtain the probability that a particular sample point will
lie within the ellipsoid. For this study, n = 2, c = i,
and the probability the error is inside the ellipse is 0.394.
In summary, the error ellipsoid can be used to
characterize the concentration of the estimate abouz the
24
.. . 0 . . .
true value of the state. A decrease in the magnitude of an
axis cf the ellipse is an indication that the error 4n -ne
estimate is decreasing in that direction.
One important item that needs to be pointed out is that
often the components of the state vector represent entirely
different types of variables, for example the components
might represent range, velocity, and depth. Since the two-
dimensional ellipses are determined by using two components
of the state vector, it is reasonable to examine submatrices
relating state variables of the same character. Doing so
will preclude most scaling difficulties when plotting the
ellipses, and provide more meaningful insight in to the S
resulIts.
B. ERROR ELLIPSOIDS AND FILTER DIVERGENCE
Thus far the discussion has centered around using
submatrices of the P(k/k) covariance of estimation error
matrix as the input for the error ellipse to indicate
filter performance. As proposed by Heffes [Ref. 6] and
Nishimura [Ref. 7], P(k/k) can be considered as a "design"
covariance matrix P d, using the assumption that the only
errors are in R and P with the followina inecualities
holding 5or all k:
a d >a pd > .a-K • k k _ k --o -:-o (3-7)
2
- ' • , , , , • . . . . . . .
with the subscript "i" indicating designed and "a" indicating
actual. Equation (3-7) implies more input noise, more
measurement noise, and more initial state uncertainty in the
design than actually exists. This conservative filter design
results in a somewhat pessimistic design error covariance
Pd(k/k). The actual error covariance Pa(k/k) resulting
from using a filter designed with Q d, Rd , and Pd is related> --- o
to the design error covariance in the following manner:
P d(k/k) > P a(k/k)
This result is particularly useful when one simply does
not know accurately the noise covariance of the input or
output, but an upper bound is known. Designing assuming
the noise covariance is at its upper bound will result in
Pa(k/k) being upper bounded by Pdk/k). In some sense a
worst case design results. Filter divergence exists when
the design error covariance P d(k/k) remains bounded while
the error performance matrix Pa(k/k) becomes very large
relative to Pd(k/k) or is, in fact, unbounded.
2
IV. iRQBLEM DEL:i3I
A. PROBLEM DESI&N
The purpose of the tracking problem is to study the use
of error ellipsoids as real-time indicators of filter
performance. In order to keep the design model realistic
albeit reasonably simplified for ease of study, a
two-dimensional tracking problem using several different
tracks has been selected.
1. Linear Tracking
All tracks are based on an x-y coordinate system
with the target moving in the x or y direction relative to
the sensor located at the origin. Thus for aircraft
tracking, altitude is considered constant, as is depth for
torpedo tracking.
Defining the state variables as:
xi = x x-coordinate of the target location.
x2 = x velocity of target (vx) in x-direction.
x3 = y y-coordinate of the target location.
x4 =y velocity of target (v y) in y-direction.
2'7
resultinq in a state vector:
X (4-1- y
*Y
The following are the state equations:
1 (t) x2 (t)
x 2 (t) = W1 (t) S
(4-2)
x3 (t) = x4(t)
4(t) = w 2(t)
where w (t) and w 2(t) are assumed to be uncorrelated,
random processes that account for unknown target accelerations
and nonlinear target motions. Writing the discrete form of
the state equations gives:
28
xi(k+l) x!(k) + T x.2(k) + M1 2
x 2(k+l) x2 (k) + T w (k)
(4-3)-2x (k~l) =x1)(k) + T x4(k) + ±-2 w k
x4(k+l) x 4(k) + T w 2(k)
or
x(k+l) ox(k) + rw(k) (4-4)
With T, the sampling period equal to 1 second, the matrices
_ and r are:
1 1 0 0 .5 0
S4- ( 5)0 0 1 1 0 .5
0 0 0 1 0 1
It is assumed that the sensor gives noisy, but
uncorrelated measurements of x and y. Hence the discrete
measurement equations are:
23
zl(k) = x,(k) + v,(k)
LS
z2 (k) x3 (k) + v 2(k)
with v (k) and v (k) uncorrelated random noise.1 2
Thus for the measurement equation:
z(k) = cx(k) + v(k) (4-7)
the matrix c is:
1 0 0 0_IC:
For the initial run and unless otherwise noted the
following values for the Gaussian random processes will be
used:
E[v(k)] = 0 for all k
-20xl10 0
T Wl2 RfoalkE[v(k)v (k)] m R 'or all k
0 20x!O
30
rm the standard deviation of
_15 0 measurement noise.
E[w(k)] = 0 for all k
100 0
E[w(k)w (k)] (m/sec 2) = coy w for all k
0 100
10
2C : m/sec the standard deviations of the-w S
_10- random forcing input.
The covariance of estimation error matrix is initialized:
-2210 2 0 0 0
0 10 20 0? =P(/-l) 0 o2 0
-0- 0 0 10 0
0 0 0 102
31
Since the filter is to be unbiased, the initialization:
x(O/-l) = x = Initial condition of the problem.
2. Nonlinear Tracking
The state equations are the same a3 f.r The linear
tracking problem. The measurement equation is considered
as a noisy range measurement by the tracking sensor and is
characterized as:
z(k) [x2(k) + x2(k)] 1 2 + v1 (k) (4-8)
Thus z(k) is a nonlinear function of the states. Using
equation (2-14) and with T = I second:
1 I (k) + x2(k)
f(x(k), u(k), k) 2
x3 (k) + x4(k)
4(k)
32
I q
Taking partial derivatives of f with respect to x i_
1 1 0 0-
A(k) = = __0 1 0 0
(x(k/k), u(k), k) 0 0 0 0
using equation (2-15):
h(x(k), k) = [x2 (k) + x 2 k) 11/21 3
and taking the partial derivative with respect to x gives:
H(k) -102 - 2 ]12-- 1 (k)+x2(k)] /
(x(k/k-1))
x3(k) 02 -2 1/2
[x 2(kl+x32(k)] x1k/-l
1 3
33
I
Using the above results with the same vali- s :: 7-i
random noise Drccesses as previously stated for the liear
case, thie exended Kal.an filter equations can noD; ae o pc lea
to the nonlinear tracking problem.
B. COMPUTER SIMULATION
1. Computer Hardware/Software
a. Hardware
All computer simulations were run on the Hewlett-
Packard HP-86 personal computer. This particular model was
chosen to evaluate its capabilities in determining its
usefulness in actual torpedo tracking at the underwater
tracking range at Naval Underwater Weapons Engineering
Station, Keyport, Washington. The HP-86 system used included
keyboard, 9 inch CRT monitor connected through an integrated
monitor interface, and two HP Flexible Disc Drives connected
through an integrated disc interface. Plotting was done on
a HP-7225B Plotter and printing on a HP-2631G Printer. These
peripherals were interfaced using a HP-IB Interface module.
Because the system uses interface select codes, the HP-IB
factory preset code was set at 7, which is the select code 5
for the printer/disc interface. This code however did not
work when interfacing with the external plotter and printer,
since duplicate select codes are not allowed. So the 5
internally set select code of the HP-IB was set to 8 for
proper system operation.
34
S
b. Software
The HP-86 has 60K built-in, useable bytes of
computer memory, expandable to 572K using either 32K, 64K,
or 128K Memory Modules. A HP-86 pilg-in ROM was required
to operate the external plotter. Also a Matrix ROM was used
to reduce program length and computer run time.
All programs were written in BASIC programming
language using REAL (full) precision, which provides 15 digit
precision. Appendix B provides an explanation of the program
options and Appendix C contains the program listings used for
this study.
2. Track Generation
To evaluate the real-time use of the error ellipse as
an indicator of filter performance, Monte Carlo simulation
runs were made. Four tracks were generated by separate
programs and one second incremental values of x, y, v , and
v were stored in data files. Appendix I contains an
explanation of the generation of tracks three and four.
3. Noise Generation
In order to simulate the random noise processes, the
computer's random number generator was used and the generated
numbers scaled accordingly. For each track and each
different value of noise sigma, a different generator "seed
number" was used. These noise values produced were added to
the applicable true track values to simulate a sensor
measurement corrupted by independent ]aissian noiSe. For all
3:
Icases where filter parameters were varied for a parzicu!er
track under a specific noise condition, one noisy trez< waC
genera-ed, stored, an: used zhrughoLut that arti1-::
simulation. This was done for ease of filter performance
comparison.
4. Gating Scheme
In order to preclude catastrophic filter failure due
to excessive measurement noise, a bound was established for
the maximum acceptable limits of measurement noise. A
three-sigma gate was designed using the covariance of the
measurement noise, R, and the predicted covariance of error
matrix P(k/k-1). The gate is defined as:
Gate(k) = 3 (pii (k/k-i) + R )/2
max
This gate is the maximum error allowable for the measurement
at time k. If the absolute difference between the actual
measurement received and the predicted measurement is
greater than the three-sigma gate, then that particular
measurement data is rejected as unacceptable. When this
occurs, the filter gain, G(k), is set to zero, resulting in
that measurement being ignored and the prediction of the
states set equal to the estimate, that is:
x(k/k) = x(k/k+l)
36
5. Collection of Statistics
In order to study er,-or elipses as an indica .
filter oerformance, statisticc were calculated, D -
after each measurement during the Monte Carlo run. The
statistics computed were relative error (in some cases the
error was normalized), error mean, error variance, and error
covariance for the positional variables. The following
equations apply:
Relative Error = e(k/k) = x(k) - x(k/k)
(filter error residual)
kError Mean = e(k/k) - i/k E e(j/j)
j=l
kError Variance =Var[e.(kWk)] 1/k Z Ee.i(j/j)]-ELe (k/k)12
'xI, x?, X 3 , x 4
Var[e (k/k)] 1/k E [e (j/j)][e (j/j)]
Positional j=1 x3
Error -[x (k/k) . (k/k)]rxCovariance
= 1 x 3
Mat rix
i/k Z [e (j/j)][ex(j/j)] VarEe (k/k)]
j=l1 xIX 3 3
-[ xl(k/k)' x (k/k)]
37
V. TARGET TRACKING AND ERROR ELLIPSE AN4ALYSIS
A. LINEAR TRACKING
Track 1 depicts a target approaching at a constant
velcciry of 223.6 feet Der second. The solid line of Figure
5.1 indicates the true track of the target and the numbers
along the track indicate the time in seconds. The target
was tracked in a measurement noise, G = 150, with the
random forcing noise a = 1. R was set for 20,000. The
dots indicate the filtered track using the linear Kalman
filter equations. Figures 5.2 and 5.3 are the filter error
ellipses for this run, computed at increments of 10 seconds.
As can be seen the ellipse size decreases with increasing
time indicating filter convergence. The computed ellipse
surface areas shown on the figures confirm this. Figure 5.4
shows the filtered track for the case where a has been
increased to 300 and all other parameters remain the same.
The error ellipses of Figure 5.5 computed for 10-second
increments show increasing area indicating filter divergence.
Figure 5.6 shows the error ellipses for the same track run
but this time the ellipses were computed using a "statistics
window" of 10. By this is meant that the ellipses were
derived from statistics computed for the last 10 data
measurements. All previous data is disregarded. Usin7 this
3 e
0
method, the ellipses of Figure 5.6 show filter conver ence
from iterations 15 tz, 25. The fil tered zrack of Fi.::-
5 con r rs t1I. .zo're 5.7 shows e r'ess z z f
same tack parameters, except in this case a statistics
window of 5 was used. The window 5 ellipse area at time
25 (5179 so ft) is much less than the area of either the
run with the statistic window of 10 (11,540 so ft) or The
run with no window at all (65,500 sq ft). This is expected
since the filter is essentially "locked on" at time 21,
and the window 5 ellipse at time 25 disregarde all data
previous to tLme 20. Figure 5.8 shows the error ellipses
for the same track but the measurement noise was increased
to 0 = 40C, while R was kept at 20,000. The error ellipses
indicate filter divergence, and indeed the filtered track
headed off in the wrong direction.
Track 2 depicts a target approaching at a constant speed
of 500 feet per second in the -y direction. Figure 5.9
depicts the solid line track and the octs indicate the linear
filtered track with ' 150, G = 1 and R = 20,000. Figure
5.10 and 5.11 are the error eloo'ses cr the run. No
statistics window was used. Other than the fact that the
ellipse area is decreasin7, t_e shase of -he ellipse proviCes
little additional infoma-ion. Figure 3.12 and 5.13 show the
ellipses for the ncrma!i-e :'rcr. -ih the same measurement
noise si-ma for both the x-oosi-oin snh V-00siton measurements,
the normalized error eliho and D reflect
0..ne he de~=-o elc
0
the target track proximity to either axis. > Fin ' .
the ellipse major axis indicates a large i:r.alize: -r-i:
t-ne ZCecsn i -o e u-Ce-- --
maintains a constant x-position of 500 feet, while the
y-position is initially 20,000 feet. However, near time 40
as the target aporoaches the x-axis, the normalized v error
increases and becomes so large at the x-axis crossing that
the normalized x error becomes insignificant in comparison.
This can be seen in Figure 5.13 for the ellipses at time
45 and 55 compared with the ellipse at time 35. The
normalized error ellipse is an excellent indicator of
target proximity to an axis, but the rapid shifts in ellipse
surface area make it difficult to determine filter
convergence.
Track 3 depicts a target approaching on a parabolic
track at a speed of 203 feet per second. Figure 5.14 is the
true track, with a linearly filtered track indicated by the
dots. For this run a =150, a =1l, and R=20,000. Figure 5.15
are the error ellipse plots for times 40, 50, and 60, the
period of the highest rate of change in x- and y-velocity.
The ellipse areas increase with time indicating divergence.
Figure 5.16 and 5.17 are the ellipse plots using a 10-data
point and a 5-data point statistics window respectively.
In both cases the ellipse areas for time 60 are less than
time 50 indicating the filter has tracked around the curve.
40
.. 6
Using error ellioses based on a statistics window in tis
tracking situation provides a better indicator of filter
performance.
For the second run of track 3, a was increased to 300
and all other parameters remained the same. Figure 5.18
shows the resulting filtered track, which obviously didn't
track around the curve. With the large amount of measurement
noise and considering that the maximum random forcing input,
i.e. the maximum acceleration in the x- and y-direction,
occurs between times 40 and 60, it is a logical place to lose
track. Another factor to be considered is the decrease in
gain as k increases. By time 40 the gains have little
influence. So a system was incorporated in the program to
"reinitialize" the filter by setting the gains to G(0), if
certain conditions were met. After several trial and error
runs, it was determined that if a statistics window of 10
were used, and the gains reinitialized if the error ellipse
area increased consecutively a certain number of times,
that the filtered track would follow around the curve.
Figure 5.19 is the filtered track using a statistics window
of 10 and reinitializing the filter if the error ellipse area
increased consecutively during five 1-second increments.
The error ellipses for that run are shown in Figures 5.20
and 5.21. As indicated on the figures, the filter was
reinitialized four times and the ellipse areas for the
10-second increments increased, until reinitializing the
.L
filter at time 52 locked the filter in. Consequently e
error ellipse areas fr t-t Le and 70 decreased, indio4 -Cng
convergence.
In the final run, the filter was reinitialized after 10
consecutive 1-second error ellipse increases, with all other
parameters remaining tihe same. ig _re 5.22 is th filtered
track; Figures 5.23 and 5.24 are the error ellipse plots
for this run. As indicated on Figure 5.24 the filter was
reinitialized only once at time 57. A comparison of ellipse
areas for the last two runs shows that, with the exception
of time 70, the areas were larger in the first run when the
filter was reinitialized 4 times. This is expected since
reinitializing results in larger gains producing more widely
vary estimates, and hence greater error variance. At time
70 the error ellipse area of the first run is less since in
the first run the last filter reinitialization occurred at
time D2 versus time 57 in the second run. The first
filtered track had more time to settle out by time 70,
resulting in less error.
0B. NO:ILI-AR TRACKING
Track 4 depicts a target moving at a constant velocity
of 50 feet per second (30 kts) in the x-direction for 15
seconds, at which time the target turns and travels in the
-y-direction. (See Figure 5.25) Using the extended Kalman
filter with a =30 and R=900, a series of tracking runs were
4 2
made for various values of COVW (a 2=from 20 to 200). In nonew
of these runs did the filter successf' Ily track the tar--
around the turn. Figure 5.26 shows the results for CaSe
when CVW=ISO. In this instance the filter lost -rack as -he
target came out of the turn. Trying to track through a turn
using a constant COVW (and hence, a constant Q) did not work.
So a scheme was devised to vary COVW dependent on information
derived from the error ellipse. After several trial runs for
this particular track, it was determined that if the error
ellipse area increased consecutively for 7 iterations of
k, COVW would be doubled, and if the ellipse area decreased
consecutively for 5 iterations of k, COVW would be halved.
Initially the the trial runs were made without a statistics
window, and the filter did not track successfully. Without
the statistics window, the old data weighted down the statis-
tics, and the error ellipses were not indicative of what was
currently happening. So it was decided to use a statistics
window. Windows of 5, 10, and 15 were tried. Window 5
proved to be too responsive and window 15 not responsive
enough. So a statistics window of 10 was chosen for the
tracking run. With P =102, 230, R=900, and COVW initially
set at 20, the tracking run was made. Figure 5.27 depicts
the filtered track output, and Figures 5.28-5.30 are the
10-second incremental error ellipses for the run. As
can be seen, the filter did track around the curve. Figure
5.29 show3 the ellipse areas are becoming less between times
55 and 65. Also indicated below the plots are the values of
4S
COVW for the k time the plot was computed. During this
particular run COVW varied from an initial value of 23 up
to 160, and then decreased to 40 by the end of the run.
Using the same filter parameters as above except a was
increased to 200, and R to 40,000, another run was made. The
filter did not track at all. During this run COVW varied
from an initial value of 20 up to 40 and decreased to 5 by
the end of the run. Obviously, the criteria for increasing
and decreasing COVW was not effective. More trial runs were
necessary to determine the optimum consecutive increases
or decreases of the ellipse areas before adjusting COVW
accordingly.
A second approach to the nonlinear tracking problem, one
that was used earlier for the linear case, is to reinitialize
the filter if certain conditions are met. Again, using
trail and error runs with and without statistics windows,
it was determined that using a statistics window of 10 gave
the best results. Using initial conditions of COVW=I50,
P =102, a =30, and R=900, several runs were made, reinitializ--o< -v -
ing the filter if the error ellipse area increased consecu-
tively for a certain number of iterations of k. Of the runs
attempted, the best results were obtained when the filter was
reinitialized if the area increased for 5 consecutive
iterations of k. Figure 5.31 is the filtered track for this
run, and FiguresS.32-5.34 are the error ellipse plots.
Indicated below the plots are the times, k, when the filter
4 L
was reinitialized. For this particular run the filter was
reinitialized 5 times. It should be noted that whern2
reinitializing the filter, P(k/k-l) was reset to i X P .-o
P was not large enough to be effective in getting the--o
filter back on track. The initial value of 102 was used for
P to reflect the high confidence in the initial state--0
conditions. Other values of P did not work as well.-0
Using the same parameters as above except the filter was
reinitialized after 7 consecutive error ellipse area
increases, another run was made with the resulting track
depicted in Figure 5.35. A comparison with Figure 5.31
reveals that using 7 area increases as the criterion for
filter reinitialization resulted in poorer filter performance,
as can be shown from the error ellipses.
The final nonlinear filter tracking runs attempted
involved simultaneously varying COVW and filter reinitializa-
tion. The results were disastrous, and highly unpredictable.
It was an interesting experiment in futility, and no
meaningful results could be obtained.
45
SY-POS Tn
10
400
39
200
Figure 5.1 Solid Line Track 1, Vel 223.6 ft/sec; DotsIndicate Filtered Track, a =150, a =1, Rz20,000
46
50(
Y-POS (FT)155
78
~X-POS (FT)
-78
-155LO ~ r- LO
COVW-1 SIGV-15 R2 000 TIME AREA(SQ T) LEO
10 1057E+001 -20 8790E+000 ....
30 7540E+000 - - -
Figure 5.2 Filtered Track 1 Error Ellipses at 10 SecondIncrements, G 'i50
47
Y-POS (FT)
78
-78
-155
COVW-1 SIGV-150 R-20000 TIME AREA(SQ FT) LEG
40 6681E+00050 5741E+000 ..
Figure 5.3 Filtered Track I Error Ellipses at 10 SecondIncrements, a v =1Q
43
Y-PO amT
am
10
20
39
aF i II I-2M0
404
Figure 5.4 Filtered Track i, a 0300, a :i, R:2OJ00
'I .
Y-POS (FT)326
163 -/ I
/ z01
-326Co CY CCVJ Co Co C'JCTn -4 mI I
COVW-10 SIGV-300 R-20000 TIME AREA(SQ FT) LEG
5 2858E+000 -15 3480E+001 ....
25 6550E+001 - - -
Figure 5.5 Filtered Track 1 Error Ellipses for a =3OO
50
--
Y-POS (FT)326
163
X-POS (FT)
-163
-326
C\J D(D C'jI I
COVW-10 SIGV-300 R-20000 TIME AREA(SQ FT) LEG
5 2858E+000 -
15 1745E+001 ....
25 1154E+001 - - -
Figure 5.6 Filtered Track 1 Error Ellipses for % =330,Using Statistics Window of 10
S
Y-POS (FT)326
S
163
0!1~X-POS (FT)
-326 tN C Co N~
I I
COVW-10 SIGV-300 R-20000 TIME AREA(SO FT) LEG
5 2858E+000 -
15 7225E+000 ....
25 5179E+000 - - -
Figire 5.7 il-tered Track 1 Error' Ellipses =Using Stazistics 'indow of 5
S q
52S
Y-POS (FrT)617
308
0 X-POS (FT)
-308 -
-617CO- CV) O -
I I
COVW-10 SICVY400 R-2000 TIME AREA(SO FT) LEG
5 5825E+001
9 8489E+001 ....
13 9507E+001 - - -
.iit ered raci < EIrror 7-liise fo- - =40
5":S
I 4
Y-POS (FT)20000
18010
12000
20
8000
*304000
-4000
U, In m
Figure 5.9 Solid Line Track 2, Vel 500 ft/sec, Dots IndicateFiltered Track for u =150, a =i, R=20,000
-il IIII I i i . . . .
Y-POS (FT)93
47
0p
11) X-POS (FT)
-47
-93--
I I
COVW-1 SIGV-150 R-20000 TIME AREA(SQ FT) LEG
5 3868E+00015 4604E+000 ....
25 3132E+000 - - -
Figure 5.10 Filtered Track 2 Error Ellipses, = 150
. . . ..4. . . . . . .
0
Y-POS (FT)93
47
0 \ .. X-P0S (FT)
-47
-93I I
COVW-1 SIGV-150 R-20000 TIME AREA(SQ FT) LEG
35 2550E+00 -
45 2137E+000 ....
55 1950E+00 - - -
Figure 5.11 Filtered :rack 2 Error -pses, u =l5C-V
Y-POS (FT)1E-001 T
SE-002
X-POS (FT)
-5E-002
-1E-001._.
I lU
I
CDVW-I SIGV,150 R-20000 TIME AREA(SQ FT) LEG
5 402BE-00715 5156E-007 ....
25 4521E-007 - - -
Figure 5.12 Filtered Track 2 Error Ellipses, =150, ErrorNormalized
5-I
E- Y-POS (FT)
5E-002 I.
1.
-5E-002I"-I
"l I.
II.
-5E35 7401. 00
l I.
] I.
I.4 l -- 1I.
I N.
I I
I I
COVW-I SIGVm15B R-n22B00 TIME AREA (SO FT) LEG
35 740 E-007 --
45 3869E-005 ....
55 3416E-005 - - -
Figure 5.13 Filtered Track 2 Error Ellipses, a z!50, ErrorNormalized
58
Y-POIT113N TFD
4MF
2m 4
04
50X-POSITION Tn
-2m9
-409
90 I
Figure 5.14i Solid Line Track 3, Vel 200 ft/sec, DotsIndicate Filtered Track for a. =150, a V =10,R=20 000
59
Y-POS (FT)150
75
6 0
-75
-150
CDVW-100 S IGY- 150 R-20000 T IME AREA (SQ FT) LEG
501291E+001 ..
60 1388E+001 - - -
Figure 5.15 Filtered Track 3 Error Ellipses, a-=150
Y-POS (FT)
56
-56
CDVW-100 SIGV-150 R-20000 TIME AREA(SQ FT) LEG
40 3284E+00050 1420E+001 ..
60 32.57E+000 - - -
Figure 5.16 Filtered Track Error --!ises, a., 1BSt at _4st -.c s Wr
Y-POS (FT)114
57
-57
-114
.4
COVW-100 SIGV 150 R=20000 TIME AREA (SQ FT) LEG
40 1448E +00 -50 2945E+00 ....
60 5517E-001 - - -
Figure 5.17 Filtered Track 3 Error Ellipses, a :15,Statistics Window 5 /
62
S
Y-POSITION T)
4M0
07
X-PSeIO IOMso
F16gure 5.1S Filtered Track 3 for azOO 4AI =, R=20,000
Y-P STION GT)
410
0 30
X-PQSITION T
-209
*170
II
Figure 5.19 Filtered Track 3 for u :300, a =i0, R=20,300,
Statistics Window 10, Reinitialized at 5Consecutive Ellipse Area Increments
54 I
. . . . . . . ,. , . . . . . . . . . .
Y-POS (FT)548
274
X-POS (FT)
-274*i-548
O N
COVW-IO SIGV-300 R"20000 TIME AREA(SQ FT) LEG RESET
30 4734E+001 - 2940 7035E+001 .... 35
50 2698E+002 - - - 44
Figure 5.20 Filtered Track 3 Error Ellipses, o =300,Statistics Window 10, Reini ±-*aze' 5
35
,I . . . . . _ -
Y-POS (FT)548
274
-- 74
I I
CCVW-100 SIGV-300 R-20000 TIME AREA (SQ FT) LEG RESET
62 1617E+002 -- 5272 6563E=-000 .... 52
F'igu-re 5.21 Filtered Track 3 Error Ellipses, o_= 3Di,S:tatz*s-ics kndow 13, Reini:ae
Y-POSITION (FM)
2M0
40
* X-POSITION Tn
CS)
Slaiitics Wi-icw V, Reirnt1.:e 11
Y-POS (FT)292-
1461
0 I X-POS (FT)
-146
CD
COVW=100 SIGV-300 R=20000 TIME AREA (SO FT) LEG RESET
30 1024E+001 - 0
40 llelE+00I .... 0
50 91886E+0z1 - - - 0
ie5.23 Filtered Tracl, 3 7Error £1iiseS, (7 =300,
Sta,:--stics Windlow 10, Reinitiali7ze 10
Y-POS (FT)22 -
146
X-POS (F-T)
-146
I4
-44
COVW-100 SIGV-30 R-20000 TIME AREA(SQ FT) LEG RESET
60 1380E+002 - 57
72 1478E+001 .... 57
Figure 5.24 Filtered Track 3 Error Ellipses, a =300,
Statistics Window 10, Reinitiaiize 10
Y-POSITION FM)
30
5050
400*
70
300
so
23900
X-POSITIDN (FD)
Figure 5.25 Track 4, Vel 50 ft,/sec
7,I
Y-POSIrION TD
4
acn in cX-POSITION CT)
Fig ure 5.26 Filtered Tr-ack 4, a,=30, R=900, a = 150--w
0
Y-POSIT!ON OFr)
flBr
v 2
X-P ITION (FT)
Figure 5.27 Filtered Track 4, a -3 O, R=900, Statistics
Window 10, Varying a 2-72
72
34T
Il
Y-POSS (FT)
-3-4
7;
1 / .1
°. 4
-68wcc C') co
SIGV -30 R 90 TIM~E AREA(SQ FT) LEG. COYW
15 9688E-002 - 225 8097E-001 ....
-35 2740-+0- - - - 160
Figure 5.29 Fiitered Trac-- 7rror~ 71-iDses, s3Zatis-icsil in w 12, 7ari;D
7 3
Y-POS (FT)
68 --
/
34/S /
//X-POS (FT)//
-34
- 8t I!I
SIGV -30 R -900 TIME AREA(SQ FT) LEG. COVW
45 2239E+000 - 16055 6303E+000 .... 16065 4625E+000 - - - 80 q
Figure 5.29 Filtered Track 4 Error Ellipses, StatisticsWindow 10, Varying COVW
7L
* 4
Y-POS (FT)68
34
. X-POS (FT)
-34
CO CT)m COI I
SIGV - 30 R 900 TIME AREA(SQ FT) LEG. COVW
75 I163E+0 0 - 4085 7916E-001 .... 40
Figure 5.30 Filtered Track 4 Error Ellipses, StatisticsWindow 10, Varying COVW
75
Y-PMITION (Fn
I
SM r
4
2W, 1
X-PSITION FT)
2
Figure 5.31 Filtered Track 4, a =30, R=900, w =150,Reinitialize 5 V - -w
76
. . .. . . ....
103 Y-POS (FT)
10
Y-POS (FT)51
7" /
/13c ..
/ /
i//
SIGV-30 R-900 CCVW-150 TIME AREA(SQ FT) LEG. RESET
15 IB8SE-001 - 025 1211E+000 .... 23
35 5743E+000 - - - 33
Fgure S.32 Filtered Track 4 Error lli-ses, =Reinitialize 5
7-
S
Y-POS (FT)103
51
B/ /
/ , X-POS (FT)
ftf
-51
-103 -
SIGV"30 R-900 COVW-150 TIME AREACSQ FT) LEG. RESET
45 4383E+00 - 3855 6852E+000 .... 55
65 6495E+000 --- 55
Figure 5.33 Filtered Track 4 Error Eliioses,Reinitialize 5
78
S
Y-POS (FT)103
51
• X-POS (FT)
-1I
-03
-4
SIGV-30 R-900 COVW=150 TIME AREA(SO FT) LEG. RESET
75 1348E+000 - 6965 2452E+000 .... 81
Fi.ure 5.34 Filtered Track E Error 3liise, 0
Reiniti~ilze 5
7 I
Y-POSITION (MT
44
304
7u5,35 ril-er,,J Track< 4, a =3 ,=R900, ~1
Reinitializ-e 7
I, T 7 1 ___ __
A. ERROR ELLIPSE
The f4ilter error ellipse proved useful as a tool for
indicating filter er.crmance. The information "r by
the ellipse, particularly surface area changes, was used to
make decisions concerning the alteration of the filher
parameters. Several approaches for using the error ellipse
were applied to both the linear and nonlinear tracking
oroblem and the results are summarized below:I
Procedure Applicable CommentsFilt er
( Linear orx t e nd ed)
tais-ics Both sUseful in keepin: the errorWiindcw elliose current and resocnsive
to cresent data. -he ellipsereflects most recent data; 3lidata is disregarded. eneficialwhen making a decision concerninzfilt7er Daramecer modification.Normally a better indicator ofi!ter c .nve...ce or dv_ _ ncethan without a statistics window.
:!ormalized Both Aid in displaying error trends
--rror as tarD-et ar-oacne o3or nate4 ~ize axis or origin. 'r -racti:-a
in ceterMnnng fiicr conver-7ence/diver-ence l:e t: scsdellicse changes in 'ocinicy ofaxes.
1 ncreas 4rir ~x e n ei- S ceisu-n- C S IEll1ipS e A r ea ?(Klkr ren -: ef .
,and '-(-e / ) ee st t -Mocre e ewnen us-_c in.con_ i .no ion wi tnh s _t i 3 cs dno w.
Err or Extended This 3'mi-u varya y -3 CQ"
Expansion! crease or 'decrease is DarticularlyCompression tuserFl in a :_raokir-Iz environmentto vary COVW cont-aining lrc variations in the(Adaptive Q) random fcrcin- input. The proce-
lure is more effective when uisedin conju nction with a st atisticswindow. Using this t eonniquewith filter reinitialiZation Wasunsuccessful.
B. COMPTER PEF.FORMANCE
The HiP-86 proved to be an extremely reliable computer
w,;-tn no "jowntime exo-erienced jur~ng the 5 month Deriod of
,Dreration. Full(Real) :)recision was use,-- throughout the
suyprovidiing iS d'igit precision, which was more th.,an
a'e quane. The use of the Ma _ri x RCM redu-ced Drogram lengt-
J: 96 n ncreased, com-.uti ng speed by a :actor o: bu 6.
~.ug ntused in this study1 a Statist cs ROM would h!ave
uincoub)tedly ru,,rther increased cmuigspeed_-. For any
rrerstudi-y using t he HP-36, it- Is recommend'ei that_ a
7tatistics ROM be Drocured.
-cctoo arprox=inately 2 secOnds f=r eachn 4ncre-e7ntcl
tme m ea7,,rement -,:ao e cuce c ' te fil1,
'Ka nns, bhfrte liear and- nornline -I- e. T_3
IL
computing time also included all statistics computations.
With an incoming data rate -F I set of measurement Uc a cer
second. this computing speed is not sufficient for n-iine 4e
processing. As previously mentioned, the Naval Underwater
Tracking Range receives a series of 4 measurement times
sequentially every 1.31 seconds. The range's three-
dimensional tracking problem will necessarily involve more
than the 4 state variables used in this study. Hence
greater matrix dimensions resulting in longer computing •
times can be expected.
The HP-86 CRT graphics were used extensively to provide
error ellipse plots during the tracking runs. Using a "no *
frills" approach to plotting, i.e. plotting without x-y
axis or labelling, it took approximately 2.5 seconds per
ellipse plot. The ellipse plotting routine used involved D
sines and cosines, plotted point by point in 30 degree
increments for a total of 360 degrees. This method was
somewhat slow. Had there been available a graphics program *
that would sketch in the ellinse around the intersected
major and minor axis, the graphics presentation could have
been speeded u-3. But since ellipse plotting for every 3 to S "
5 increments of time Drovided sufficient "real-time"
information, the time of 2.5 seconds per plot was tolerable.
Summarizing, the HP-36 could be used to comoute •
statistics and Drovide 7raDhics in -he real-time underwater
tracking environment, if the graphics were required not moreS 4
S q
often than 3 to 5 seconds. However, before the HP-86 can e
considered feasible for real-time Kalman filtr procssinz,
more investigaz ion i: needi in findin wa173 to swe-v
computer processing time such as parallel processing,
additional use of manufacturer-provided ROMs and machine
language programming.
I8
AI
APPEfIiX A
TRC G,.ZNB.FATIO I
1. TRACK THREE
Target ovement is in the x-y plane with the tracking
sensor located at the origin of the cartesian axes. The
target follows a parabolic track (see Figure A-i) at a
constant speed of 200 feet per second.
y
h x
-V .Vx
Figure A-i Track 3
3?
0
The parabolic equation is:
!v
y 4pD(x-h)
where p = 1000 and h = 1000, resulting in:
2y 4 4000(x-1000)
Initial target location is (x,y) = (8000, 5291.5). Target
x-direction velocity is given by:
v = v cos(Angle) (A-1)
and target y-direction velocity is:
v = v sin(Angle) (A-2)Y
where the Angle is obtained from:
Angle = tan-1 (-X)LX
and v 4s obtained from:
S(v 2 + v2 ) 1/2
x 7
03
4
where the argument of the inverse tangent is the slope _f
small increment (less than 1 second) at each successiu:e sa=a
point. Usinlg a sampling period of one second, the 3a
points (x,y) can be obtained from:
x(k+l) = x(k) + v (k)X
y(k+l) = y(k) + v (k)y
Table A-i gives the numerical values for the four states.
2. TRACK FOUR
Target movement simulates a 30-knot torpedo (velocity 50
feet per second) at a constant depth. The target's initial
position is (x,y) = (3250, 5000), and it is moving in the v
direction with v =0. (See Figure A-2.) The target remainsy
on a straight course for 15 seconds, and then executes a 90
degree turn and travels in the -y direction at v = -50 ft/sec.y
The trajectory of the target turn is described by a 90 degree
arc of a circle of radius, R = 1000, with the circle centered
at (x,y) = (4000, 4000). The 90 deg arc will be traversed
in:
27rR/4v sec. 107 sec.(where v = 50 ft/sec)
37
TABLE A-i
NUMERICAL VALUES OF STATES FOR TRACK THREE
K X X-VEL Y Y-VL
1 eO00. CO -186.90 5291.50 -71.152 7813.10 -186.93 5220.38 -71.133 7626.17 -186.60 5148. 27 -71.98
•4 7439.5E -186.25 5075,26 -72.87c 7253.33 -185.89 5001.33 -73.79
6 7067.44 -185.51 4926. 43 -74.747 6e81.93 -185.11 4850.54 -75.73
8 6696. 62 -184.68 4773. 60 -76.76
9 6!12.14 -184.24 4695.59 -77.83
10 6327.90 -183.76 4616.45 -78.94
11 6144.14 -183.26 4536. 14 -80.09
12 5960.87 -182. 73 4454.60 -81.30
13 5778.14 -182.17 4371.79 -82.56
14 5595.98 -181.57 4287.65 -83.8715 5414.41 -180.92 4202.10 -85.24
16 5233.49 -180.24 4115.09 -86.68
17 5C53.25 -179.51 4026.54 -88.19
18 4873.74 -178.72 3936.37 -89.78
is 4695.C2 -177.87 3844.49 -91.44
20 4517.15 -176.96 3750.81 -93.1921 4340.19 -175.97 3655.24 -95.04
22 4164.22 -174.91 3557.65 -97.00
23 3989.31 -173.74 3457.93 -99.06
24 3815.57 -172.48 3355.93 -101.25
25 3643.0c, -171.09 3251.52 -103.57
26 3472.00 -169.57 3144. 52 -106.05
27 3 02.43 -167.89 3034. 75 -108.68
28 3134.54 -166.04 2922.01 -111.50
25 2S68.5C -163.98 2806. 06 -114.51
30 2804.52 -161.67 2686.65 -117.7431 2642.E5 -159.09 2563. 47 -121.21
32 2483.76 -156. 17 2436. 20 -124.94
33 2327.59 -152.86 2304.42 -128.98
34 2174.74 -149.07 2167. 70 -133.3335 2C25.66 -144.72 2025.50 -138.05
36 1880.95 -139.67 1877. 18 -143.15
37 1741.28 -133.78 1721.95 -148.67
38 1607.50 -126.83 1558.85 -154.64
39 1480.67 -118.59 1386. 61 -161.05
40 1362.08 -108.70 1203. 46 -167.88
41 1253.38 -96.73 1006.73 -175.05
42 1156.65 -82.01 791.58 -182.4143 1C74.6 4 -63.44 546. 41 -189.67
44 1011.20 -37.24 211. 63 -196.50
45 1coo. CC 0.00 0.00 -200.00
88
TABLE A-1 (CONT.)
K X X-VEL Y
46 1011.20 10.57 -211.63 -199.724-0 1C21.77 25.14 -295.07 -198.4148 1046.90 35.82 -433. 13 -196.7749 1082.72 48.89 -575. 23 -193.9350 1131.61 61.85 -7251 57 -190.2051 1193.46 74.49 -879. 69 -185.6152 126 7 .95 86.36 -1035.28 -18C.3953 1354.31 97.25 -1190.48 -174.7754 1451.56 107.04 -1343.96 -168.9455 1558. 6C 115.75 -149U. 80 -163.1056 1674.35 123.43 -1642.38 -157.3757 1797.7E 130.16 -1786.37 -151.3558 1927.94 136.06 -1926.60 -14o.5859 2C64.01 141.24 -2063. 01 -141.6160 2205.24 145.78 -2195.67 -136.9261 2351.02 149.78 -2324. 67 -132.5462 2500.80 153. 31 -2450. 14 -128.436 3 2654.11 156.45 -2572.25 -124.6364 2810.56 159.24 -2691.14 -121.0165 2969.8C 161.73 -2806.99 -117.6666 3131.52 163.97 -2919.95 -114.5267 3295.4S 165.98 -3030. 17 -111.5868 3461. 47 167.80 -3137. 1 -108.826 3E29.2 7 169.46 -3243. 01 -106.2370 3798.73 170.96 -3345.88 -103.7971 3S69.6s 172.34 -3446.56 -101.4972 4142.03 173.6C -3545. 15 -99.327-Z 4315.62 174.76 -3641.77 -97.2674 4490. 3 175. 82 -3736. 51 -95.3275 4666.21 176.81 -3829. 47 -93.4876 4843. 02 177. -3 -3920. 72 -91 .7377 5C20.74 178.57 -4010.36 -90.0678 5199. 32 179. 36 -4098.45 -88.487S 5378. 6 180.10 -4185. 06 -86.9780 5558. 78 180. 79 -4270. 26 -85.5381 5139.57 181.44 -435U. 11 -84.1582 5S21. 01 182.04 -4436. 67 -82.8383 E 103.05 182.61 -4517.99 -81.5784 6285.66 183. 14 -4598. 11 -80.3685 6468.80 183.65 -4677. 09 -79.2086 6652.45 184. 13 -4754. 98 -78.0987 6E36.5E 184. 58 -4831. 80 -77.0188 7021.16 185.00 -4907.61 -75.98e9 7206. 16 185.41 -4982. 43 -74. 990 7391. 57 185.79 -5056.31 -714.03
89
*Yx
I ..y.. V
A-
0
Figure A-2 Track ~4
So each second:
2-74 =.05 radians will be traversed 4
a Using the trigonometric identity:
sin 2 (A) + cos 2 (A) 1
39
And the equation for a circle:
2 n 2X +y =C2
It follows that the arc of Figure A-2 is described by:
x(k) = 4000 + 1000sin(.05k)
y(k) = 4000 + 1000cos(.05k)
where the angle argument is in radians and k=0,1,2,...31
seconds. The velocities v and v can be obtained from:x y
v (k) = SOcos(.OSk)
yv (k) =-50sin(.05k) 4
using the same angle argument.
The track values for the track arc are contained in Table
A-2.
91 3
J I
TABLE A-2
NUe1ERIAL VALUES CF STATES FOR TRACK FOUR
K X-VEL Y Y-VEL
12 3049.9S 49.98 4999. 38 -1.2513 3099.S6 49.94 4997. 50 -2.501a 3149.46 49.86 4994.38 -3.7515 3199.E7 49.75 4990. 01 -4.99 S16 3249."5 49.61 4984. 40 -6.2317 3298.88 49.44 4977.54 -7.4718 3348.22 49. 24 4969. 45 -8.7119 3397.34 49.00 4960. 13 -9.9320 3446.2 1 48.74 4949. 59 -11.1621 3494 .1 48.45 4937. 82 -12.37 •22 3543.C . 48. 12 492U. E5 -13.5323 3591 .C4 47 77 4910. 67 -14. 7824 3638.62 47.38 4895. 30 -15.972 3685.8C 46.97 4878.75 -17.1426 3732.55 46.53 4861.02 -18.3127 3778.E4 46.05 4842. 12 -19.4728 3824.64 45.55 4822. C8 -20.622s 3869.93 45.02 4800. 89 -21.7530 3914.68 44. 46 4778.59 -22.8731 -958.E5 43.88 4755.17 -23.9732 4002.43 43.27 4730.65 -25.0633 4045.37 U2.63 U705.05 -26.1334 4087.67 41.96 4678.38 -27.1935 4129.28 41.27 4650.67 -28.2336 4170. 19 40.55 4621. S3 -29.2537 4210.3- 39.80 4592.17 -30.2638 4249.7S 39.04 4561. 41 -31.2439 4288.44 38.24 4529.68 -32.2140 4326.27 37.42 4497.C0 -33.1641 4363.28 36.58 4463. 38 -34. 0842 4399.43 35.72 4428. 84 -34.994- 4434.1 34.84 4393.41 -35.87u4 4469. 10 33.93 4357. 11 -36.7345 4502.56 33.00 4319.97 -37.5646 4535.C9 32.05 4281.99 -38.3841 4566.65 31 .09 4243. 22 - 9. 17
:2*
11 A-2 (CONI.)
X X-VEL Y Y-VEL
48 4597.24 30.09 4203.67 -39.9349 4626. E3 29. 0 4163. 37 -40.6750 4 655. 4C 28.06 4122. 34 -41 .3951 4682.S4 27.02 4080.60 -42.07
52 4709.43 25.95 4038. 20 -42.7453 4734. E5 24.88 3995. 14 -43.3754 4759. 18 23. 79 3951. 46 -43.9855 4782.41 22.68 3907. 19 -44.56
56 4804.54 21. 56 3862. 35 -45.1157 4825.53 20.42 3316. 97 -45.6458 4845.38 19.28 3771.09 -46.1359 4864. C8 18. 12 3724..72 -46.6060 4881.61 16. 95 3677. E9 -47.0461 4897.S7 15.77 363). 64 -47.4562 4913. 14 14.5E 3583. CO -47.83E3 4927.12 13.37 3535.00 -48.1864 4939.eS 12. 17 3486.66 -48.50
j E5 4951 .45 10.95 3438. 01 -48.7966 4961.79 9.73 3389. 10 -49.04
67 4970.SC 8.50 3339.93 -49.2769 4978.78 7. 26 3290. 56 -49.4769 4985.43 6.03 3241.01 -49.6470 4990.83 4.7E 3191.20 -49.7771 4 994. 3.54 3141. 47 -49.3772 4997.90 2.29 3091.56 -49.9573 4999.57 1. 04 301. 59 -49.9974 4999.98 -. 21 2991. 59 -50.00
9
A , P -H
COMP'JTER PROGRAM EXP LANATIDN
1. i±HMAL
The L7, KAL ro-gram comcutes the filter ?A:;(.2)
for the 4-state system and stores the gains in "LINGAIN
.STORAG". The theoretical covariance of error matrix, 7I_,(4,4),
is also computed and stored in "LINCOV.STORAG". Several
different sets of gain and covariance values were conmputed
and stored for various values of measurement noise, RMAT(2,2),
and random forcing noise, COVW(2,2).
2. 1,!NE ST
T :he LI.EST program retrieves the appropriate ja nn
sc.qedule from storage and computes the optimal estimate,
X.AT(4,l), and the optimal one-step prerczion, X.H!l('-I)
The following capabilities are contained in the LINST program:
a. Gating Scheme
If the absoluze difference (DIF-) betw een the one-s* ....
Drediction, XHKIK and the noisv track value, ZHAT, is greater
than tne three-s 7ma gate, then GAIN matrix is di..sre.:.
and XHAT is set ecual to XHh1IK.
. ra ck Noise
'i, :ITA AiK 1, the r-ndom rumber aenerator,
,ID, and the ..- utin.. simulated, n'i2z - v1 n: ;,
4
are bv!,:,a ssed a nd th-e -r ack -. alIue s corotd;iti7, noisl'UZ
arc etr e f-_crn" dft fil 3 RiKS " NT: 7-7.
programn run, producing a different set of noise values
resulting in a unique noise corrupted track for each run.
C. Stati:-s t ic s W indow
When WINDOW is set to 0, the filtcer stat-e statistzocs
are computed after each Monte Carlo run. The error mean,
variance and position covariance are computed. If WINDOW
is set to an integer, I, such that max k.>I>0, the stat istics
will be computed based on the data compiled during the 'Last
I iterations of the simulat:ion. If, fo r e xam ple , 'AlINJD 1~
the computati on of statistics will be based on the c ata
obtained during the last ten iterations of k, and al!
:)rev 4ous data is disregarded.
d . Error 'TormaliZation
By setting NORM = , the error (ERR), whico is
de fi-Ln ed as the difference between the tru-e t-rac , valu.e (TR.AY)
adthe e stimate (XHAT), is normaliz7ed For all other value; s
of NORMI the relative error is S Sed in comp utn te
statistics.
e. F ein it ial1i z ing EO7a_4n s
T he program has the opt ion D." reinit iacizccg toe .
-to G0) T h is c an b e don e by7 set-_tin -L~ n- Hn nt et
suco, toat max JI .. If the sur* ace a-rea
ILiPNOT OF NIL-flU ERROR ELLIPSEW AN M fW-OF KELIMN FILTER PEWFORUANCECU) NAVAL POSTIRROUNTESCHOOL MONTEREY CR J JANOS NA 64
UNCLASIFIED F/G M4 ML
EEEEEEEEEE-E
-~~~~ - - -- - - - -
------ i 332 1112 20
L
1.25 ~ 1.
MICROCOPY RESOLUTION TEST CHARTNATIONAL BLIR4AU OF SANDARO 19bi A
increases I consecutive times, indicating filter divergence,
the gains are reinitialized to G(O). If reinitializing
gains is not desired as an option, then HIGH is set :o some
arbitrary large number greater than max k.
3. EXTKF
The EXTKF program computes the optimal estimate, XHAT,
and the optimal one-step prediction, XHKIK, for the 4-state
nonlinear tracking problem. The EXTKF program has the
options: gating scheme, track noise, statistics window,
and error normalization as described for the LINEST program.
EXTKF has the following additional options:
a. Reinitializing the Filter
The program has the option of reinitializing the
covariance of one-step prediction error matrix, PKIK, by
setting it to 10nxPo, where n is zero or some small integer.
If HIGH is set to an integer I, such that max k>I>0, and the
surface area of the ellipse increases I consecutive times,
then the PKIK matrix will be reset to 10 .xP--0
b. Adaptive Q
EXTKF has the option of automatically increasing or
decreasing the state excitation matrix, 2, under certain
conditions by changing the value of the covariance of
excitation noise, COVW, INCREASE and DECREASE are set to
integers I and J respectively, such that max k>I,J>0. If
the surface area of the error ellipse increases I
96
consecutive times, COVW is doubled or increased by some
factor. Conversely, if the surface area of the error ellipse
decreases J consecutive times, COVW is halved or decreased
by some factor. if COVW is changed, increased for example,
the value of COVW can be changed again later in the run,
if the criteria for either increasing or decreasing is met.
If the adaptive Q is not desired then INCREASE and DECREASE
are set to some large number greater than k.
97
APPENDIX C
PROGRAM LISTING
*N
=#-4 %%w 4*
mw t t - * *.
o)g %* 04. * *-uON-0 ft) * 0
U) iii a a* *0 W
trk 04'-* W4 4.N 4
~O) %4 %*Z 0 4t 41.v lb to. (o
4pra %C4 O OjOc %M, IC
+10 404 Q') %MM 01 o*AI(
06 0 %4 W 0* k q-4
aL w 41 0 41SrqCV~- 11.4 SJ4 llt1 *0 4- WO*aLI-i
"q4. o~'.r 0v a1 Mv* rv 40WW6L 4 0- % . n LH a aa* 0 1 JM J b40401 41c 04 is0 4=I . i C79 # m Inr* VV4 M thw
w (U 44 - ull l b 1M- 0ql p a0 MntPE "-4J o 0 00 CQ : tU1= U * Q .) *W0f~ 4 t (C W1 %"CV %c jA .1 * * V1 * M140+G) m 0 . i~l3= V ) ft %, a. AU (*.,I 0lP A4)L=
=. U.i'JO4 - 00o% % % + =0a -40 Wi )fiU to ' 4 .4 %0 %0. 3 4 '.1-44 #UO 4 J Q'. ~ 11 W
orf~ fC*'U)4 0 04.40%- * 4-4 *M AC-. 4J 4w.7- %"* *zz 0 Uor %aK tifB 04-rI' o+J(NBWBq 1.4714 0"1 f
+' o %24 Wr i - C.0- V%0 k%0 %* 11 4J It CM V '_ rl u* r''
o~~~~ -o CM' 3 ON*QO. 41*4 NUUJ-' **Y).4 4
1.4 V _ aU U4-1-aQ (d. %to P 3 to mu) to * co 8 a V * 1 ci wai 0) T4. 04* *70-4.C0 0'4) 4)bv Wv+i '.q W. 0 (1) 4 CO*0 0WB* M. W. (04) 00 4J "D40*V
toWr U&4 0%+'.-) C W W 04- Wt $ 4 Q 041 B~i-W A*4~tW
ig. u. 0 4 4- toc- &' MU to * . gg Ig Li *r.Q+.JJ4+* U4 tilU Q(3 4-)0 4* 0 - 4N 4-*W
-A ~ ~ ~ ~ ' o4 10 d o1(4U)t(0M1 0CCQto M o * ' (0 0I* 0-4CC o( t~-o o* *.-1t- 0 MO 1 0)U 1USU 110 fV V q VA1 v 0 0)4 U)4 NIB it)4 6- %4# l 8 I 0c: 0f
U.0.c-'-UG) 0 .O .U(0tf0410.-UUO1oc93a400 o4*
'- .*0'4b41'W' W'0 C40* ~~0)* .~C j49.80
as ,)4J 140 4 ' U10 t - *-414. 4- Q CU +J fl
Wr-4 M *14 -Cs 0! .40~ adq 0 (11 -
0J44 Ii*$- V 4 2 c.) (1 l 0 C: 0 ) Uwo a) ( Ou*0Oc: 0 IU .1 %, U.
(d0fd 4 .;P di H. f-4wO V 'D - ONof-I 050 0 4Q to .
-A X r. % %* WS 3110 1) 4.4( f/r4 0ri 9 W.m0W
W..-A 94 t W *i.44-4 Ad 011 " a ) -P-tI-W+)S.4 41-M~( m0O o 1t U 1tI M54W W0 144r4 fV.0 fl ( CLS-IS9% a,4 ('4.U A Or4 W,4 -CO
0 M0( M l etX4 1 I 4J-r4 W0) 0 Fw 4 V -W+ -a4 o '0+) :PC w t:0 u f0 'P a W'4f W~P tr 014
*4 Ara - t cuoo NfV 0 D W) 10 3k -'P.X P..+'
CUP Ar" T o0uS0gm .,tg (CO 4 U4 ' U.~ i u
to0 U-. *CJlWtY /U P CuuW INW C%;;-4~U- P(JO oi :.
wO DoC MC4 a 0%a43 .-H-4 W-4W.WWC$ W'H a +) 0 w 40W
U 4J -W .....- 0C'a f-Ce-4 P M J2.-' W4J'U W 29 bP I.IV W 0 Wev A4) :w~ ' 4'0 (0 +) 0OojOU) wer4 oj)U1)
M t toCwU ~.Ui?73w 0 Q Cu LSJ W4JH1-4I+-rt W Wr-Il Wa 4.) -. Ur*5U VIA0nW1.1 W WW~ w -ON14 4-4 4) vq MH-
-HC4) W % am . 0waffolp -4 W 0wQ)4.4Q 0ow u c~4.-4 V(14 %M 4O4-04J 0 (uv-4 w 0&.O10dL4V 0 M
WW .,0U2 elOU) w- PJ(M 01 0.4' U)W r-1 WCot ..-. ' r4) 0 t 4-4 0 -..-C- 4 Q4114 I WA+J WP 0 W-90) a to % e44 00 $.-4 0-4UW-H (V 4J 4) U.=-.-
01= W o aN5-4U00$-4 30 00 C)r.CV+J U D00t)4..W4. U 04 UN WCOW 0) 4 WO-C.'=I ri. W(U-4 44U4
0 . %OpW oo 44444J4-.UVI4 U,-I%4-1 U~ W rHr(Mifl 9 u w- U p4-:2 3r34+I. *-frl.."-H 1 - 1~. 'p i .,-5+.) 0
(~l C &' 1F0 %113W 4 +fVua WfC)U 0 w QI*4U Of.) a W-= I I Z u C XZZOZr.-4. C,-4 9%43V CPU) +a f"44 w) la MO 9 14 >- I9-w- "'+J0+q JC4-4 "Q~~ X "'T
a. a- tl104Up' %W3;:44o
f/SId 4 X t %0 V, t0;2 .19-CSI.-9- RIMW 9to 14 %cr %(J, (. 0 q9-5-dL *4ja)* V_ - .Mai ) ( (1 ft* OrV~ (d moo ii 110 U, 0 C)
0+ ~oa0w0 0q0 q_ It UN II I 05.4-4-f4 -tCM Cu bo W 06 10 Ad 0~M I 0M.4A it
*4jQJW4,-4I4 CuA 0MXX 01111 4Wvr-I to 11 3 - () to 0- iV1 441 C NCL)11u 9-0cn W UA LI. W- U 4J.' f-4--'II11 -+'+
0WU '-440 Hv-e-4 IN MSUMS4J +JAQ FO U01 I 11 tv if _ cw tioMAr% I o1 11) It V4.C+J rd *0% Q w 0) (355 0 W W 0
'r4.4 0 WuJ0 0 0 W 0 0 *M Of-4.-44) 0*0 4 W4 -1 1 W-C+') W .f40O-I-N
v-14i0 0.41 W W WW0IA 0 00J 0 W CMV CQ-29%4r U Q00ww U-OW4
00
* **
to * *tr CI*
* *
* .**- * 0* r 4 0*
III Q lQ L 4 v Q A * W* V
*br C C . A *0*
441* *r4rI "0IZ( N ~ . -* =* fr
+1 ~ ~ ( 4j Q Ar c1- t.VUQUm * *W
%-- I(U--9* O% WD m-- wI *1 *N0-Cid 1)N ) f * Ad 0'f ~ f4- tI
if if Q * A fn)LMaO+ Lid W* .,4,1w*
0: a' InO (a * 4 1 )t ) IUQ4 l-0 ! 1 ) Udr% 4)t U ) ))41' 0
4)' 1.=49 0 r3 a.0 93 = l0w uw*M ljowom cd f r* * 11 1 '4
0VJ0JU 4.1 Ul66 W 4 t4 * * m (
*4 00 m 0 (04 m p** w 1- k
(N U LL W( '1 *- 101 *I
C S
0
*j 41* 0* *
*0 4* * * *4
*~~~~ * w
9: * * * * * * ** ** ~ * * ** * *
* ** ** ** * *.. j -(* * ) 4J to * * *U - * * i-* u * 41 wr.l) * W a~ * -* %* * b 414. 414 ..m* 41 * * 4* 441 r.
41 N*W* 0, W *Ad * 4 * **4 * S,* AI * * * 41* 04-1441 1 *d . *3* 4 4 * w1 to
dl ftw* C1 4J * 4 -*4 4) VQ:".- 41 * 04 9L *. 4) *) as41
% 1 * 4 * *4 41 4D sl o(21 41 * *VJMJ r * 41*4 0)O- 0 4J N0J U,. 41 * 4- w 414 *4 * 14o .- sU i
*11q- 0, 414141+ = * *.. * 44q 0 CV lW )14.d * U 41 41 w it 41 +-, 40.D' r
00*. *1 U 41 1 *1441N VI
M J -W10 4 0 MAC *l M* 41r. 14*K.4 lb %%14 M UQ"HM d0 tfC4* 41 4.1 M4. N1~ U=I-e4 - W 4W UD
N aU * tr 11 *4 X 4* 41*B U 4-J~ ()) ' 004J4" * 41 A o1N4 *I 41 to v^1 U W
14U441 ~ fn * 4-' * 41* 21* I^- 0 11 x + 0
't -1414 CO * ...8M 4D*XU * . * W P1 w rz*NM a*NO *-N 444 1AS 2-I * (04 *WW 44.S4r *t4q w=4)
4J Ob* * '44JX * di *14* X * 4)0f 4) 4flUU WE9 rO 444* *~0* m -N *1 *J4 0 V= 1*1' 4 .' 444 * 4d4J4 44'U * U) 4 1+' *44.'0 4J~ W. X+ J-rl 4' U 41~~
W *H4* W41 G M *-IW M* *0d 41 W* CC~1.1 $4U4 MV N M 4+...M640W* alo 6*41 ~ 4 04 0* 0*V M M41 4 0 W +' A*N M I
4-Nm *0%of41 * CDC% a"* 41.. 4nk U4a r 00000000JU) f0.. WUH141
0 -. 1m~*04 ~.1 0 4r-4%01 1 02 .4*~0 ~4~1-0 N 41
HI+* * * ** * * *)
* * N *4
* 0 4- * d
of 0- 41 vi VI-,:- * W. U) P
%* 0d VO 4J V A
0- C.2 *' %", rc* Vr IV ( 0 104A
(4) %00 r IM o0) A4* .54 4 0A9.4l wo oW 00%4 )C
9C4 $414 t- ri -r 0 0 A C- (1 4
WWv 4J t1 1 1-4 01 &A4
D- 0- w44U
W Cr0 0*r-I J W L-r*)L I 1 I
IfH WPr 4J r. 4-4.) _-r ca. --r454J- %. -- c -. 4r' Vh Ad xi A *
WEI 4d .- It W~4 0I~ 10 4 14 t: W 4)l 9Jr -- *5.4 * LAW= H- 0 0v VI -r JV
9z 4J U I 0 4J ON W 3 - = (0(1)* 0 ' ), 'H0 0 T-N (- 1X U-4 '--* SH l--
0 0J jj0. 44* VU 0 ..4 p *.- 4. 400 M W* W 4WMWX~WP*WX 14X
to10 +11 ()- 14 v lii V *.9 )44 0 W Q -Q ) W * W Qbo . A* 44 RI- '44 r.~ V%4 (~~l%- c: 014 r- 4- c; 4
~4I .. 1 0 r14
r4 ~ a 103'
0I
** * * 0 *
41 *3 A * -4
At * 00 *- * 3 *
H *) r.*r4 * *1 W* -
a *) 3o *0 *) 0 r m* 3
*3 a * *. %U) *t (d m*Q-1 -4 * * -4 a trtm r
631 X) 9* %Wt w 831 rf4 - W- R* o ( . * =1)
"Ad 1-1 to * 19 w v =c
Ad U) %3: . ,+ 9 i * - +11
=- W ( * 0WWto0M a +d U 1-4 1A-4 w to rr 4JW
1!1 m fa 1=W) i - b 0 0..itr r v~ 13to 4 U +) - c 30 4J *d M-fl t 44D .
w* a Ill *0 .w x a U) 3jw X ) rW- '0*- 0-.Of %..r- P
(U( 4J Ag - W U 't4( 0' M IWWW110.. ) 0 %- 0 N 30 x.i a %A4
3. an 0'C. 1.4 ~0.. U& of vE -.40% I Co64 a In It 1-4 a 3 u It
-. * Q.I "O +0 W U) %.M .4. U*
J4 % if V- 1-11 A 64 %, . 3AM I 14 H
0 -r *-V U--M * 3111 0 *0 % 0-H- .r4M Ow.A*"r+ % -)M * If-= -P *Jr 4M..4 +) %O +3t4DW i
4 -W).- W1- Wic. +JACQ 10 M~ '0j-.4 4-1If.
IU) V U) I- 04J Wr G;- W - -aI U -P=
a(ID a) cu (U W J 'w 0 -4 :3 W a (a 4(a je-It 3- 11= 1 Uw() (a 0ft, 0 0.ra 11 3 11W *bMfwIV( A
Ad ~ ~ VI Vl40 4J + 1 ot .4o* 41 4* 4, 4*II W1 0W 4M M W m A W 144J** W MMM W bX- W * C6)1.1.
-r14 0%4 r. S.. 30 I 4.-' 1 14 R 4 4 R '44 r.4.- *3
CD .i 0i=Wa '4 % 0'0 *j (p 03 41co r*O 3*OC..-4-.o- % 4m 0 C2 0 3a0C(N(n t 4 r c a%~ 40 (4 z7 *) %or.. %4 v *-t V d *-' DO 4q14 ) 7V)% -W o *N a INI%0.4 c a% a% .cr ' -r V' r'. 11..' * - - P 0 oO c o o0 co 00 00 41 % 41 0 (J r 41% 0%a Is N00C
9-) 9- w 1-j:; *- q-q -q -W -q -9 -V q- 4 W,- V. ". r- r- I- *.r --P - 4(
'~ U 0 E (UtU-4 MI-0 ~.'4( 10~41 U * N UU M.
IIw*
0*
4J
+ * )
4.( 4*
(a( - ZX0) * J4
Jw to 00 *__ % r 1 U) m
wJ Vj 0 r.
a:* ~ ~ A o M I
04 * .-4-Ab4.)f 4)4- *4c 0 "-
% o(a U) * cl 'U) 7I Or - 00t 0 -I w w CMJ x IvI l+j4JM Ha
NX XMI 0 P-0QW W R-+4 J+ %. -.4J c
0 N M 1 *.0 p-I9M 0 H-4 V 0 41 J PU)4 -. + ri 11 Ad4 09 Pbb 0- lC
ol U44 .4 W"3 V, aI UU)-+ UV(rN" fa v
ja % t 0 Ot "1 %J iII (0MVMo 0- 0*00%-1+ )= D.W.C ) It( 09 6-4 I t'IW99 % f 1 n- ItI P .
to ~ I ,q 0%4- 4-'$- b94 rxU _* ' ,-I54-'I (INQ CJbW4 ~ 1 *(4 0 1 - fo ("m ucjd.0
UfdU 4-) 4Qj - 4.'OW(0-)+- 0+J'-(j 4-' *( (
93:0 -:0h 4- 'W 4 a.41 ~ ab 0 N.K .-r QV J(
6 0 0 0EW U 0 4I 4 J.A X.0- P6 81".44%4O4+0 044.40
NNNN C- N NN N4V.0 N 54NN NN CNN N""N N"NNNN- Nc"N
Hul 0(~A ~ II .~4J~ m r~-4 0 (10.5~
3o x
v0 1ge% 0o
% (1 ()00 cu-w
-7 N "M c m0b 0: 14.41
N w (aC old 00) u r- .-e
o Ini ."- 1"0-U,*-i 00 :gt- a I- L ) crr
o -Nvt 0+ UU M.-% ) (fHU o, Cn rM ~ t
Lnw ad W.. 0 N, t * . r -9w 0v ft2 . 4 Al t=: ,,rft- bi w ~ (9- .9 :::-1I) )
q- 0 -tri 4 4+ 4) A 4-)(G r- 000W =-- Dr o 0-'- V '-U oW U
j ) q-4 0j t. go W U) - a !(Nr4-
(P~~~+ WO JWtJ.. '4NN
(44 to ) Q --l LVIU l4-t4 XO.. U)Jto AjQA
r-4 Uf~Q w3 (M~%. %r1* * - - AUl 00 bQ -ID0 ..-. toQN 43 wa .0) w 40 %WO .-443 U):)t * VwM% -14
w II wvD 4 W-CD V 1
-': UW)AUJI % W1
41 H" CU1 '- i uf = q- vi-+'t 4JtUo Wal -4; t==O-o R rvj* "~ %u %4--AC p w V +, g; ==:
o 0 m I() (Ne - F44s t -n -
w~~~~ ~ ~ ~ ~ D(St f 4 * J.4W 00 0v4 )0U0U0Ut7 A- oTU1 f tLALJ-mw 1 MW r + QA-4 M 9J-I4 "i-'*0Ulr.% = U) -9040 It 1
a W-r #j 4+ 00 1W 0t .dUrf 4-.4 V "4 WWW % C
M W4 4J() +11 1 c *rd(nu x *( :.4 + * .w % 0 W4'Q' * ~ tr0 uIt0 ,* ) Y 0I r- A 0- gA Ut.- - '- CN 0 V t MI t f-l P W 10N CO r- )
0tmo 41 11 O 1M 4~) *cUW 'Jr V W1 1 0 9 UU NV 11a( %- 1(400
InIn0- 0 Hri* - to, 4 *..*-+v-4 t-4 #-4C'4J, 4 4 - ,4 f-4.-4 w I)(N 0to ,.j-lid( x( aC6 1 120- w *1 w~~a U 0U)0 0) 91- (V ~ 4) tr9 u '-(N'lq4V 0U
11i p J( U) *U.. m3 CP0U U00QA. OOPtit :w'0pQt V : ntW444 Q. f 0 W W44~3 4J 4 U4 MQto 0 tV 0 t 11444 04.4~ OJ-4*J GLW4%4- l- 0"444
M 4o't-+j tAE II:*+ "0%? U)OUX*g a" + ar- ar-i 4 U\ i* g*H 4''.i,~4r-4 .4-i
1414 0U) AA 0-~AU)(N~o'*~M~or-AC'14(.0(N
0'I5It4.H~(NI OO v~t-gI I VO. . . . U NVII 900r4 fl(N o0(3) n -4 0 Hr4r-44 0.fr- 91)-4 vI v- .-4 ..-443 )~-4 ..-4-4 -.- 4 Ii '4c40
0 .M1. 0P400 .. I 00~- (3 3 ) ( 1 )4 0 I 3.~- 4 '. .
l~lt0 +'00 ~ 0 Q~Q0 4.' U)0U4-
.14 44
0 4
(0
4 J 4 -
Co co9 r -
r- .,4.H) 4
N U) In US
0: Cfl H*r4 * ) 0W00 5.*4 0U) ... dj LLA bP4
US 44 - 1 4 In Cr-4 ONO) CL.. 0 ti
070 14 . -i-.-ri US +)4 H P-.- 4 ~ ~ (. (v'%~ a-,4-4 '.W ) 0 -
avc ~ ~ a4-u r.QCC -
:: ft U 4 J t'4fn:& I .* *Y .0 c* U . 4q:00W001 0 0-r4X >.X WV-~ .0 .14 00000000
I to P I fi It 61 JP~ of vI USSlp *. 44+++4
011*0 4Jd+ 11 4. 041 II -- i r"zv rrJN IV) m %r- r 10 r- II-dDIIIO0'Uff g 4- -14 U j ~
OO W Hir- 000cX;0A 011- f-0 Adf0 U -4 UU41 0 4.' 11 U u 0 m$'4..
'44 U44 0 iar.- %3-4.444 Q W Q.-Ir,4+J4-4 %6 I 4 4-4 .43MU US US~ US W US US) USru"448-ri--H44r- OV V rri00 W U U)U S()U) US R
Ma% or- rqmwc)ODocoo~~C ' ~
C) ~ 9( 0-H 4) 0 44S .-
r= Sd -- 0U) C 0V 0r0 4-' 0.. 0H-4 L 4.to (1 t- % 0 .,-1 - rW C) -) WH 4
0 4-) (NI4e~. di 0 u 9:1 0 +j m . : i- 4 [-)
C6" 0... 0.~( U -4 * W- 0! *-4 '- l ,(11 J0 U-1 C%' mp- .1 1' .,-1 )ul U) ON, I ff2 4- 6:
W -- j (1) Q (u l . X U) 0) (1) IV :x *-(ato * 8~i~U %A-D W b ) -rr a) p0 5-4J (421- d> Fs %a 0.tje 31(l) TH W-( 01 +1'-011 1 1o W) i-i 0 ...- &-' 1:r %95 rO () 14-) 0 r.-J-4J 0*4-J' 1= W)
ur U4 ( F-1- % for 4-) -) W- -)- (04- 0 4- r-1C.0-4 Q)H
u;W U) (o' -- .flU0 m0 a00 4J- W '..0 W:,q0)O--0 - :---U ) w *-4 ft ) 0 u. 0 4-4 mm M' M . 4- (1) 0
Q1 "~ --Z4 %W-- M' W-4 (VWO) r- A r9 .>10 z0Ab(U - 0-. r ("1 -- n a.- %0 w 202 0 -,qoo 0 Wb. 0 P"0.
Inc. UI~ 0) (JO *rt4H 4-) 10 9 1-0:3.0(j) )4-1(Ua . o ' . ~ T t .0 02-H -r-r-- i (o a) .,-4 (V1.0(11.0 0 : 0$4 ft O '.-I%0) WIC (A 0 >. .- r-Ir-4 W 4J U~ 39 0
-ri4+ % . '- %-.'..T 0 6,W En 4-'W (D (1 W)) (1- =-,- tW WoLdO-Hj>,tv %.T.- .4tvC04 *dI O1W I 4J QA0)04-' .a-U+J>.U
LAO- '. %=Jlu W ~.Wl " CPU. sMIj W -OW -H Si (4.= 144 "I W Wlf14'JF is '.Wo( V.'~ t QUCJ0 vr1+OEU(a0) jr-l US " :0+J
U) (U W- cr%'' -- W 31: x f7W2-r4 r-' -W 1-W4 0) U(J 04J (Ivl~ 4 C 0.4'W 4-) - QA4.4. (UW4JoO(WWW ,0 (044-H U)W .4 *r'4
+) V) 0CU( 8 0 0(+ 04+ W )W4) *.0 40)WW.W 0--0W
OW c rato mm.'-0 0 ) .- (1. .0 '.(-or. .0 0.- 0.. 1-(0 - .0 r- 000 +--'~)- ~ 00) 0. 00) 444-1 W -- '-- W = WaW ,,1
aU %U.~(.4 Wt-I (U(U ~ .w-4 04 (0 (Ubd 0-.Wj U '*- 0 a- 0+ 'I-I)U..-4 j .-4.
0-4W 0 -2 ,nr--4~ ~0 M14-4 4J r- Mid U::I. If W0 0
u0 .0 ow 04 Ow. +'. =Irf MaO W an . W~.nm .. . nn ,
%-- W- 1+ CO~ (J)44 W -4)44US+ -W J 44 Q- 1+414 0 0
(4) - -0 r=C 3 (-- r.+f r(+40 U 0 4- 0-W 00 41 Ac 1 QN Cr - W Ns P l 02 U l."- (a (Ji W I m (10t
c .4. +'V-UU~W'1) oi 0o "CW'-30 4 l =II -r S4 clC - r4WrJq, -V.
01(a (A W aa o =3 01- wt r- PX>-I 1 (II- ZZ-l4 4-4 '-ZW HVW-V 0'.a' F-Q. 4-' 0t WJ HI- Sit .- 3QIjr -Ce- ' 0 ~ ( X9. %4 W. ft-j0 % (0-.04'
1 00 j 0+000000m US 0 0000 0 W0 0 00000000oWm o~ouol'w 0,-mOP')W~W it it C) o(0
44+ r.N'4'( (N WN( wx o ) najr ('q ttoi ( - i
-W r- It 3 00-1 (a 8
44 * -1.4 47 * 04
() .J) * 1CC: cn*
U) fu 0(0) *Io
(v* w
w (R0in rJ 4-u
rA U
-~t U,
+)t m-N 7 0'-40u ) 0 W(W1
. U) tb9 ) - U Jr
0- *n0W0W ; - -
5JO to+, t N itu (N (a c41D U)0 000 4-' .'1 wU Is "
t4 ,4- @4cc' to r-10 -00u00 + (a+ 4 w 0a amur9
QU) 15X U) 14 -H II U r-4 WU =* 4JUJ.4~jIID+. 4 J H to- 4-04CV J3C4 W.M W ccri = 0' .="Wi-) :PC-
* .W4-1 4.r-0 0 4- mr- 00 BbcUo4 Q-. Qj ~w ox CD0 V) *-- + U (a0to (.L1-4::::-)-'-0w (D 11011 MWL
w N () 'iW.dgJv0 -- U 1O 4-)*)u) 0C N NVOUOUO-fl I44.'00 If CO 114 11
4*4J4J4 +J 4- It X4 a I I 11 If It 14J * U" -- r-1 tP -If )I O il Oil 1 It 11 H * oo444 C4 C6 ..4-'
V* 0f:: Lr -~~ Q f- 4-) .... *.... ...- * Hr, u.4i a V%'0 0 C) OH - 115U go +1 4)l' -A .4J 0fjN-TNV- *44r.4r0 oj)i
Un (d M1, WU4*V*o)-v.-+HN 3 %ll* --40 4-- 1t(13 tOy" ~ =I)c1E~ =' Mtn'-N*~U rcw 'U4+' to a'u) U) w100U0O0r i-i 14 Lam or - 4 b P.4J (0
4-' W W..it it f-r4 r.~ Q -4 Y9rI*4
fll- ,r0 fO **.4*JXX * * q** QW 0* 0 Wto00 442 = ~ 4 W 4J 4 U)U () UJ U)04J 04J400 1 P W* --1--4 U-4 8 8.4a r WHC0.0 0 0 ( ( ) in U w(4 ) ( a( 4 ua aa0 0*"If.- q: r 1 Vt Vf 0( o1 . 4R -QUV4-
r" (~
0 .
*r 4J** * *
I- * ** * *
Ad + *
*J =+ * *
*r a -..- mm* a a:3 ~ ~~ ~ ~ *N V*V -1 _ 4 jU L )q "V
m fa x~r co - %4 10 *.W-b 4
W4' U I-_% l * * NN.I J S * *+o-4~
W * -l %-( - = XX4 . + tAd*rlr * N q *114 N.4 0 *d WA"_ 4) CA* i x * 11it1
o~v ~ - W J) W *J 0 i .- 4 1 t-
o r. f* .4. - O *'Dq *.a*
-1 a* ifI 4m(o N fJ .1*aa *-a5(
--- * -"r tU 1t4)'V VV 1 )d 1 0 0) X. ) N)C b
4'~ ~ 0 0404...4) W r 90 4j+) Q I- -r
N dH X W4 M- +;4 W 4' J*r JI aM4 JJ4 J4 44J+4- M4 4-+ JI 4 'r dm 3
14=aW U W 1 IV.4 CCJ CC4
W 4-)( V(0M0( 0 ( ( 0 ( at
'4 &* 8- as S as*Sa * a80 Xx 0A
COC OC CD Ceoo)a o o q~N -'ul %Q,.D V43 -VL0)O a- NM--t
*~~ ~~~~ 10 -4- . (
4JI
4I
* * * *4-)
* * * * *0*~~ 14 *** * * *
*~~ U-4** -
* *w .4 0
* * * 0* '4-4
**0 uI V 0
*- *Z (+ N"
41 .1 P 0
go CD *~y %- r4 r.4J * -i IV 0to 4-) '- Ui 1w' 9 m
+) 0r Ur -Ir 0
:2 44 4j J ) r1 0 wo 1 4-oa, tun .- 0 +.- swoo 'N1 M H -P
CL* 10 -U - 4J to * **- (1, 0T 0I
U) q +j . *rl1-M j +-.dl %- d
+4r- -W " 0 w4 4-10) () 0 WOW w (U H :L*W4 PA +. O.0p*, *4~ t4J U+, 11 w 00 CL V 4-4
M. ft 4 I (a ofr= - It ( UJw Q "M5 O5-4 '-MMu+' Ow 00'-!- 11M - H 11 cro UuI. 11 , b."18 0(-
CL" d(I - M V 4J a' * (II UU u 14 g:4- +j Iu-8 8-4J 0 .4.'4M 11)WO * (* W ~IS tO(0 fl* 1\,. x -i(0, m (-i (4 ft y~ 44 4w.vs* * w 0 (D of cu 0 (t8. 4 *) 4-
*4-.C +3J P.5.4 0-- + a)H - . U)-)* 0~* .(P.q* U) N i 9 +'.* -r14 M *-U *.59 mw w w * M .* 4- %1I* ~* N54 500jI~*I WJ
* +) 4j+'5 W .' WU W-H 4.* W 545J * 'D * V 4- 4J5 *111*M M* (CU N0 (0 S4 WI WI W k 14 0 WI (0( * f* CC~UJ (0 CCM * CN'.. &.h ) *~ Q O(
8 N~--41i54 ""UII N-r 4-4 I;a * M* 8 +J11 4r* 4 -1* 4
4J5
Ad6
% V A
* W P *,* r
*do Ad* 0H .
63 %* ) 3 - A t~ r -V J4
0 M 1'l W ()it11 V- $4 wa w w w 3 ; '
* VI .0 (P 0 u : i v 0
O- 1-4 __X IL Wv
If k4.4 AC n- % 1 )
__ J4 ZT* ZTiW *0.9. 4- dW id %. % , *ar
0 0- rl 4 0 - :1+ rVA'
__ 0 0 (0 w- is4U -- 0ew j
W ) C 1 ) I W 11 H $to a' 11 0 W 0
4 *' 3 A 4. 01-i * .0a * -A x1W.44-'3 xx00 0 -W. 4J 01 4J J J9 4J0 41 4
P44, wJ xj M.4 X M X u- W- " ~ 4w XWW .W ** 0 *00 V *VQ W4VWUW *4
0i V*4 0 1-i4 1 .4 Q-* $3* U 4 H 1i-'4+-
1.4" cj CJW 00 ac U via 0c)
1.4 ~ U II k 112.
w in * II '- i---bosom"o
* * '*a *C* * *1 *
* .. s * 0F **1 2 * * 0 *
*0 09 4- * 11 P 4* (I C: * U-14*
*) I~ *ww a do * * !KK 4J 0 IN to l* * * en
* Otr * () x (e+ if a * =-V 5.c IV 5I *- VI 0 MH 4q (n * 4J)
*. a AC 4 4 -A *0i 4 - T- 0 &J* It
54i *r SD o' tc0 r 0 -4 V' W
U) (V 0r- aa) n) Q, ') (aifO c - LL~ to
U 1 W .' Mm A*~M 6:.' 10 u v v44* Rmu- al I-i r4 mi * ''0 I "cj 2~ If+ Pu .4 5*
0 1 w4. -1 -. W .. ma 3 4-4 - 0*e- toI *l Ad 3w .+Vr 0 *01 4J OVr 0 M 0$4 0 040 0 oo. -%Ua 5ir~.d W-0 1. (0 40o tox v C1110C Sr 4.. 01 *.I Stud a 1 W) W1SI *OW. LA SI
31 Ad 04 (1 %% aO 9-H22 0 '. 0 = I01-ommes 04 wIsa w (il (U "I4 t * -r4 .-W g 0 * 4U0 4) wj-. a.1
- 0 -P 4 .. q. 5j -i 3 Wtr . 0 * 9= 2 I r...J4 O'
61- '0 14 0 1 Q - gr * a 4* r.... I...0 a4- 000) 0 01o 3 w w %,41w 0 .u-f* 0 WQ 0 W '002~j 1E0"U
AC-4 * w w 2 3 m 1 0 1 4J W W r- o* W %- V-4 2I M w U*1 21 ON 0 4)' W. 64. LW012- I c~I 4'W W4.tS:1 $3 '0 If Wm '0-r a0 m-0 *.-0~ = -0w OuW 0 '0 vi
0 W 0 4N u- 11-W 0 PJ* 2* 4.1.4.)W 4WJJ S M
&' 0 S 1 0~t II w I.Ad in v 44 0 0mW K Uj. * 04)4 V
+1 -)4 Wa - X4' If 1i0'; U U Q U5 QNE2 CC*2J -4*a1 0k 0 - * U#-V n 3 1 U 0w- WW.M W 11SH" * 4JCj
4J-1 0* * * *0..r4J 0%' 4* t a 11 - OS)AU Oi *P- *1.~- *4J 4.JC U L *9r llgg +4 ..... )0A U' 4J * U)
II lI J - ( *-l * l '01 Wr. 9 * II-a-ft4 all +' -f- -4 W
= * (0 * o WE (40 *. U 111'97 -~ Q. *W4*0'f'*I *1IfW4 1 * W~ 0 10 *) ('0bP :211 C4-1 -el
'-WN "'0C4* tn.-i*4 * 0 US-r- ff P -.4* '0 04'+'*d 11 0 0 V .d 10 tW- Itfa '01- &4 -Hb *u4 *'4 X 3 -- 42M22 * O10(0V0 0 00U.-O 41 -*V0 0 0 L0* 4.)* 4.1* +) Jw 0 m 0'E.54U Zoa U) a:* *M5-s4-'5WIC"4-'*3W pq*S M* H Of w )4* O.W= 0 0:10u -
Q .. U" " a S * Q W * ( W * v w U W * 4 ' w . S U U " 0 V 0 4 4 4 * 4 * 4 W * 4 4 4 .
0-t -- -rm0OOQ00~00~3~04.0 OL0OJOJ0CJ NCJ0 .) 4
113
0)40
0 4--'I
Ad 4. co r.
ue IV r f)- U
0- 4'.di- rn U)I
-4 (N .~ 05 1100U r4 .n(
=I U1 wi UJ) A.
0 044-)+i rl co4-4 0 +J4 %
orq r_ P -N J%=r- 4 M-r4 Q(
('1so 14 %C 4J ('4 ._%.N * -4 0 QQ) W=W cz I
V o: ),= 0 'S--P z. % *Q.0 Ln IV4ro 0~ 00 ) r4-J 40 U-1 --- " r i 40,- me' =4 5. ('4
It ~ . U it a ofm M0tpNr#-o ft fW( rff0W U 0~f4.D +'sWwV~er- ". ,U .44-'O +2+*I At I('4g~ cb 0 wU' W.~ ~ ( rouI 1 11 .4- 11(
M 11 A(d4M4 .(4 - -4J V t-iC' 114 +) 11:C(IUV0441 i'0 0 *48 J9J. Tvh )4 1 XwM r- r44-- r-4-1U4Z 4u-W+ 4 ~l 4J 47 4-0 ' 4 It It W ~ . 4j l ir" (
04. W M- 014 044444- *44 44 0 W4444%64 00 114 *U4*4 44
NN'1 000 0 r.N4r4 NNNCN (N4J0 NN CNN %C4JY4 NN 0 (U(%J NN NC..0 0)SC
U~'Q 0.~ r A 0A O . ~ ~ ) ~ (~4J 114,**
id 40 (0o
Cb 0: CI4.- 0
.0 44
tu .)N H w wCZ in ::t:wza= -
0mn- cn ..... +J 4J
IQ LL w r
.*0 0 30 id trnt-7 0L -X ~4j %4.r-) tpoo r" : r 1:z 00) T r4 .93= 0000,J- w) 4) d
H1 In44r fu' 4)u u "Poqn ;q
F-4%. .-40 .. U)U 0jDad I+ dGj- .CIlG UJA 0. *r4.lI*.+. cgo w w 4 * 04N
0 -f4 * 44-l E 00 U) (V (UCCMKiCI-~ol2 t9K V*N sz aS-% to ,- 11U lb 4J4J4 .Dr- %V4)r4C
(9- * oC ' w n 4p 0 &;r".- rvIww(oL
r" J*4 V- a(1 0 1 44 Q) 1.QO ! 04100% *.+.-.(J * v =.
W W" W4-) W r. to w .0.-ri to *',.-*e' as .'+.' ua p p~ u uj'- 6 =~ cm
40u0). 00 0 0 0 NEV7* UO 11 -4 ro 0c0U :o ul "Uj Q . In V- VA 0 -4' o 0 11 of it 11 to if 11 so .- Vr X w .~,
Lad$ w.. 6H~k-WOC- IMWWW % c-i % 1 U) x P 0 0+4 # -
A* X* K.* 4. -4 '..DA* 01174 to 11 -+3j++4J qoo 4 j+jj V t"- 0 r- -rV~tJQ 0IgY% o 1OVO1N W(OCJ P- 9 ~ ~ I r. =n OC4%r - l 1 1110aM -44.f11 r
* * *0 "4 oi00.-4 (I'4000W O03030 *r-4: "40M c ,u oe'F- -4 -4(l- MM ,4-4 MM~ 000(000 4Krp,-m0 '-0
;OQW 40L +)' Do U)M04.' (a P, M- )4-1 U) ' 4.'4.4 U M0 0 W 044+4 0)44 V4444r-I o- U44-444 M 0444444 0 WV 0-l Q14.J444 U"I b...4 tr-I E-i"' V'I E-'f OD"r4 MHr4.'4.dr4"r EM-r4-Hr4.144 vu oirU r a,1s.10~
-M ocoooI
.a ocoooo
N116
LIST OF REFERENCES
1. Sorenson, H.W., "Kalman Filtering Techniques", Advancesin Control Systems, v. 3, pp. 219-292, Academic Press,1966.
2. Kirk, D.E., EE4413 Class Notes "Optimal Estimation: AnIntroduction to the Theory and Applications", (unpublished),Naval Postgraduate School, 1975.
3. Jazwinski, A.H., Stochastic Processes and FilteringTheory, Academic Press, 1970.
4. Cramer, H., Mathematical Methods of Statistics, PrincetonUniversity Press, 1961.
5. Mitschang, G.W., An Application of Nonlinear FilteringTheory to Passive Target Location and Tracking, PhDDissertation, Naval Postgraduate School, 1974.
6. Heffes, H., "The Effects of Erroneous Models on theKalman Filter Response", IEEE Trans. Automatic Control,v. 11, pp. 541-543, 1966.
7. Nishimura, T., "Error Bounds of Continuous Kalman Filtersand the Application of Orbit Determination Problems",IEEE Trans. Automatic Control, v. 12, pp. 268-275, 1967.
11
117
S
BIBLIOGRAHY
Anderson, B.D., and Moore, J.B., Optimal Filtering,Prentice Hall, 1979.
Meditch, J.S., Stochastic Optimal Linear Estimation andControl, McGraw-Hill, 1969.
Sage, A.P., and Melsa, J.L., Estimation Theory withApplications to Communications and Control, McGraw-Hill,1971.
Ladas, J.E., Interactive Monte Carlo Simulation Programfor Evaluation of Estimators and Predictors for FireControl Systems, Master's Thesis, Naval Postgraduate
Schol,19 81.
Ilal
~I~l9~I
118
0
0
S
0
INITIAL DISTRIBUTION LIST
No. Copies
1. Defense Technical Information Center 2Cameron StationAlexandria, Virginia 22314
2. Library, Code 0142 2Naval Postgraduate SchoolMonterey, California 93943
3. Chairman, Code 62 1Department of Electrical and ComputerEngineeringNaval Postgraduate SchoolMonterey, California 93943
4. Professor A. Gerba, Jr., Code 62Gz 1Department of Electrical and ComputerEngineeringNaval Postgraduate SchoolMonterey, California 93943
5. Professor H. Titus, Code 62Ts 1Department of Electrical and ComputerEngineeringNaval Postgraduate SchoolMonterey, California 93943
6. Commanding Officer 1Naval Underwater Weapons EngineeringStationKeyport, Washington 98345
7. Commander Joseph Jaros, USN 2Command and Control Engineering CenterCode G520Reston, Virginia 22090
112
p
I-T
4-#
4 W4
!R f