2008 Internatioal Symposium on Telecommunications

Improvement of Extended Kalman Filter

Frequency Tracker for Nonstationary Harmonic Signals
H. Hajimolahoseini #1, M. R. Taban #2, H. R. Abutalebi #3
#
Signal Processing Research Lab, Electrical Engineering Department, Yazd University, Yazd, Iran

Abstract—In this paper, we have considered a frequency

tracking method based on Extended Kalman Filtering (EKF).
The method uses a state space model to estimate and track the ⎧ x1 (t + 1) = x1 (t ) cos( x 3 (t )) − x 2 (t ) sin( x 3 (t ))
⎪
frequency of a harmonic signal embedded in broad-band noise. ⎪ x 2 (t + 1) = x1 (t ) sin( x3 (t )) + x 2 (t ) cos( x3 (t ))
This signal is generally characterized by time-varying frequency ⎨ (1)
and amplitude (nonstationary noisy harmonic signal). In this ⎪ x3 (t + 1) = x 3 (t ) + w(t )
paper we have proposed a new state-space model for estimate ⎪⎩ y (t ) = x1 (t ) + v(t ).
and track the frequency of such nonstationary signals. The
conventional EKF frequency trackers are typically characterized Here, w(t ) and v(t ) are zero-mean uncorrelated white
by a vector of three tuning parameters. With some assumptions
and without any performance loss, we have proposed an EKF noises. Also, The state variable x 3 (t ) represents the unknown
frequency tracker that only uses one tuning parameter. This frequency ω (t ) .
simplification allows an easier and more transparent tuning of
the EKF tracking behaviour. In the above-mentioned works ([5]–[9]), it has been
assumed that the harmonic signal s (t ) has a time-variant
I. INTRODUCTION frequency but they have not any assumption about variation of
The problem of determining the frequency of a signal is a signal amplitude. In general case, the amplitude of harmonic
common one; however, its inherent nonlinearity has also made signal can be time-variant or may be corrupted; so, the third-
it difficult. There is a long history of the application of order state space model (1) no longer fulfills this situation.
extended Kalman filter (EKF) methods in frequency In this paper, we show that by adding one additional state
estimation and tracking going back at least to Snyder [l]. variable (that represents the variations of signal amplitude) to
This paper deals with the problem of estimating the third-order state-space model (1), the performance of EKF
frequency ω (t ) of an harmonic signal frequency tracker will extremely increase for nonstationary
harmonic signals (harmonic signals whose both amplitude and
s (t ) = A(t ) cos(ω (t ) × t + ϕ ) (with time-varying amplitude and
frequency are time-variant i.e. s (t ) = A(t ) cos(ω (t ) × t + ϕ ) ).
frequency), given its noisy measurement y (t ) = s (t ) + n(t ) ,
The EKF frequency tracker will be applied on the model to
where n(t ) is a broad-band stationary signal (e.g., a white estimate and track the frequency ω (t ) of s (t ) , given its noisy
noise). This problem is frequently encountered in real-world measurement y (t ) = s (t ) + n(t ) where n(t ) is a broad-band
applications, especially in the fields of adaptive control, signal
processing, radar and communication. Numerous techniques stationary signal (e.g. a white noise). Then, by some
have been developed for the treatment of this problem (for e.g., assumptions, we will simplify the parameterization of EKF
see [2]-[9]). In this research, we have focused on the use of frequency tracker. Finally, by means of simulation, it will be
the Extended Kalman Filter (EKF), an approach proposed by shown that in comparison with other procedures, our proposed
Bitmead, et al. in [5]-[8] for the design of a frequency tracker. algorithm has the advantage of better working in following
In particular, [5] proposes the EKF frequency tracker and non-stationary conditions:
• noisy environments,
provides some heuristic guidelines for the tuning of its
• variations of signal amplitude and frequency.
parameters. In [7], the stability analysis of EKF is developed,
whereas the case of high-noise environments is considered in II. SIGNAL MODEL
[6].
The goal of [8] and [9] is to go a step further in the analysis Kalman filter design and, by implication, EKF estimator
of the EKF frequency tracker, in order to gain more insight in design proceeds from a state space signal model of the process
the relationships between its behaviour and its tuneable to be estimated. The signal model dynamics describe a
parameters. Especially, it is shown in [8] that the performance mechanism for how the process may be evolving. Thus, the
of EKF frequency tracker depends on the relationship between initial stage in the filter design is to model the signal and
its parameters. In [9], the parameters of the EKF frequency derivation of state space of it.
tracker are optimized via genetic algorithms. As it was described in previous section, our signal model is:
The basic idea behind these papers ([5]–[9]) is the
modeling of measured signal y (t ) = s (t ) + n(t ) by means of a y (t ) = s (t ) + n(t ) (2)
third-order state space model, as follows:

978-1-4244-2751-2/08/$25.00 ©2008 IEEE 592

where n(t ) is a broad-band stationary signal (e.g. a white w(t ) , u (t ) and v(t ) are zero-mean uncorrelated white noises,
noise) and: having variances q , p and r , respectively. The state variable
x 3 (t ) represents the unknown frequency ω (t ) .
s (t ) = A(t ) cos(ω (t ) × t + ϕ ) . (3)
Given the state-space model (11), the EKF technique can
be applied straightforwardly. The complete set of equations of
For derivation of state space model, we can write: the EKF frequency tracker derived from (11) is here presented
(see, e.g. [10]-[11])
s (t + 1) = A(t + 1) cos(ω (t + 1) × (t + 1) + ϕ ). (4)
 
Using the cosine expansion: ⎧ x (t | t ) = f ( x (t − 1 | t − 1))
⎪ 
+ K (t )( y (t ) − H ( f ( x (t − 1 | t − 1)))
⎪  
cos( P + Q ) = cos( P ) cos(Q) − sin( P) sin(Q ) (5) ⎪ ω (t ) = Cx (t | t )
⎪⎪
we rewrite equation (4) as: (
⎨ K (t ) = P(t ) H T HP(t ) H T + r
⎪
−1
)
s (t + 1) = A(t + 1)[cos(ω (t + 1) × t + ϕ ) cos(ω (t + 1)) ⎪ ⎢⎣
(
⎪P (t + 1) = F (t ) ⎡ P(t ) − P (t ) H T HP(t ) H T + r )−1
HP (t )⎤
⎥⎦
(6) ⎪
− sin(ω (t + 1) × t + ϕ ) sin(ω (t + 1))]. ⎪⎩ × F T (t ) + I qp

According to third equation of state space equation (1), we (12)

assume that:
where
ω (t + 1) ≅ ω (t ). (7)

f ( x (t | t )) =
For amplitude of s (t + 1) , A(t + 1) , we also assume that:     
⎡ x 4 (t | t )[ x1 (t | t ) cos( x 3 (t | t )) − x 2 (t | t ) sin( x3 (t | t ))]⎤
⎢     ⎥
A(t + 1) ≅ α (t ) A(t ) (8) ⎢ x 4 (t | t )[ x1 (t | t ) sin( x3 (t | t )) + x 2 (t | t ) cos( x 3 (t | t ))]⎥
⎢ x 3 (t | t ) ⎥
⎢  ⎥
where α (t ) is a coefficient that is considered as a new state ⎣⎢ x 4 (t | t ) ⎦⎥
variable (as it will be shown later). Substituting (7) and (8) in
(6) yields: H = [1 0 0 0] C = [0 0 1 0]

s (t + 1) = α (t ) A(t )[cos(ω (t ) × t + ϕ ) cos(ω (t )) and

(9)
− sin(ω (t ) × t + ϕ ) sin(ω (t ))]. 
⎡ x 1 (t | t ) ⎤
If we define state variables as the followings: ⎢ ⎥
 x 2 (t | t )⎥ ∂f ( x) 
x (t | t ) = ⎢  , F (t ) = | x = x (t −1|t −1)
⎢ x 3 (t | t ) ⎥ ∂x
x1 (t ) = A(t ) cos(ω (t ) × t + ϕ ) = s (t ) ⎢ ⎥
⎢⎣ x 4 (t | t )⎥⎦
x 2 (t ) = A(t ) sin(ω (t ) × t + ϕ )
(10)
x3 (t ) = ω (t ) I qp = diag (0,0, q, p )
x 4 (t ) = α (t )
Recursive equations in (12) are suitable for on-line
the fourth-order state-space model of measured signal frequency tracking. In particular, the fourth equation in (12),
y (t ) = s (t ) + n(t ) can be written as: giving a recursion for the auxiliary matrix P(t ) of dimension
4×4, is the celebrated Riccati equation [12].
⎧ x1 (t + 1) = x 4 (t )[ x1 (t ) cos( x3 (t )) − x 2 (t ) sin( x3 (t ))]
⎪
⎪ x 2 (t + 1) = x 4 (t )[ x1 (t ) sin( x3 (t )) + x 2 (t ) cos( x3 (t ))]
⎪
⎨ x3 (t + 1) = x3 (t ) + w(t ) (11)
⎪ x (t + 1) = x (t ) + u (t )
⎪ 4 4
⎪⎩ y (t ) = x1 (t ) + v(t ).

Figure 1. Measured signal y(t).

593
 The structure of the EKF frequency tracker is If we also define the ratio of q and r as:
diagrammatically depicted in Fig.1. As with any frequency
tracker, from an external point of view it can be seen as a q
device receiving the measured signal y (t ) as input, and λ= , (17)
  r
delivering the estimated frequency ω (t ) = x3 (t | t ) as output.
The internal structure is constituted by two main sub-blocks: it can be shown (see [8], [12]) that the Riccati equation and
the first performs the updating of the state equations, the gain (i.e., K (t ) ) of EKF change in the form of:
according to the first expression in (12); the second has the
task of providing the gain vector at each time step and is based
on the third and fourth equations of (12).
⎢⎣
(
P(t + 1) = F (t ) ⎡ P(t ) − P(t ) H T HP(t ) H T + λ )
−1
HP(t )⎤
⎥⎦
Notice that our EKF frequency tracker is parameterized in
{ q, p, r }. These three parameters are the “knobs,” by means × F T (t ) + I11
of which the designer can obtain suitable tracking
performance. (
K (t ) = P(t ) H T HP(t ) H T + λ )−1
. (18)
III. PARAMETERIZATION ANALYSIS
Therefore, the performance of the EKF frequency tracker
In [5] and [7] a preliminary analysis of the roles played by
the parameters has been given, and some guidelines for their q
depends only on the ratio λ = and so, the problem of
tuning have been proposed proposed. Such analysis, having r
the objective of explaining the effects of each parameter on tuning of three design parameters { q, p, r } can be reduced
the EKF performance, is mainly based upon Monte Carlo to tuning of just one design parameter λ . With these results,
simulation trials. Even though some useful and interesting the EKF equations (12) can be rewritten as:
indications have been obtained in this way, the tuning of
parameters still remains a difficult task due to the unclear  
cross-relationships between such parameters. ⎧ x (t | t ) = f ( x (t − 1 | t − 1))
⎪ 
In this section, considering the works of [8] and [12], with ⎪  + K (t )( y (t ) − H ( f ( x (t − 1 | t − 1)))
some assumption and without loss of performance, we 
⎪ ω (t ) = Cx (t | t )
parameterize the extended Kalman frequency tracker by ⎪⎪
means of only one design parameter instead of three (
⎨ K (t ) = P(t ) H T HP(t ) H T + λ
⎪
−1
)
parameters { q, p, r }. This simplification provides an easier
and more transparent tuning of the EKF tracking behaviour. ⎪ ⎢⎣
(
⎪ P (t + 1) = F (t ) ⎡ P (t ) − P(t ) H T HP(t ) H T + λ )
−1
HP(t )⎤
⎥⎦
If we assume that: ⎪
⎩⎪ × F T (t ) + I11
q= p (13) (19)

and define the I11 matrix as follow: where


I11 = diag (0,0,1,1) , (14) f ( x (t | t )) =
    
⎡ x 4 (t | t )[ x1 (t | t ) cos( x 3 (t | t )) − x 2 (t | t ) sin( x3 (t | t ))]⎤
according to definition of I qp in (12), we can conclude: ⎢     ⎥
⎢ x 4 (t | t )[ x1 (t | t ) sin( x3 (t | t )) + x 2 (t | t ) cos( x 3 (t | t ))]⎥
⎢ x 3 (t | t ) ⎥
I qp = q × I11 ; (15) ⎢  ⎥
⎣⎢ x 4 (t | t ) ⎦⎥

H = [1 0 0 0] C = [0 0 1 0]
therefore, with this assumption, the Riccati equation (the
fourth equation) in (12) changes in the form of: and

(
P(t + 1) = F (t ) ⎡ P(t ) − P(t ) H T HP(t ) H T + r
⎢⎣
)
−1
HP(t )⎤
⎥⎦

⎡ x 1 (t | t ) ⎤
⎢ ⎥
 ⎢ x 2 (t | t )⎥ ∂f ( x) 
× F T (t ) + qI11 . x (t | t ) =  , F (t ) = | x = x (t −1|t −1)
⎢ x 3 (t | t ) ⎥ ∂x
⎢ ⎥
(16) ⎣⎢ x 4 (t | t )⎦⎥

594
 I11 = diag (0,0,1,1) In order to compare performance of proposed technique
with that of [8] for estimating of frequency of the harmonic
Tracker (19) is a significant simplification of (12); the signal embedded in y (t ) , the EKF frequency trackers (19)
vector of the “tunable” parameters of (19) is now constituted and [8] are used.
by a single parameter λ . In conclusion, the task of tuning the In order to obtain more reliable results, the value of λ is
tuning parameters of the EKF frequency tracker (12) chosen near zero ([8], [12]). The corresponding frequency
(parameterized with { q, p, r }) boils down to that of tuning estimation results for EKF frequency tracker [8] and our
the tracker (19), where a single parameter { λ } has to be proposed EKF frequency tracker, are displayed in Fig. 3 (a)
chosen. In this way, when λ increases, the error variance and (b) respectively (where f1=10(kHz), f2=20(kHz)).
diminishes, however, its tracking ability deteriorates; in As shown in Fig. 3, since the algorithm of [8] has no
opposition, when λ decreases, the tracking process becomes assumption about the variation of signal amplitude, our
more robust for state variations, but the error variance grows proposed EKF frequency tracker has resulted to superior
(see [8], [9]). performance compared to that of [8].
In the previous simulation, we assumed that f1=10(kHz) and
IV. SIMULATION RESULTS f2=20(kHz) and therefore f2=2f1. Now, if we increase the
deference between f1 and f2, for example f1=4(kHz) and
The following simulations demonstrate the advantages of
f2=30(kHz), and therefore f2=7.5f1, the simulation results will
our proposed technique in comparison with other techniques
be in the form of Fig. 4. The corresponding frequency
(specially the method proposed in [8]) for frequency tracking
estimation results for EKF frequency tracker [8] and our
in two following nonstationary conditions:
proposed EKF frequency tracker are displayed in Fig. 4 (a)
• noisy environments,
and (b) respectively.
• variations of signal amplitude and frequency.
First, assume a harmonic signal, characterized by time-
varying frequency and amplitude and embedded in white
noise. Specifically, y (t ) = s (t ) + n(t ) for 0ms ≤ t ≤ 10ms ,
where n(t ) ≈ WN (0,0.02) and s(t) is defines by means of
following two equations:

s(t ) = exp(−500t ) cos( f1 × 2πt ) + n(t )

for 0ms ≤ t ≤ 5ms ,

and

s (t ) = exp(−500(t − 500)) cos( f 2 × 2πt ) + n(t )

for 5ms < t ≤ 10ms .

The measured signal y(t) for f1=10(kHz) and f2=20(kHz) has (a)
been shown in Fig. 2.

(b)
Figure 2. Measured signal y(t). Figure 3. (a) Frequency tracking of the EKF frequency tracker proposed in
[8]. (b) Frequency tracking of our proposed EKF frequency tracker (19).

595
 (a) (a)

(b) (b)
Figure 4. (a) Frequency tracking of the EKF frequency tracker proposed in Figure 6. (a) Frequency tracking of the EKF frequency tracker proposed in
[8]. (b) Frequency tracking of our proposed EKF frequency tracker (19). [8]. (b) Frequency tracking of our proposed EKF frequency tracker (19)

As shown in Fig. 4, the EKF frequency tracker [8] can not The corresponding frequency estimation results for EKF
track the frequency jumping of y(t) whereas our proposed frequency tracker [8] and our proposed EKF frequency tracker
EKF frequency tracker has tracked it. are displayed in Fig. 6 (a) and (b), respectively. (In order to
obtain more reliable results, the value of λ is chosen near
zero ([8], [12])).
Finally, consider a harmonic signal embedded in white
noise, characterized by time-varying frequency and random
amplitude. Specifically, y (t ) = s (t ) + n(t ) for 0ms ≤ t ≤ 8ms ,
where n(t ) ≈ WN (0,0.01) and
s(t ) = A(t ) cos( f (t ) × 2πt ) + n(t ) where A(t) is a random
process, f(t) is 10(kHz) for 0ms ≤ t ≤ 2ms , f(t) is 13(kHz) for
0.6ms ≤ t ≤ 0.8ms and for 0.2ms ≤ t ≤ 0.6ms f(t) increases
linearly from 10(kHz) to 13(kHz). The measured signal y(t) is
shown in Fig. 5.

V. CONCLUSION
In this work, by adding one additional state variable (that
represents the variations of signal amplitude) to third-order
Figure 5. Measured signal y(t).
state space model proposed in [5]-[8], we extremely improved
the performance of EKF frequency tracker for nonstationary

596
harmonic signals (harmonic signals with time-varying
amplitude and frequency). Since most of harmonic signals
typically have time varying characteristics (such as time
varying amplitude), the proposed technique is more useful and
superior compared to the conventional methods.

REFERENCES
[1] D. L. Snyder, The State-Variable Approach to Continuous Estimation
with Applications to Analog Communications Theory. Boston, MA:
MIT Press, 1969.
[2] B. Boashash, “Estimating and interpreting the instantaneous frequency
of a signal,” in Proc. IEEE, vol. 80, no. 4, pp. 520–568, 1992.
[3] L. Hsu, R. Ortega, and G. Damm, “A globally convergent frequency
estimator,” IEEE Trans. Automat. Contr., vol. 44, pp. 698–713, 1999.
[4] S. M. Kay, Modern Spectral Estimation: Theory and Applications.
Engleweood Cliffs, NJ: Prentice-Hall, 1988.
[5] B. La Scala and R. Bitmead, “Design of an extended Kalman filter
frequency tracker,” IEEE Trans. Signal Processing, vol. 44, pp. 739–
742, 1996.
[6] B. La Scala, R. Bitmead, and B. G. Quinn, “An extended Kalman filter
frequency tracker for high-noise environments,” IEEE Trans. Signal
Processing, vol. 44, pp. 431–434, 1996.
[7] B. La Scala, R. Bitmead, and M. R. James, “Conditions for stability of
the extended Kalman filter and their application to the frequency
tracking problem,” Math. Control, Signals Syst., vol. 8, pp. 1–26, 1995.
[8] S. Bittanti and S. M. Savaresi, “On the Parameterization and Design of
an Extended Kalman Filter Frequency Tracker,” IEEE Trans.
Automatic Control, vol. 45, pp. 1718–1724, 2000.
[9] L. D. Avendano-Valencia, L. E. Avendano, J. M. Ferrero, and G. C.
Dominguez, “Improvement of an Extended Kalman Filter Power Line
Interference Suppressor for ECG Signals,” Computers in Cardiology,
ISSN 0276−6574, 34:553−556, 2007.
[10] B. D. O. Anderson and J. B. Moore, Optimal Filtering. Englewood
Cliffs, NJ: Prentice-Hall, 1979.
[11] M. I. Ribeiro, “Kalman and Extended Kalman Filters: Concept,
Derivation and Properties,” Institute for Systems and Robotics,
Instituto Superior Tcnico, Av. Rovisco Pais, 1, February 2004.
[12] S. Bittanti, A. J. Laub, and J. C. Willems, Eds., The Riccati equation,
New York: Springer-Verlag, 1991.

597