I I66 IEEE TRANSACTlONS ON ACOUSTICS. SPEECH. AND SIGNAL PROCESSING, VOL. 31. NO X.

AUGUST 1989

Robust Adaptive Kalman Filtering with Unknown

Inputs

Abstract-The standard optimum Kalman filter demands complete known statistics for modeling error compensation, and
knowledge of the system parameters, the input forcing functions, and systems for which we choose to use a filter of lower di-
the noise statistics. Several adaptive methods have already been de-
vised to obtain the unknown information using the measurements and mension than that of the actual system state.
the filter residuals. Methods which a r e designed to simultaneously es- In general, a linear system can be modeled by
timate the input forcing functions and the noise statistics have subop-
timal performance and a r e potentially unstable. This problem is com- Xk+l = 4kXk + BkUk + w k . (1)
mon in target tracking where the acceleration inputs and the noise Here, the subscripts index the time sample: x k is an ( n X
statistics are unknown.
In this paper, we propose a method to adapt the Kalman filter to
1) system state vector, & is an ( n x n ) transition matrix,
the changes in the input forcing functions and the noise statistics. The uk is an ( I X 1) vector of the input forcing function, Bk
resulting procedure is stable in the sense that the duration of diver- is an ( n x I ) matrix, and wk is an ( n x 1 ) vector of a
gences caused by external disturbances a r e finite and short and, also, zero mean white sequence. The discrete vector measure-
the procedure is robust with respect to impulsive noise (outliers). The ment z k is described by
input forcing functions are estimated via a running window curve-fit-
ting algorithm, which concurrently provides estimates of the measure- Zk = HkXk + vk, (2)
ment noise covariance matrix and the time instant of any significant
change in the input forcing functions. In addition, an independent where z k is an (rn x 1 ) measurement vector, Hk is an (rn
technique for estimating the process noise covariance matrix is sug- X n ) matrix, and v k is an (rn x 1 ) measurement error
gested which establishes a negative feedback in the overall adaptive vector which is assumed to be a zero mean white noise
Kalman filter. This procedure is based on the residual characteristics
of the standard optimum Kalman filter and a stochastic approximation
sequence uncorrelated with the wk sequence. The covari-
method. ance matrices for the w k and v k vectors are given by
The performance of the proposed method is demonstrated by sim-
ulations and compared to the conventional sequential adaptive Kalman
Qk; i =k
E{WkW;} = (3)
filter algorithm. 0 i f k
Rk; i =k
I. INTRODUCTION AND BACKGROUND E{VkVr} = (4)
0 i + k
T HE optimum Kalman filter is a recursive algorithm
which is developed to solve the state estimation prob-
lems of a known system based on certain assumptions
E { wkvT} = o for all k and i (5)

about the system's mathematical model. These assump-

-
where E { } is the statistical expectation, superscript T
denotes transpose, Qk is the process noise covariance ma-
tions include complete information about the input forc-
trix, and Rk is the measurement noise covariance matrix.
ing functions and the noise statistics. In practice, this in-
The optimum Kalman filter estimator and error covari-
formation is usually not totally known, and assumptions
ance update equations are given by [I],
about the parameters of the noise statistics and the math-
ematical model of the system are necessary to allow adap- i k = 4k-lfk-l Bk-lUk-1
tive estimation of the unknown information. This problem
is commonly observed in any control or tracking system
which involves state estimation in an environment where
Pk = ( 1 - & H ~ ) ( ~ ~ - I P , - I ~+~QTk --i )I? (7)
noise characteristics and/or deterministic input forcing
functions are unknown. Such models can be used to study where K k , the optimum Kalman gain at tk, is given by
systems subject to component failures, systems involving T T
small nonlinearities, systems with process noise with un- K k = (4k-iPk--id'-i + Qk-,)H~[Hk(4kk-iPk-i4k-l
Manuscript received October 5 , 1987; revised November 2 8 , 1988. -k Qk-1)Hl &]-I. (8)
A . Moghaddamjoo is with the Electrical Engineering and Computer Sci- In the above equations, ikis an estimate of the system
ence Department, University of Wisconsin-Milwaukee, P.O. Box 784,
Milwaukee, W1 5 3 2 0 1 . state vector x k , and P k is the covariance matrix corre-
R. L. Kirlin is with the Electrical Engineering Department, University sponding to the state estimation error defined by
of Victoria, Victoria, B.C., Canada V 8 W 2Y2.
IEEE Log Number 892873 I . Pk = E { (Xk - i k ) ( x k - ik)T]. (9)

0096-35 18/89/0800-1 166$0 1 .OO 0 1989 IEEE

MOGHADDAMJOO AND KIRLIN: ROBUST ADAPTIVE KALMAN FILTERING I I67

This optimum recursive filter starts with the a priori in- proaches use the augmentation method of the state vector
formation about the initial state vector, fo,and its corre- to simultaneously estimate the input forcing functions and
sponding error covariance, P o . The algorithm is based on the state vector. These procedures are in general based on
the following assumptions: the assumed knowledge of the noise statistics.
1) the mathematical model of the system is truly rep- Myers and Tapley [ l 11 have proposed a sequential
resented by (1) and (2); adaptive procedure which attempts to concurrently esti-
2) the system parameters, &, B k , and H k , are perfectly mate the unknown input forcing functions and noise sta-
known; and tistics. There are, however, several problems associated
3) the input forcing functions, uk, and the noise statis- with this method. The major one is that the estimates of
tics are fully available. the unknown parameters are related directly to the esti-
In practical situations, the above-mentioned assump- mates of the state vector which itself is again a function
tions are not totally valid and the standard optimum Kal- of the unknown parameters. Existence of this closed loop
man filter cannot be applied directly. This problem has in the algorithm is a source of instability which negates
been studied since the development of the Kalman filter its applicability to situations involving long time runs.
and, as a result, several different adaptive procedures have There are also several modified versions of this algorithm
been devised. Here we intend only to address the prob- [12], [13] which are robust to outliers and consider only
lems associated with the uncertainties in the input forcing pulse-like input forcing functions. Although these robust
functions and the noise statistics. procedures are insensitive to the measurement error out-
Various adaptive filters have already been proposed for liers, and tan be run for longer periods of time, they still
the estimation of the noise covariance matrices [2]-[7]. have the same source of instability.
Some of these methods are briefly reviewed by Mehra [3] Here, we propose a running window (limited memory)
who has divided the adaptive filters into four categories: curve-fitting algorithm which is robust to the measure-
Bayesian, maximum likelihood (ML), correlation, and ment error outliers and provides optimum (in the weighted
covariance matching. Generally, Bayesian and ML meth- least-square error sense) estimates of the input forcing
ods demand more involved computations and are based functions and the measurement noise covariances (under
on the assumption that the noise statistics are time in- the assumption that the mathematical model of the system
variant. In correlation methods [2], [4], [6], the basic idea is correct). In this work, not only are the admissible input
is to autocorrelate a known linear processing of the sys- forcing functions broadened to include any of determin-
tem output and derive a set of equations relating the sys- istic form, but also estimates of the measurement noise
tem parameters to the observed autocorrelation functions. statistics and the input forcing functions are calculated in-
The resultant equations are then solved simultaneously for dependently. The significance of this approach is that es-
the unknown parameters. timates of the unknown parameters are not linked to the
Sinha and Tom [6] have proposed a procedure in which Kalman filter state estimates. In addition, by applying a
the initial value of the Kalman gain is calculated using the stochastic approximation method, the state noise covari-
algorithm proposed in [4]. Then, based on a stochastic ance, Q, is changed in such a way that the statistics of the
approximation method, as new measurements become filter’s residuals approach that of the optimum Kalman fil-
available, any change in the Kalman gain is adaptively ter. This method originates a stabilizing negative feed-
estimated. In this procedure, estimation of the noise co- back in the statistics of the residual sequence of the filter.
variance, which is only required for the gain calculation, In the next sections, first, the conventional sequential
is eliminated and changes in the Kalman gain are directly adaptive Kalman filter [ l 11 is reviewed, and then our ro-
evaluated. Although this modified algorithm is computa- bust adaptive procedure is formulated. Finally, perfor-
tionally more attractive and brings further improvement mance of the proposed method is demonstrated by simu-
in the overall results, like other methods in this category, lation of a simple one-dimensional tracking problem.
it is only applicable to the stationary time-invariant sys-
tems. Initialization of this algorithm, and in general any 11. CONVENTIONAL ADAPTIVE KALMAN FILTER
method in the correlation category, requires suboptimal In this algorithm, sequential estimators are derived for
filtering for a long period of time, which may not be per- suboptimal adaptive estimation of the unknown noise. sta-
missible in some applications. tistics, deterministic input, and the system states. First-
Covariance matching techniques are based on some and second-order moments of the noise processes are es-
schemes which make the filter residuals consistent with timated based on the estimates of the state and the obser-
their theoretical covariances [7]. These methods are very vation noise produced by the Kalman filter algorithm. A
restrictive and their convergence is questionable. limited memory algorithm is developed for adaptive cor-
Adaptations to unknown forcing functions have been rection of the a priori statistics, which are intended to
studied only for the target tracking problems [8]-[ 101. In compensate for time-varying model errors [ 1 11.
most of these methods, the deterministic input term in (l), In this procedure, N, most recent samples of the state
Bkuk,is dropped and the state noise covariance, Q , is in- error sequence
creased to allow the filter to follow any unpredictable
change in the input forcing functions. Some other ap-
I168 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, A N D SIGNAL PROCESSING, VOL. 37. NO. 8. AUGUST 1989

are used to estimate an unknown constant forcing func- suboptimum state estimates. It has been shown [13] that
tion. This is there is a positive feedback in this sequential algorithm,
which is mainly due to the estimates of Rk and Q k . In
Bk-1ak-l = -
1 Ns
N, J = I
c
fk-J. (11) addition, Since a combination of BkUk and rk can only be
observed through zk, it is impossible to simultaneously
This bias is then included in the state estimator estimate these unknown parameters as indicated. This
problem has been shown and discussed in [ 131.
i!,= 4 k - l i k - l + Bk-lak-1 - &[zk - H k ( 6 k - l i k - l The robust version of this adaptive filter is proposed in
[13] where r k , the measurement bias, is assumed to be
f Bk-lGk-1) - ‘ k ] . (12)
zero. In this method, the input forcing function is as-
Similarly, in (12), is an adaptive measurement bias es- sumed to be a step-like function, and robustness is with
timator based on the N, most recent errors in the predicted respect to the noise outliers. The fact that this method is
measurements insensitive to the state and measurement noise outliers re-
sults in less variation in the estimates of Rk and Q k , due
1 Nz to an abrupt change in the unknown input forcing func-
‘k =
N, J = I [zk-j+1 -
- Hk-j+I(6k-jik-j + Bk-~ak-~)]. tions. In addition, estimates of the time instants of the
abrupt changes in the input provide a tool (refiltering from
(13) the time of change to the present time) to adjust the pre-
Myers and Tapley [ 111 show that for linear Hk and opti- vious estimates of the unknown parameters. This algo-
mum state estimates, the bias measures (11) and (13) are rithm can be run for a longer period of time without sig-
unbiased. Groutage [12] also shows that for linear Hk and nificant filter degradation. However, since the skeleton of
the optimum state estimates, the covariance estimator this robust algorithm is derived from the proposed method
by Myers and Tapley [ 1I], there is a potential of insta-
1 Ns T bility in the procedure.
Qk ---
N, - 1 J = I [(fL-J+l - Bka;)(fL-J+I - Bka;) ] To counter the problems inherent in the methods pro-
posed in [ I l l and [13], we propose an adaptive Kalman
. N,
filter with the following properties: a) it is robust with
(14) respect to impulsive noise, b) it provides optimum esti-
mates of the unknown parameters, and c) it is stable, in
the sense that its divergence from the filter’s nominal path,
due to any external disturbances (caused by unknown sud-
den changes in the input forcing functions and/or noise
statistics) is a time decreasing function. Differences be-
tween this method and the method proposed in [13] are
that this method is not based on the Myers and Tapley
[ 113 algorithm and it can handle any general form input
is unbiased. Similarly, the measurement noise covariance forcing functions. Also, estimates of the unknown param-
estimate eters in this method are independent from the Kalman state
estimates. In addition, the proposed stochastic approxi-
mation for Qk inserts a negative feedback in the estimation
process of this parameter which stabilizes the filter under
any unknown external disturbances (i.e., sudden changes
N,
-~ -1 T in the input).
Hk-j+ l(6k-jpk-j@k-j + Qk-j)
Nz
111. ROBUSTADAPTIVE
KALMAN
FILTER

T 1
Hk-j+l 9
A. Estimation of the Unknown Input Forcing Functions
The instability problems of the conventional adaptive
procedure, as discussed earlier, are mainly due to the de-
pendency of the adaptation routine on the byproducts of
the Kalman filter. It is therefore desirable to search for an
is shown to be unbiased. independent adaptive procedure to estimate the unknown
There are some problems associated with this algo- parameters. In this procedure, the unknown input forcing
rithm. All these estimators are good as long as state esti- functions are estimated by using only the N most recent
mates are optimum. During an abrupt change in the input measurements. In consequence, it is necessary to discover
forcing function, however, state estimates will degrade a direct relationship between the system measurements
rapidly from optimum values, causing rapid increases in and input forcing functions. The general deterministic in-
the estimates of Rk and Q k . This in turn will result in more put-output relationship is derived as follows.
MOGHADDAMJOO AND KIRLIN ROBUST A D A P T I V E K A L M A N FILTERING I169

The system equations in noise-free (denoted by super- sion of uk, 1 be less than or equal to the dimension of zk,
script *) situations are m.
The problem now is reduced to that of fitting an appro-
xf+l = + ~ +f B u ~ (19) priate deterministic function of time to each component
of measurements in order to obtain estimates of error-free
zf = H x t , (20)
measurements, z:'s. Given the complete knowledge of the
where it is assumed that the system is time invariant. Note system characteristics, it is always feasible to find these
that in the case of time-variant systems, the derivation is deterministic functions. Let us assume that the following
similar but much more complicated. The z-transforms of measurement model is appropriate for the system under
(19) and (20) are study:
X*(z) = (zl - +)-'BU(z) (21) zr = Po + Plfl(tJ +PZ"f2(tJ + - * + Ppfp(t1) + €1,

Z*(z) = HX*(z). (22) (30)

Substitution of X * ( z ) from (21) into (22) yields the final wheref,(t);j = 1, 2, . , p are determined based on
input-output relation in the z-domain the knowledge of the system and 0,;j = 0, 1, 2, * 7P
Z * ( z ) = H(zZ - +)-'BU(z). are m-dimensional vectors of coefficients which are as-
(23)
sumed to be constant. E , is the measurement error at time
In general, for any linear system, we can assume t , , which is assumed to be white Gaussian noise with po-
tential outliers.
To obtain estimates of the error-free measurement, z:,
it is necessary to find the best fit (in the weighted least-
wheref(z) is an nth-order polynomial in z , and A ( z ) is square error sense) to the N most recent measurements
an (n x n ) matrix whose elements are polynomials in z . using the measurement model in (30). This curve-fitting
Substitution of (24) into (23) yields procedure should be conducted for each individual com-
ponent of z k regardless of their interdependency. There-
f ( z ) Z * ( z )= H A ( z ) B U ( z ) (25) fore, for simplicity and without the loss of generality, we
which in the discrete time domain is written as can assume scalar measurements in the following curve
I1 -1 n- I fitting formulation.
To simplify the notation, the following variables are
introduced:
Here we have assumed XI1 = fl G I )

n-l
x12 = f 2 ( t r ) ; i = 1, 2, . - ,N

and
n- I

HA(z)B = c DJ',
r=o
(28) where t N corresponds to the present time (the time of the
most recent measurement). Assuming scalar measure-
where D,'sare (m X 1 ) constant matrices. Note that the ments, (30) can be written as
dimensions of the unknown vector uk and the measure-
ment vector zk play an important role in the estimation z=XP+E (32)
process of the unknown input forcing functions. Assum- where
ing that z f ' s are known and the estimation process is at
step q , from (26) we have
n-I n-2

~ n - 1 ~= 4
r=o
arz;-n+r+l - C DIUq-n+l+l (29)
r=O E = tz, e ., (35)
where u ~ - , , + ~i +=~ 0,
, 1, * , n - 2, are substituted

by the estimates of the input forcing functions in the pre-

vious steps. If m = 1, we would have a unique solution
for uq. But if m is less than I, the number of equations is
then less than the number of unknowns and, in general,
there is no unique solution for uq. In the case of m > I,
there is a typical minimum MSE solution to (29). To ob-
tain a unique estimate of the unknown input forcing func- Therefore, z is an ( N x 1 ) vector of the N most recent
tion at step q, it is therefore essential to have the dimen- measurements, X is an [ N X ( p + 1 ) ] matrix of the lev-
~

I170 IEEE TRANSACTIONS ON ACOUSTICS. SPEECH. A N D SIGNAL PROCESSING. VOL. 37. NO. 8. AUGUST 1989

els of the regression variables, 0 is a [ ( p 1) X 1 1 + regression coefficient vector before a change in the input
vector of the regression coefficients, and E is an ( N x 1 ) is 0,and after the change is 6 = 0 +
A@, where A@ is
vector of white measurement errors with zero mean and the change in the coefficients due to the change in the
variance 13. However, it is assumed that the measurement input. This problem has been studied for different cases
errors include some wild points (outliers). by Hudson [19]. In this work, only the case with two-
The objective is to estimate the vector of the regression segment curves is considered. For N data points in the
coefficients, 0,while reducing the effects o,f the outliers. window, it is important to determine whether or not a two-
We use the weighted least-square estimate, b, which min- segment curve is appropriate and has the weighted least-
imizes square error. If so, it is necessary to estimate the point j
N at the intersection of these two curves. Also, it is essential
c
S( 0) = I = 1 W,E? = ETW€ = ( z - XP)'W(z - xp) to test the hypothesis for the significance of changes in
the coefficients, Ab. If none of the components of Ab is
( 3 7 ) statistically significant, then the estimated j should not be
considered as an estimated point of change in the input.
where the matrix W is diagonal with elements w l ,w 2 ,
. . . , w N ,which are the corresponding weights of the The weighted least-square solution sought is now briefly
stated for the case of two submodels, f l (x; 0)andf2(x;
errors. The weighted least-square estimator of 0 [14] is
6 ) joined together at x = a.It is necessary ty find vectors
8 = (x~u.x)-lxTwz. (38) p = 8, 6 = 8, and real values a = &, j = j , which min-
imize the N weighted squares of errors in the window.
Measurement errors with small w, have larger variances
That is, minimize
than errors with large w,.
Application of the weighted least-square method re-
quires the knowledge of the weights, w,.Sometimes, prior
knowledge, experience, or information from a theoretical N
model can be used to determine the weights. In some
cases, it is necessary to guess the weights, perform the + i=j+l
X
wi[zi- f 2 ( x i ;S)]' (41)
analysis, and then reestimate the weights based on the re-
sults. Several iterations may be necessary. subject to the following relationships among the parame-
In general, a class of robust estimators may be defined ters:
that minimizes a function p of the residuals, for example,
f l (&; 8) = f 2 ( % 8) (42)
N N

min C p(e,/s) = min C p[(z, - xT~)/s] (39) and

p r=l p ,=I
X'J <
- & Iy + 1 . (43)
where xy denotes the i th row of X. The function p is re-
lated to the likelihood function for an appropriate choice The parameters @, 6, a,a n d j are clearly not independent.
of the error distribution. In (39), s is a robust estimate of The estimatedj is accepted as a point of change in the
scale. A popular choice f o r s is absolute median deviation input when A 0 , which is related to the change of the in-
given by put, is statistically significant. Based on Gaussian noise
I 1
s = median e, - median ( e , ) /0.6745 (40)
assumption (aside from the additive impulsive noise which
is rejected in the estimation process of @ b y using the
which is insensitive to outliers, and its statistical behavior weighted least-square method), the random variable 0 is
is well studied [ 141-[ 181. A number of popular robust cri- approximately normal [ 141. Therefore, test of the hypoth-
terion functions are discussed in [ 141. Based on an exten- esis (significance of A b ) can be conducted using Gauss-
sive Monte Carlo simulation study in [22], we have found ian distribution. For example, for 95 percent confidence,
that Tukey's biweight method and the square of the ab- AB is statistically significant if it is greater than 1.96 times
solute median deviation given in [16] are the most ac- its corresponding standard deviation. If A p is statistically
ceptable methods for providing a robust estimation of insignificant, the process then proceeds without consid-
variance. In [22] it is shown that these methods have the ering ;as the point of change in the inputs.
least bias and fewest variations. On the other hand, it is
also shown that Spearman's rank-correlation [ 171-[21] is C. Estimation of the Measurement Noise Covariance
the most appropriate method (again due to its least bias Matrix
and variation) for the robust estimation of correlation coef- An important byproduct of this curve-fitting procedure
ficient. is that it is possible to simultaneously estimate the mea-
surement noise covariance matrix, R , independent of the
B. Estimation of the Point of Change in the Input Kalman state estimates. We estimate R by calculating the
Forcing Functions covariance matrix of the curve fitting residuals.
For the purpose of refiltering, estimation of the points Let v k represent the vector of residual at time tk. The
of change in the inputs is required. It is assumed that the conventional estimate of R (if vk is white Gaussian), using
MOGHADDAMJOO A N D KIRLIN: ROBUST ADAPTIVE KALMAN FILTERING I171

r
the N most recent residuals, is given by It can be shown that [2]

1
N HP-HT + R, k =0
N - 1 c (Vi
R = - v)
i=l
- 5)(Vi - (44)
k =
c Hd[I - KHIk-’$[P-HT - K(HP-HT + R)],
where k > O (47)
where
v = -
-
l
Ni=1
cN vi. (45) P- = q5PdT + Q. (48)
Note that the optimum value of K , i.e., K =
Robust versions (with respect to outliers and the as- +
P - H T ( H P - H T R ) - ’ , makes C, vanish for all k > 0
sumption that vk is white Gaussian noise with potential (the optimum filter property).
outliers) of (44) and (45) can be found in [14]-[18] and In this paper, we propose that instead of trying to esti-
[20] and [21]. As mentioned earlier, Tukey’s biweight mate Q completely, Q can be used as a negative feedback
method and the square of the absolute median deviation to control and adjust the statistics of the filter’s residual
are the most acceptable methods for the robust estimation sequence. In other words, Q is modified and changed in
of variance, and Spearman’s rank-correlation is the most such a way that causes ck’s to approach their optimum
appropriate method for robust estimation of correlation values. It is therefore necessary to recognize the direction
coefficient [22]. of changes in c k due to variations in Q. Based on (47), Q
must be adaptively changed so that c k vanishes for k >
D. Estimation of the State Noise Covariance Matrix 0.
The optimal estimation of Q is a very difficult task. By careful consideration of (7), (8), (47), and (48), it
Mehra [2], [3] has shown that under special conditions, it is evident that if any component of c k is positive fork >
is possible to optimally estimate Q adaptively. His stated 0, then proper correction requires an increase in those
conditions are that components of K which correspond to the positive com-
1) the system should be controllable and observable; ponent of c k . This can be achieved by an increase in the
2) process noise should be stationary, and there should corresponding components of Q; similarly, negative com-
be no simultaneous input estimation; ponents in ck, k > 0,will be corrected by a decrease in
3) for large sample size covariance estimation, the Kal- those components. In other words, if Ck is positive for k
man filter with suboptimal gain should be run for a long > 0, Q should be increased and, as a result, C, will be
period of time to ensure steady-state suboptimal residual decreased (ck approaches zero). On the other hand, if Ck
sequences; and is negative for k > 0, Q should be decreased and, as a
4) in order to have a unique solution, the number of result, c k will be increased ( Ck approaches zero). Thus,
unknown components in Q must be less than ( n X r ) , the effects of c k on Q are direct and, for C, = [O], k >
where n is the dimension of the state vector and r is the 0, it is necessary to keep Q unchanged.
dimension of the measurement vector. The relationship between Q and C, cannot be derived
These conditions are generally not satisfied in practice. analytically, but several different empirical relationships
The role of the noise covariances, Q and R , in the Kal- may be proposed; for example,
man filter is to adjust the Kalman gain in such a way that
it controls the filter bandwidth as the state and the mea- Q , + l = Qiexp [ A ( c , + c2+ - + C,)C{’B~]
*

surement errors vary. Direct estimation of R , as men-

(49)
tioned before, is attainable via an adaptive curve fitting
procedure while, in general, direct estimation of Q is im- where matrix Q is ( n X n ) , and matrices Ci’s are ( Y X
possible. r ) , A and B are coefficient matrices which can be found
For indirect estimation of Q, we propose to use the re- experimentally, and their dimensions are ( n X r ) . The
sidual properties and adjust Q in such a way that statistics number of autocorrelations k that can be used depends on:
of the filter residual approach those of the optimum Kal- 1) the maximum number of data points that are necessary
man filter. In this regard, it is therefore necessary to de- to yield acceptable estimates of Ci’s,and 2) the maximum
rive a relationship between Q and the autocorrelation number of permissible data points in the moving window.
structure of the Kalman filter residual sequence y i . Q must In this procedure, due to our findings in [22], we suggest
then be modified in such a way that reduces the deviation that small sample size estimates of correlations (like rank-
of the autocorrelation structure of yk from its optimum correlations) be used.
values 121, 131. By assuming that the filter is in a steady- Although the stability formulation of this method, due
state condition with a steady-state Kalman gain K, the re- to its complexity, is not readily accessible, it can be easily
sidual sequence yi is a stationary Gaussian sequence [2]. understood by the following arguments. In this method,
The autocorrelation of yk is defined by Q is the only unknown parameter which controls the Kal-
man gain, because R and U are independently estimated
and they are not associated with any feedback loop in the
I172 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, A N D SIGNAL PROCESSING, VOL. 37. NO. 8. AUGUST 1989

filtering process. Let us assume that, due to some distur- The estimation of U is
bances (i.e., unknown sudden changes in R and/or U),
ci = 2 & . (54)
the Kalman gain becomes less than its optimum value.
The resultant residual sequence will then have positive (This is also valid, with some approximation, if U is a
autocorrelation, due to (47). Detection of the positive au- slow time varying function.)
tocorrelation demands an increase in Q , which will, in For an abrupt change in input, there would be a maxi-
turn, increase the Kalman gain ( K changes in the direction mum N-step delay in time until the estimation of U be-
which approaches its optimum value). On the other hand, comes correct. Therefore, it is better to select N as small
if the Kalman gain is larger than its optimum value, the as possible while still getting a reliable estimation.
residual sequence will develop a negative autocorrelation For the calculation of R , the weighted sum of squares
which requires a decrease in Q . Reduction of Q will then of fitting errors has been used.
decrease the Kalman gain ( K changes again in the direc- The following algorithm is used to estimate the point
tion which approaches its optimum value). This correc- of change in the input.
tion continues until K reaches its optimum value where it 1) Compare b2( k ) , the final estimate of the parabolic
oscillates. Therefore, deviation of K from its optimum coefficient at time t k , to b2( k - N +
1). If their difference
value, due to any disturbances, not only is controlled by is greater than 2 * Var ( & ( k ) ) , there is a significant
Q , but also will be reduced in time. This behavior is what change in b2;otherwise, there is no indication of change
we refer to as a negative feedback which has a stabilizing in the input.
role in the overall performance of the algorithm. 2) Calculate

IV. SIMULATION RESULTSOF THE ROBUSTKALMAN SS,*(k) = SS&) + SS,(k -N + 1) (55)

FILTERI N A TARGETTRACKING SYSTEM where SSE(k ) is the current estimate of the sum of squares
The one-dimensional target tracking system is used in of errors.
this section as an example to demonstrate potential im- 3) As the window moves, estimate the point which
provement of the proposed method over the conventional makes SS,* minimum. If k’ is the pqint corresponding to
adaptive Kalman filter proposed by Myers and Tapley the minimum SS;, then the estimate of the point of change
[ 1 13. It is assumed that measurement of position is di- in the input would be
rectly available.
System and measurement equations in this case are ;=k’-N+l. (56)
For estimation of Q , it is suggested that only p 1 =
CI/Co and p2 = C 2 / C , where Ci= E { yjy:-;} be used.
In this example, yj’s are scalar; thus, so are the Ci’s, and
r -
(49) simplifies to

where rl and r2 are Spearman’s rank-correlation estimates

In this equation, T i s the sampling interval, X l k is the po- of p 1 and p 2 , respectively,
sition of the target at time t k , and X 2 k is its corresponding k
velocity. The acceleration (maneuver) of the target at time
t k is described by the scalar U & . Measurement z k at time t k
6 c
i=k-N+l
(rank ( y i ) - rank ( y i - l ) ) L
r,(k) = 1 -
is also scalar, and w k is a vector white noise sequence, N ( N 2 - 1)
which is independent of the measurement white noise se-
quence z / k . Both of these sequences are considered zero
k
mean with potential outliers. 2
The noise-free equation which relates continuous mea- 6 (rank ( Y J - rank ( Y i - d )
i=k-N+l
surements to the continuous constant input U with initial r2(k) = 1-
N ( N 2 - 1)
position zo and initial velocity vo is
(59)
in which N = 15 has been used. Selection of the param-
eter (11has a direct effect on the speed of convergence and
A similar relation would exist for any polynomial-form the rate of fluctuation about the optimum value. The speed
input u ( t ) . must be traded against the fluctuations around the opti-
Assume that N measurements Z k - N + I , Z k - N + 2 , * * , mum value.
zk are available in the moving window of size N . A par- In order to study the convergence property of (57) for
abolic curve is then fitted to these data points: estimating Q, several simulations have been conducted.
In these, the Kalman filter routine uses true values for R
2; = Bo + bliT+ b 2 i 2 T 2 (53) and U. In the first simulation, the true value of Q is 101
 /
MOGHADDAMJOO A N D KIRLIN: ROBUST ADAPTIVE KALMAN FILTERING I173

;
100,

60

50
. ,
- A--TRUE VALUE
--8--RO0 ADPT
COMPARISON OF O(1.1) EST
I

Fig. 1. Estimation of Q ( 1, 1 ), position's process error variance, where

Q, = 1001and CY = 0 . 3 . A = true value, B = estimated value.
0
3
e+
2
E
11

10

7 :

6 -

5 -

4 :
3 -

2 .

1 -

-0

1
-
~

'
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

'
-A--TRUE
-B--RO0.
-C--CON

~ '
VALUE
ADPT.
AOPT

~ ' ~ '

Fig. 2 . Conventional adaptive and robust regression estimates of the input

sequence. A = true value, B = robust estimated value, C = conven-
"
fi* 1
A0

~ ' ~ ' ' " ' ~ " "

UYm""--.
" " " '

tional. Ensemble average over 100 independent trials.

and initial value Qo is chosen to be 1001. The result of 118

COMPARISON OF R ESTIMATION
. . . , . , . , . . . , . . . . . , . . . , . , . , . , .
this simulation is shown in Fig. 1 . The coefficient a in ----TRUE VALUE h
(57) has been set to 0.3. In a second trial, the initial value
of Q has been changed from 1001 to 1. The result of this 110

simulation is not shown, but convergence and fluctuation

about the true value are similar. Although convergence is
slow, the method does yield convergence, and more ex-
perience may improve the rate. These simulations show
that deviation of Q from its true value, in any direction 96 - Y

and with any magnitude, will be decreased in time. If any

unknown sudden changes occur in U and/or R , during the
adaptation lag period, the correspondence between Ck's
8.

82

80- '
~

-
0
8

20
'

40
' ' Ir':
60 80
, ,
100
, ,
120
, ,
140
,

TIME
,
160
,

180
, . ,
200
,
220
, , ,
240 280 280 300

and Q may change and, as a result, Q may deviate further

from its true value. But this deviation will start to be de-
creased and remain in its correct convergence trajectory
after the adaptation lag is over.
Measurement data points are simulated for the target 340
320
1, i
tracking system based on the following parameter values:
Q ( 1 , 1) = Q ( 2 , 2 ) = 0.1, Q ( 1 , 2 ) = Q ( 2 , 1 ) = 0.
For 5 percent contamination at scale 3, additive noise has
R = 100, 95 percent of the time, and R = 900, 5 percent
of the time. Window size used for curve fitting is N = 10
and for estimation of Q, N = 15.
Since we want to show the performance of the proposed -[noi ' '2 1 ' 8:' I ' n ' . 1
method in adapting to sudden changes of the input, while 0 20 40 60 BO 100 120 140
mic
160 180 ZOO 220 290 260 2Bo 300

R and Q are unknown, the simulated input forcing func- (b)

tion is chosen to be a pulse with 10 m/s2 height and 60T Fig. 3. (a) Robust regression estimate of R , measurement error variance.
s duration. Fig. 2 shows the true and average conven- A = true value, B = estimated value. Ensemble average over 100 in-
tional and robust input estimates (over 100 trials). The dependent trials. (b) Conventional adaptive estimate of R , measurement
error variance A = true value, B = estimated value. Ensemble average
robust estimate exhibits less delay than the conventional over 100 independent trials.
estimate by a factor of 2 / 5 . This factor remains almost
the same under any similar conditions, i.e., decreasing
the size of the moving window, N , for reducing the time that this estimator is considerably worse, with Q( 1 , 1 ),
delay. as shown in Fig. 4,ranging up to 450 by the 300T mark.
The robust regression estimate of R is more accurate With only casual study of these results, it becomes clear
than the other procedures tested. Fig. 3(a) shows the ro- that conventional estimates of Q have no relation to its
bust estimate of R. The conventional adaptive estimate true values. Existence of the positive feedback in this es-
shown in Fig. 3(b) ranged up to 400 at 130T and stabi- timator yields an increasing bias in the estimates of Q .
lized around 130 after 200T. It is obviously unrelated to This causes the filter bandwidth to become infinitely wide.
the true measurement errors. Overall performance of the proposed procedure can be
The conventional estimation of Q is calculated using measured by calculating the mean-squared error (MSE) of
the Kalman filter state estimates. Our simulations show the estimates of position and velocity. MSE of the esti-
 I174 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, A N D SIGNAL PROCESSING. VOL. 31. NO. 8, AUGUST 1989

comparison of the average behavior of these procedures.

In all cases, the results are averaged from over 100 in-
dependent trials.

V. CONCLUSION
In this paper, we devised an estimation algorithm for
I- determining the state estimates of a known system, when
there is less than full knowledge of the necessary a priori
information, i.e., noise statistics and the system’s forcing
functions. All estimators in this algorithm are robust to
potential outliers (impulsive noise). Adaptations with re-
spect to the input forcing functions and the measurement
Fig. 4. Conventional adaptive estimate of Q( 1, I ) , position’s process er-
ror variance. A = true value, B = estimated value. Ensemble average noise covariances are independently carried out via a
over 100 independent trials. moving window, curve-fitting algorithm, using the
weighted least-square method. Estimation of the state
noise covariance, Q,is indirectly directed toward its op-
COMPARISON OF ERROR IN X(1)
. , . . . . , . . timum value by demanding optimum Kalman filter resid-
n
3400 . . , . , . , . , . , . I . . . , .
Uylllx“-;i
3200
I”””
“1“”
- A--OPT VALUE
uals.
1 -0--ROB ADPT.
2800
2600 - - C--MEAS. ERROR
The overall algorithm, with regard to adaptations to the
-
=
2400
2200
.-
.
noise covariances and the input forcing functions, is ro-
z 2000 1 bust to the noise outliers and stable in the sense that de-
x ,800 -
z 1800 - viation of the Kalman gain from its optimum value, due
5 1400
to the lack of knowledge of the necessary a priori infor-
E 1200 1
1000
600
-
-
mation, will be decreased in time (except during transi-
600 - -B-
tion period, when a sudden change occurs in one of the
.oo -
unknown parameters). This algorithm is examined by
0 20 40 80 80 100 120 140 180 180 200 220 240 260 280 300
simulation of a one-dimensional target tracking system,
TIME and the results verify its superiority over the conventional
Fig. 5 . MSE of the estimate of position using the robust adaptive Kalman adaptive method under the same conditions.
filter algorithm. A = optimum MSE, B = robust adaptive, C = mea-
surements MSE. Ensemble average over 100 independent trials.
REFERENCES
[ l ] R. E. Kalman, “A new approach to linear filtering and prediction
problems,” Trans. ASME, J. Basic Eng., vol. 82, pp. 35-44, 1960.
mates of position using the proposed adaptive filtering [2] R. K. Mehra, “On the identification of variances and adaptive Kal-
technique is shown in Fig. 5 . Our method shows consid- man filtering,” IEEE Trans. Automat. Contr., vol. AC-15, Apr. 1970.
[3] -, “Approaches to adaptive filtering,” IEEE Trans. Automat.
erably improved results in comparison to those of the con- Conrr., vol. AC-17, Oct. 1972.
ventional adaptive method, except during transient pe- [4] B. Carew and P. R. Bellanger, “Identification of optimum filter
riods. The MSE of the robust estimate is close to the steady-state gain for systems with unknown noise covariances,” lEEE
Trans. Automat. Contr., vol. AC-18, Dec. 1973.
optimum value in the steady-state condition, but the MSE [ 5 ] P. R. Bellanger, “Estimation of noise covariance matrices for a linear
of the conventional adaptive estimate approaches that of time-varying stochastic process,” Automatica, vol. 10, pp. 267-275,
the variance of the measurement. This means that effec- 1974.
[6] N. K. Sinha and A. Tom, “Adaptive state estimation for systems with
tively no filtering has been done. unknown noise covariances,” Inr. J. Sysr. S c i . , vol. 8 , no. 4 , pp.
The MSE’s of the estimates of the velocity using both 377-384, 1977.
robust and conventional adaptive filtering have also been [7] A. P. Sage and G. W. Husa, “Adaptive filtering with unknown prior
statistics,” in Proc. Joint Automat. Contr. Conf., 1969, pp. 760-769.
analyzed. Their results (not shown) are similar to those [8] R. A. Singer, “Estimating optimal tracking filter performance for
for position. However, the MSE of the conventional manned maneuvering targets, ” IEEE Trans. Aerosp. Elecrron. Syst.,
adaptive filtering method remains high during input step vol. AES-6, July 1970.
[9] N. H. Gholson and R. L. Moose, “Maneuvering target tracking using
duration and does not approach its optimum values there- adaptive state estimation,” IEEE Trans. Aerosp. Electron. Syst., vol.
after. AES-13, May 1977.
This example clearly shows the superiority of the pro- [lo] C. B. Chang and J. A. Tabaczynski, “Application of state estimation
to target tracking,” IEEE Trans. Automat. Conrr., vol. AC-29, Feb.
posed adaptive Kalman filter over the conventional, linear 1984.
adaptive Kalman tracking filter in a variety of similar sit- [ l l ] K. A. Myers and B. D. Tapley, “Adaptive sequential estimation with
uations. In all of our simulations, 5 percent of the noise unknown noise statistics,’’ IEEE Trans. Automat. Conrr., Aug. 1976.
[12] F. D. Groutage, “Adaptive robust sequential estimation with appli-
sequences are outliers (i.e., same Gaussian distribution, cation to tracking a maneuvering target,” Ph.D. dissertation, Dep.
but 9 times larger variance). In general, as the noise vari- Elec. Eng., Univ. Wyoming, May 1982.
ance or percentage of the outliers increases, the variabil- [13] R. L. Kirlin and A. Moghaddamjoo, “Robust adaptive Kalman fil-
tering for systems with unknown step input and non-Gaussian mea-
ity of the estimators in both methods increases. Simula- surement errors,” IEEE Trans. Acoust., Speech, Signal Processing,
tion results presented here, however, are related to the vol. ASSP-34, Apr. 1986.
MOGHADDAMJOO AND KIRLIN: ROBUST ADAPTIVE KALMAN FILTERING 1175

[14] D. C. Montgomery and E. A. Peck, Introducrion to Linear Regres- gineering and Computer Science of the University of Wisconsin, Milwau-
sion Analysis. New York: Wiley, 1982. kee, which he joined in September 1986. His teaching and research are
[I51 R. L. Launer and G. N. Wilkinson, Robusfness in Statistics. New primarily in the areas of digital signal processing, adaptive filtering, pattern
York: Academic, 1979. recognition, and image processing.
[16] F. Mosteller and W. Tukey, Data Analysis and Regression. Read-
ing, MA: Addison-Wesley , 1977.
[17] M. Hollander and D. A. Wolf, Nonparamefric Statistical Methods.
New York: Wiley, 1984.
R. Lynn Kirlin (S’66-M’68-SM’82) received the
1181 W. W. Daniel, Applied Nonparametric Srarisrics. Boston, MA: B.S. and M.S. degrees in electrical engineering
Houghton Mifflin, 1978. from the University of Wyoming in 1962 and
[19] D. J. Hudson, “Fitting segmented cuwes whose join points have to 1963, respectively, and the Ph.D. degree in elec-
be estimated,” Amer. Stat. Assoc. J . , Dec. 1966. trical engineering from the University of Utah in
[20] P. J. Huber, “Robust regression: Asymptotics, conjectures, and 1968. He has also had coursework in statistics.
Monte Carlo,” Ann. Sfat., vol. 1, pp. 799-821, 1973.
He has analyzed data transmission circuitry and
[21] H. A. David, Order Statisfics. New York: Wiley, 1981.
an EMP pulse simulation system at Martin-Mar-
[22] R. L. Kirlin, “Robust adaptive Kalman trackers for systems with un-
ietta, Denver, CO, from 1963 to 1964. At Boeing
known step inputs and non-Gaussain delay measurement errors,”
(1965-1966) he performed analysis and design
Tech. Rep. for 1984 to Office of Naval Research, Contract N00014-
tasks for various space communication projects.
82-K-0048, Elec. Eng. Dep., Univ. Wyoming.
At Datel, Riverton, WY, from 1968 to 1969, he had experience in com-
puter peripheral design and analysis. At Floating Point Systems, Beaver-
ton, OR, from 1978 to 1979 he specified and developed a basic image
Alireza Moghaddamjoo (S’84-M’86) was born processing library and a signal processing library addition for the array
in Tehran, Iran, on March 16, 1953. He received processor AP-120B. There he also analyzed 1/0 times for the processor
the B.S. degree in electrical engineering from the with attached disk. From 1969 through 1986 he was on the Faculty at the
University of Tehran, Tehran, Iran, in 1976, the University of Wyoming, in the Electrical Engineering Department, attain-
M.S. degree in nuclear engineering from the Mas- ing the rank of Professor in 1978. During this time he taught nearly 30
sachusetts Institute of Technology, Cambridge, in different courses in electrical engineering, establishing several in commu-
1978, and the Ph.D. degree in electrical engi- nication theory and signal processing. He has also published nearly 40 jour-
neering from the University of Wyoming, Lara- nal papers in control theory applications, signal demodulation and detec-
mie, in 1986. tion, pseudonoise applications, and speech and signal processing. His recent
At the University of Wyoming he was a Re- research and consulting involves image, sonar, and seismic signal process-
search Assistant and a Western Research Institute ing. These have been partially supported by U.S. Naval organizations and
Graduate Scholar. During his Ph.D. program he was involved in several a number of oil companies. He has continued these interests at the Uni-
funded projects, namely, robust Kalman tracking, robust step detection, versity of Victoria, where he has been employed as a Professor of Electrical
covariance estimation and eigenstructure variability, and eigencoding for and Computer Engineering since January 1987. Additionally he is pursuing
seismic data. He held a Lecturer position in the Department of Electrical speech coding for enhancement in noise, spectral analysis, array process-
Engineering at the University of Wyoming for the Spring semester of 1986. ing, and image quality measurements, being supported by NSERC funds
He is currently an Assistant Professor in the Department of Electrical En- and the National Telemetry and Information Agency, Boulder, CO.