A HYBRID KALMAN FILTER-FUZZY LOGIC

MULTISENSOR DATA FUSION ARCHITECTURE

WITH FAULT TOLERANT CHARACTERISTICS

P. J. ESCAMILLA-AMBROSIO, N. MORT
Department of Automatic Control and Systems Engineering
University of Sheffield
Mappin Street, Sheffield S1 3JD, United Kingdom
Phone: +44 (0) 114 222 5619, Fax: +44 (0) 114 222 5661


Abstract- In this work a novel Multi-Sensor Data
Fusion (MSDF) architecture with fault tolerant
characteristics is proposed. This MSDF architecture
is based on Kalman filtering and fuzzy logic
techniques. First, the measurement coming from each
sensor is fed to a fuzzyadapted Kalman filter (FKF).
The adaptation is in the sense of adjusting the
measurement noise covariance matrix R using a fuzzy
inference system (FIS) based on a covariance
matching technique. Second, another FIS, here called
a fuzzy logic observer (FLO), is used to monitor the
performance of each FKF. The FLO assigns a degree
of confidence to each one of the FKFs. The degree of
confidence indicates to what level each FKF output
reflects the true value of the parameter being
measured. At this level transient sensor faults are
eliminated. Finally, the fused estimated measurement
is obtained through a defuzzification process based
on these confidence values. At this level persistent
sensor faults are eliminated using a voting scheme.
To demonstrate the effectiveness and accuracy of this
hybrid MSDF architecture, an example with four
noisy and faulty sensors is outlined. The results show
very good performance.

Key words: Multi-sensor data fusion, fault tolerant
systems, fuzzy logic, adaptive Kalman filtering.


1. Introduction
The Multi-Sensor Data Fusion (MSDF)
approach is described as the acquisition,
processing, and synergistic combination of
information gathered by various knowledge
sources and sensors to provide a better
understanding of a phenomenon under
consideration [1]. Different MSDF techniques
have been explored recently. These techniques
vary from those based on well-established
Kalman filtering methods [2, 3] to those based
on more recent ideas coming from soft
computing technology [4, 5]. However, little
work has been done in exploring hybrid
architectures that consider both these
approaches. In this work a novel MSDF
architecture with fault tolerant characteristics is
proposed. This architecture is based on a hybrid
structure integrating fuzzy inference systems
and Kalman filtering techniques.
The general idea explored here is the
combination of the advantages that both
techniques have. On the one hand, Kalman
filtering is recognised as one of the most
powerful traditional techniques of estimation
[6]. On the other hand, the main advantages
derived from the use of fuzzy logic techniques,
with respect to traditional schemes, are the
simplicity of the approach, the capability of
fuzzy systems to deal with imprecise
information, and the possibility of including
heuristic knowledge about the phenomenon
under consideration.
The remainder of this paper is organised as
follows. Section 2 describes the Kalman filtering
technique and summarises the proposed
approach of fuzzy-adapted Kalman filtering [7].
Section 3 describes the proposed new hybrid
MSDF architecture. To show the effectiveness
of this approach and its fault tolerant
characteristics, in section 4 an illustrative
example with four noisy and faulty sensors is
outlined and results are discussed. Finally, the
conclusions of this work are given in section 5.

2. Kalman filtering
The Kalman filter is an optimal recursive data
processing algorithm [6] that provides a linear,
unbiased, and minimum error variance estimate
of the unknown state vector
n
k
x at each
instant k = 1,2,, (indexed by the subscripts) of
a discrete-time controlled process described by
the linear stochastic difference equations:


k k k k k k
w u B x A x + + =
+1
(1)

k k k k
v x H z + = (2),
where x
k
is an n 1 system state vector, A
k
is an
n n transition matrix, u
k
is an l 1 vector of
the input forcing function, B
k
is an n l matrix,
w
k
is an n 1 process noise vector, z
k
is a m 1
measurement vector, H
k
is a m n measurement
matrix, and v
k
is a m 1 measurement noise
vector.
Both w
k
and v
k
are assumed to be
uncorrelated zero-mean Gaussian white noise
sequences with covariances:

{ }

=
=
k i
k i Q
w w E
T
i k
0
,
k
,
{ }

=
=
k i
k i R
v v E
T
i k
0
,
k
{ } 0 =
T
i k
v w E for all k and i (3),

where E{} is the statistical expectation,
superscript T denotes transpose, Q
k
is the
process noise covariance matrix, and R
k
is the
measurement noise covariance matrix.
The Kalman filter algorithm [8, 9] is
organised in two groups of equations:
i) Time update (or prediction) equations:


k k k k k
u B x A x + =

+

1
(4)

k
T
k k k k
Q A P A P + =

+1
(5).

These equations project, from time step k to step
k+1, the current state and error covariance
estimates to obtain the a priori (indicated by the
super minus) estimates for the next time step.
ii) Measurement update (or correction)
equations:


1
] [

+ =
k
T
k k k
T
k k k
R H P H H P K (6)
] [

+ =
k k k k k k
x H z K x x (7)

=
k K k k
P H K I P ] [ (8).

These equations incorporate a new measurement
into the a priori estimate to obtain an improved
a posteriori estimate.

2.1. Fuzzy-adapted Kalman filtering
Assuming that the process noise covariance
matrix Q is known, here an innovation-based
adaptive estimation (IAE) algorithm [7, 12] to
adapt the measurement noise covariance matrix
R and prevent filters divergence [10, 11] has
been derived. In particular, the technique known
as covariance-matching is used [13]. The basic
idea behind this technique is to make the actual
value of the covariance of the residual consistent
with its theoretical value. This is done in two
steps; first, the innovation sequence or residual r
is defined by:

) (

=
k k k k
x H z r (9),
and its theoretical covariance is defined by,

k
T
k k k k
R H P H S + =

(10),

obtained from the Kalman filter algorithm.
Second, if it is found that the actual covariance
of r
k
has a discrepancy with its theoretical value,
then adjustments are made to R in order to
correct this mismatch.
Given the availability of the innovation
sequence r
k
, its actual covariance
r
C

is
approximated by its sample covariance [12]
through averaging inside a moving estimation
window of size N,

=
=
k
i i
T
i i rk
r r
N
C
0
1
(11),

where
1
0
+ = N k i
is the first sample inside the
estimation window. The window size N is
chosen empirically to give some statistical
smoothing.
Now, a new variable called the Degree of
Matching (DoM), is defined to detect the
discrepancy between S and
r
C

. This is:

rk k k
C S DoM

= (12).

The basic idea of adaptation used by a Fuzzy
Inference System (FIS) to derive an adjustment
of R is as follows. It can be noted from Eq. 10
that an increment in R will increment S, and vice
versa. Thus, R can be used to vary S in
accordance with the value of DoM in order to
reduce the discrepancies between S and
r
C

.
From here three general rules of adaptation are
defined:

1. If DoM 0 (this means S and
r
C

match
almost perfectly) then maintain R
unchanged.
2. If DoM > 0 (this means S is greater than
its actual value
r
C

) then decrease R.
3. If DoM < 0 (this means S is smaller than
its actual value
r
C

) then increase R.

Thus R is adjusted in this way:

k k k
R R R + =
1
(13),

where R is the factor that is added or
subtracted from R at each instant of time. R is
the FIS output and DoM is the FIS input. A
graphical representation of the Fuzzy-adapted
Kalman Filter (FKF) is shown in Fig. 1.

Sensor
Kalman
Filter
Fuzzy
Inference
System
Compute

rk
C

R
k-1
k
z
z
k
S
k
DoM
k
r
k
R
k
R
k
FKF

Fig. 1. Basic structure of the Fuzzy-
adapted Kalman Filter (FKF).

3. Hybrid multi-sensor data fusion
architecture
The objective of the proposed MSDF
architecture is to obtain fused measurement data
that determines the parameter being measured as
precisely as possible. The main characteristic of
this architecture is its fault tolerance.
In the proposed MSDF architecture it is
assumed that there are n different sensors
measuring the same parameter. Each sensor has
its own characteristics of noise and
measurement errors. The measurements
obtained from these sensors are fed to a fuzzy-
adapted Kalman filter (FKF), one for each
sensor; thus there are n sensors and n FKFs
working in parallel (see Fig. 2). A FIS, here
called the Fuzzy Logic Observer (FLO), is used
both to monitor the performance of each FKF
and for the detection of transient sensor faults.
To achieve this, the FLO was designed using
three variables: the absolute value of DoM
(AbsDoM), the value of R, and a variable called
the Residual Compatibility rC.

Sensor 1 FKF 1
Fusion
centre
FLO 1
w
1
Sensor 2 FKF 2
FLO 2
w
2
Sensor n FKF n
FLO n
w
n

2 2
w z
1 1
w z
n n
w z
1
z
2
z
n
z
z
DoM
R
DoM
R
DoM
R
1
z
2
z
n
z
r
S
r
S
r
S
Fig. 2. New hybrid MSDF architecture.
The variable rC is obtained using the
following relation:


k
k
k
S
r
rC =
(14).

This value gives a measure of the actual
amplitude of the residual r compared to its
theoretical value S . For a sensor without
transient faults the value of rC is around 1, if
this value changes abruptly, this means that a
transient fault on the corresponding sensor is
present.
The FLO assigns a degree of confidence w, a
number on the interval [0, 1], to each one of the
FKF (see Fig. 2). w indicates to what level each
FKF output reflects the true value of the
measurement and if this value is coming from a
non-faulty sensor (transient faults). At the same
time, the degree of confidence acts as a weight
that indicates to the fusion centre at what
confidence level it should take each FKF output
value.
The fusion centre (see Fig. 2) obtains the
fused estimation of the parameter being
measured based on the confidence values. First
of all, the fusion centre applies a voting
technique to investigate if there are persistent
sensor faults. If one of the estimated
measurements differs markedly from the others
this means a persistent sensor fault is present in
that sensor, thus that value is not considered for
fusion purposes. Finally a defuzzification
process is applied to obtain the fused estimation
of the parameter being measured. Fig. 2 shows a
graphical representation of the proposed MSDF
architecture.

4. Illustrative example
Consider the following linear system, which is a
modified version of a tracking model [14, 15],

(
(
(

+
(
(
(

(
(
(

=
(
(
(

+
+
+
3
2
1
3
2
1
3
1
2
1
1
1
75 . 0 00 . 0 05 . 0
25 . 0 75 . 0 25 . 0
00 . 0 0.20 0.77
k
k
k
k
k
k
k
k
k
w
w
w
x
x
x
x
x
x
(15a)
[ ]
k
k
k
k
k
v
x
x
x
z +
(
(
(

=
3
2
1
0 0 1
(15b),

with initial conditions 0
0
= x ,
3 0
01 . 0 I P =
,
where x
1
, x
2
, and x
3
are the position, velocity and
acceleration, respectively, of a flying object. In
Eq. 15, the system noise sequence {w
k
} is
uncorrelated zero-mean Gaussian noise with Q =
0.02I
3
.
MATLAB code was developed to simulate
the process described by Eq. 15 and the
proposed MSDF architecture considering four
sensors measuring the position of the flying
object. The simulation was carried out for 500s
with a sample time of 0.5s. Q was fixed as
0.02I
3
. The actual value of R for each sensor was
assumed unknown. The starting value of R in all
sensors was selected as 1. Each sensor
measurement is corrupted by the noise profiles
described in Fig. 3. Additionally, persistent
faults are introduced on sensor 1 at time ranges
[100-200], [300-310], and [400-410]. During
these ranges the measurement of sensor 1 is
fixed as 15. Transient faults are introduced on
sensor 4 at times 150, 250, 350 and 450. At
these times the measurement of sensor 4 is fixed
as 20.





Fig. 3. (a) Noise on sensor 1, (b) noise on sensor
2, (c) noise on sensor 3, (d) noise on sensor 4.

In subsequent sections, the implementation
of each one of the components of the hybrid
MSDF architecture is described.

4.1 Fuzzy-adapted Kalman Filter (FKF)
Each FKF was implemented considering five
fuzzy sets for DoM: NM = Negative Medium,
NS = Negative Small, ZE = Zero, PS = Positive
Small, and PM = Positive Medium; and five for
R: IL = Increase Large, I = Increase, M =
Maintain, D = Decrease, and DL = Decrease
Large. The membership functions for DoM and
R are presented in Fig. 4.



Fig. 4. Membership functions for DoM and R.

Following the general guidelines given in
section 2.1 five fuzzy rules have been
formulated in order to adjust R, these are:

1. If DoM = NM, then R = IL
2. If DoM = NS, then R = I
3. If DoM = ZE, then R = M
4. If DoM = PS, then R = D
5. If DoM = PM, then R = DL.

Thus using the compositional rule of
inference sum-prod, R is adjusted as mentioned
in Eq. 13. The window size N was selected
empirically as 15.

4.2 Fuzzy Logic Observer (FLO)
The membership functions for AbsDoM, R and
rC are shown in Fig. 5. Here the fuzzy labels
mean: ZE = Zero, S = Small, L = Large, NF =
No Fault, and F = Fault. For the output w, 3
fuzzy singletons were defined with the labels:
G=1=Good, AV=0.5=Average, and P=0=Poor.
Thus 18 rules complete the FLO rule base, as
shown in Table 1.

Table 1
rC = NF
R
AbsDoM

ZE

S

L
ZE G G AV
S G AV P
L AV P P
rC = F
R
AbsDoM

ZE

S

L
ZE P P P
S P P P
L P P P
(a)
(b)
(c)
(d)



Fig. 5. Membership functions for AbsDoM, R,
and rC.

The above rules are based on two simple
heuristic considerations. First, if the AbsDoM is
near to zero and R is near to zero and no
transient fault is present, then it means the filter
is working almost perfectly. Second, if one or
both of AbsDoM and R increases far from zero
that means the filter performance is degrading.
At the same time if rC increases abruptly, this
means that a transient fault on the corresponding
sensor has been detected.
Thus, using the compositional rule of
inference sum-prod and the centre of area
defuzzification method, the FLO obtains a
degree of confidence for each FKF.

4. 3 The fusion centre
As a first stage in the fusion centre (see Fig. 2) a
voting technique [16] is used to detect persistent
sensor faults. This is done by calculating the
factors:

2 1 12
z z C = ,
3 1 13
z z C = ,
4 1 14
z z C = ,
3 2 23
z z C = ,
4 2 24
z z C = ,
4 3 34
z z C =
(16),

and then applying the following crisp rules:

1. If C
12
and C
13
and C
14
are then sensor 1
has a persistent fault.
2. If C
12
and C
23
and C
24
are then sensor 2
has a persistent fault.
3. If C
13
and C
23
and C
34
are then sensor 3
has a persistent fault.
4. If C
14
and C
24
and C
34
are then sensor 4
has a persistent fault.
In these rules is a threshold factor used to
indicate the degree of disagreement between
each estimated measurement and the others. If
one of the estimated measurements differs
markedly from the others then this signal is
eliminated and the corresponding sensor is
assumed to be in a persistent fault state. Here
was selected as 1.2. Finally, a defuzzification
process is applied to obtain the fused estimation
of the parameter being measured. In this case the
centre of area method is used,

=
=
=
n
i
ki
n
i
ki ki
k
w
w z
z
1
1

(17),

where
ki
z is the output of the i-th FKF
(i=1,2,3,4) and
ki
w is its respective degree of
confidence at instant of time k. Thus, each FKF
output is weighted according to its
corresponding degree of confidence w, and if it
is found in the persistent fault state, then this
output and its corresponding weight are not
considered for fusion purposes.
In order to prevent possible conflicts, the
following crisp rule was incorporated: If the sum
of the degrees of confidence is equal to zero,
then the fused output is simply the average of
the non-faulty FKF outputs.

4. 4 Results
For comparison purposes, the following
performance measures were adopted:

=
=
n
k
k k zv
z za
n
J
1
2
) (
1
(18)

=
=
n
k
k k ze
z za
n
J
1
2
) (
1
(19),

where za
k
is the actual value of the position; z
k
is
the measured position; and
k
z is the estimated
position at an instant of time k; n = No. of
samples.
Table 2 shows the performance measures of
each individual FKF and those obtained by the
proposed MSDF architecture where the fusion
of the four sensors is made for both cases: with
non-faulty and faulty sensors (NFS and FS).
From the analysis of the data, it is noted that the
most accurate estimation of the position is
obtained with the MSDF architecture for both
cases NFS and FS. For the case of NFS the error
in the estimation is 6.24% less compared to that
obtained with FKF 1 (for sensor 1), which has
the best individual performance measure. At the
same time this error is 58.86% less compared to
that obtained with FKF 2 (for sensor 2) which
has the worst individual performance measure.
Fig. 6 shows the actual and fused estimated
position and the corresponding error in the
estimation for the NFS case.

Table 2
Without faults With faults
Sensor J
zv
J
ze
J
zv
J
ze

Sensor 1 0.7285 0.2693 7.8075 7.4804
Sensor 2 3.9066 0.6137 3.9066 0.6137
Sensor 3 2.5633 0.4200 2.5633 0.4200
Sensor 4 2.5928 0.3973 2.9549 0.6371
Fused 0.2525 0.2831



Fig. 6. Actual and fused estimated position and
the corresponding error obtained with the MSDF
architecture for non-faulty sensors.

For the case of faulty sensors the
performance measure for the fused estimated
position is only a little larger than that for the
case previously analysed (see Table 2). The
error only increases by 12.12% in spite of the
faults introduced on sensor 1 and 4.
Additionally, this estimation error is 32.6%
smaller than that observed in the best individual
FKF (in this case number 3). Thus both fused
estimated measurements are more accurate than
those obtained with any individual FKF. Fig. 7
shows the estimated position obtained with FKF
1; here the effects caused by the introduction of
persistent sensor faults can be seen. Fig 8 shows
the estimated position obtained with FKF 4; here
the effects caused by the introduction of
transient sensor faults can be observed. Finally,
Fig. 9 shows the actual and fused estimated
position and the corresponding error for the
faulty sensors case.


Fig. 7. Estimated position obtained with FKF 1.


Fig 8. Estimated position obtained with FKF 4.



Fig. 9. Actual and fused estimated position and
the corresponding error obtained with the MSDF
architecture for faulty sensors.

5. Conclusions
A novel hybrid MSDF architecture integrating
Kalman filtering and fuzzy logic techniques has
been presented. The main characteristic of this
architecture is its tolerance to persistent and
transient sensor faults. This characteristic is
obtained by exploiting the advantages that both
approaches have: the optimality of the Kalman
filter and the capability of fuzzy systems to deal
with imprecise information using fuzzy sets and
common sense rules.
In this novel approach the linear estimations
of the individual Kalman filters are improved
through the adaptation of the measurement noise
covariance matrix R by means of a FIS. This
prevents filter divergence and relaxes the a
priori assumption of the value of R. It is
noteworthy that only five rules were needed to
carry out this adaptation.
The role of the FLO in the proposed MSDF
architecture is of great importance. This is
because the fusion of the information is carried
out based on the degrees of confidence
generated on this element. At this level transient
sensor faults are automatically removed.
The results obtained in the illustrative
example are promising. They show that this
novel hybrid MSDF architecture is effective in
situations where there are several sensors
measuring the same parameter and each sensor
measurement is contaminated with a different
kind of noise. This approach is also capable of
tolerating transient and persistent sensor faults
in an efficient way. Both fused estimated
measurements (with faulty sensors and non-
faulty sensors) produced better approximations
to the actual value of the parameter being
measured than that obtained with any single
FKF. Thus the general idea of exploring the
combination of traditional together with non-
traditional techniques appears to be a promising
avenue of investigation.
The system employed to illustrate the
effectiveness of the approach presented is
simple and only one parameter is considered as
being measured. However the approach can be
easily extended for systems with more than one
parameter being measured. In fact this is the
subject of current work being done by the
authors.

References
[1] Varshney, P. K. Multisensor data fusion.
Electronics and Communication Engineering
Journal, 9 (6), pp. 245-253, 1997.
[2] Luo, R. C. Multisensor integration and
fusion in intelligent systems. IEEE
Transactions on Systems, Man, and
Cybernetics, 19 (5), pp. 901-931, 1989.
[3] Gao, Y., Krakiwsky, E. J., Abousalem, M.
A., McLellan, J. F. Comparison and analysis
of centralized, decentralized, and federated
filters. Navigation: Journal of The Institute of
Navigation, 40 (1), pp. 69-86, 1993.
[4] Kuo, R. J. and Cohen P. H. Multi-sensor
integration for on-line tool wear estimation
through radial basis function networks and
fuzzy neural networks. Neural Networks, 12,
pp. 355-370, 1999.
[5] Kobayashi F., Arai, F., Fukuda, T.,
Shimojima, K., Onoda, M., and Mauri, N.
Sensor fusion system using recurrent fuzzy
inference, Journal of Intelligent and Robotic
Systems, 23, pp. 201-216, 1998.
[6] Maybeck, P. S. Stochastic models estimation
and control. Volume 1, Academic Press, New
York, 1979.
[7] Escamilla-Ambrosio, P. J. and Mort, N.
Adaptive Kalman filtering through fuzzy logic.
Proceedings of the 7
th
UK Workshop On
Fuzzy Systems, Whirlow Grange Conference
Centre, Sheffield, U. K., pp. 67-73, 2000.
[8] Welch, Greg and Bishop, Gary. An
Introduction to the Kalman Filter. The
University of North Carolina at Chapel Hill,
TR95-045, 1995.
[9] Brown, R. G. and Hwang P. Y. C.
Introduction to random signals and applied
kalman filtering whit matlab exercises and
solutions. Third edition, John Wiley & Sons,
New York, 1997.
[10] Mehra, R. K. On the identification of
variances and adaptive kalman filtering. IEEE
Trans. Automatic Control, AC-15 (2), pp. 175-
184, 1970.
[11] Fitzgerald, R. J. Divergence of the Kalman
filter. IEEE Trans. Automatic Control, AC-16
(6), pp. 736-747. 1971.
[12] Mohamed, A. H. and Shwarz, K. P.
Adaptive kalman filtering for INS/GPS.
Journal of Geodesy, 73 (4), pp. 193-203, 1999.
[13] Mehra, R. K. Approaches to adaptive
filtering. IEEE Trans. Automatic Control, AC-
17, (Oct), pp. 693-698, 1972.
[14] Paik, B.S. and Oh, J.H. Gain fusion
algorithm for decentralised parallel Kalman
filters. IEE Proc. Control Theory Appl., 147
(1), pp. 97-103, 2000.
[15] Chen, G. and Chui, C. K. A modified
adaptive Kalman filter for real-time
applications. IEEE Trans. Aerospace and
Electronic Systems, 27 (1), pp. 149-154, 1991.
[16] Caspi, Paul and Sale, Ryam. Threshold and
Bounded-Delay Voting in Critical Control
Systems. Lecture notes in Computer Science,
1926, pp. 70-81, 2000.