You are on page 1of 3

UNBIASED FREQUENCY ESTIMATION

Yolanda Vidal
LaCàN, Departament de Matemàtica Aplicada III, Escola Universitària d’Enginyeria Tècnica Industrial de Barcelona,
Universitat Politècnica de Catalunya, Comte d’Urgell, 187, 08036 Barcelona, Spain.
yolanda.vidal@upc.edu

Leonardo Acho, Francesc Pozo


CoDAlab, Departament de Matemàtica Aplicada III, Escola Universitària d’Enginyeria Tècnica Industrial de Barcelona,
Universitat Politècnica de Catalunya, Comte d’Urgell, 187, 08036 Barcelona, Spain.
leonardo.acho@upc.edu, francesc.pozo@upc.edu

Keywords: frequency estimation; observer; adaptive system.

Abstract: This paper presents an alternative solution to the problem of frequency estimation. An analysis of the stability
and a dynamic estimation are obtained for a dynamic system using the Lyapunov theory. The robustness of
the proposed method is tested numerically.

In this paper a new option is proposed that employs


sinusoidal frequency
input estimation
Lyapunov theory to find the stability of the system.
frequency
estimator We assume that the sinusoidal signal is unbiased; a
study on the biased case is given in (Bobtsov, 2008).
We present numerical experiments with a corrupted
sinusoidal signal to illustrate the robustness of our
Figure 1: Block diagram of a frequency estimator. system.
The contents of this paper are as follows. Sec-
tion two presents a dynamic system for the estimator
design; Lyapunov theory is employed to develop the
1 INTRODUCTION design. Section three presents the numerical experi-
ments and robustness analysis. Conclusions on this
On-line frequency estimation has been ap- work are drawn in Section four.
proached using various techniques (see (Bobtsov,
2008; Xia, 2002; Hsu et al., 1999) and the references
therein). Frequency estimation is an important topic 2 FREQUENCY ESTIMATION
and has many applications in engineering, such as
vibration control and the design of disk drivers and
DESIGN
magnetic bearings (Hsu et al., 1999). These appli-
Consider the problem of on-line frequency esti-
cations are in the time domain, which is often sim-
mation for a pure sinusoidal signal (for instance, see
pler than spectral analysis. Spectral analysis requires
(Hsu et al., 1999))
complex numbers and many numeric calculations; the
data for spectral analysis is contained in the time- s(t) = a sin (ωt) (1)
domain. Using the time domain data, frequency esti- where ω is the frequency and a is the amplitude.
mation consists of modeling the dynamic system with These parameters are assumed to be unknown positive
a sinusoidal input signal with an unknown amplitude numbers. The objective is to find a dynamic system
and frequency; then, estimations are made on the fre- that gives an estimation of w. Consider the following
quency of the system (see Figure 1). A globally con- system,
vergent adaptive notch filter is studied in (Hsu et al.,
ẏ = ṡ − θ(y − x2), (2)
1999), and an alternative method, employing an adap-
tive observer design, is presented in (Bobtsov, 2008). x˙2 = s − y. (3)
Suppose that there exists θ = θ∗ such that y = 0,
s p w ẏ = w − θ(y − x2 ) y, x2 θ
which implies that ẏ = 0. Then, the system (2)-(3) θ̇ = ρy(y − x2)
reduces to βp + 1 x˙2 = s − y

θ∗ x2 = −ṡ, (4)
x˙2 = s. (5)
So, θ∗ x˙2 = −s̈, which is equivalent to Figure 2: Block diagram of the proposed frequency estima-
tor.
θ∗ s = −s̈, (6)
corresponding to a linear oscillator with the oscilla-
tion frequency: θ∗ = w2 . Now, the objective is to find (ii) x2 is bounded. As y converges to zero then by (2)
a dynamic system for θ such that a combination of y we have that x2 converges to − θṡ∗ as t → ∞. So x2
and θ gives a stable equilibrium point (in the sense of converges to a periodic motion. This is verified in
Lyapunov theory). To fulfill this objective, consider the numerical simulations.
the following Lyapunov function: The main difficulty with system (2)-(3) is the time
1 2 1 1
 

2 derivative of s: differentiation is undesirable due to
V = y + (θ − θ ) + ∗ 2
x2 − − ∗ . (7) noise amplification. However, to avoid this diffi-
2 2ρ 2 θ culty, we can use the following stable approximation
Here, ρ is a positive constant. The last term in (7) (Spong and Vidyasagar, 1989):
captures the convergence of x2 to a periodic motion, p
as expected. Then, the derivative of V along the tra- w= s∼= ṡ
βp + 1
jectory of system (2)-(3) is,
1
 ∗
θ x2 + ṡ
 

 where p is the Laplace operator and β is a small pos-
V̇ = yẏ + (θ − θ )θ̇ +

ẋ2 − − ∗ itive number. Therefore, our frequency estimator is
ρ θ∗ θ
,
given by:
and using the equation (6), yields
ẏ = w − θ(y − x2), (10)
1
V̇ = yṡ − θy(y − x2) + (θ − θ∗ )θ̇. x˙2 = s − y, (11)
ρ
θ̇ = ρy(y − x2), (12)
Defining θ̃ = θ − θ∗ , and using (4), p
w= s. (13)
1
V̇ = −yθ∗ x2 − θy(y − x2) + θ̃θ̇ βp + 1
ρ
A block diagram of the frequency estimator (10)-(13)
1
= θ̃yx2 − θy2 + θ̃θ̇ is given in Figure 2.
ρ Remark: In practical computations the derivative
1 approximation (13) can be rewritten as a differential
= θ̃yx2 − θ̃y2 + θ̃y2 − θy2 + θ̃θ̇
ρ equation for the state variable w as
θ̇ βẇ + w = ṡ,
 
= θ̃ − y(y − x2) − θ∗ y2 . (8)
ρ
which finally gives
Finally, defining
1
θ̇ = ρy(y − x2), (9) ẇ = (−w + ṡ).
β
the equation (8) transforms to
V̇ = −θ∗ y2 .
3 NUMERICAL EXPERIMENTS
Thus, V̇ is negative semi-definite. Also, the equations
above imply that y, θ, ẏ, θ̇ ∈ L∞ and y ∈ L2 . Invoking In this section we present numerical experiments
the Barbalat’s lemma we can ensure that V̇ → 0 as for the system (10)-(13). The simulation results are
t → ∞. Therefore, we can conclude that y converges shown in Figure 3 for the case where
to zero as time passes. Note that this implies that:
s(t) = sin (2t), ρ = 10, and β = 0.001.
(i) θ is an estimation of θ∗ . The convergence of y to
zero after a period of time implies, using (9), that As expected, the frequency estimator gives θ∗ = ω2 ≈
θ̇ converges to 0 as time passes. So θ converges to 4 and the variable x2 is periodic. To test the robust-
a constant value θ∗ (as y converges to zero). ness of the estimator we use the following corrupted
1 1

0.5 0.5
y(t)

y(t)
0 0

−0.5 −0.5
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
time time
0.5
0.4
0.2
x (t)

x (t)
0 0
2

2
−0.2
−0.4
−0.5
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
time time
5 5
4 4
3 3
θ(t)

θ(t)
2 2
1 1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
time time
2
4
1
2
w(t)

w(t)

0
0
−1
−2
−2
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
time time

Figure 3: Numerical experiments for the unperturbed case. Figure 4: Numerical experiments for the perturbed case.

2
signal (see Figure 5), which is based on the signal in 1
(Bobtsov, 2008) without the un-biasing term:
s(t)

s(t) = sin (2t) + 0.1 sin (10t) −1


+ 0.05 sin(20t) + 0.03 sin(30t). −2
0 2 4 6 8 10 12 14 16 18 20
time
The simulation results for the perturbed signal are
shown in Figure 4. For the perturbed signal the Figure 5: Perturbed signal.
frequency estimation approaches the expected value
θ∗ = ω2 ≈ 4 and the variable x2 remains periodic.
REFERENCES
4 CONCLUSIONS Bobtsov, A. (2008). New approach to the problem of glob-
ally convergent frequency estimator. Int. J. of Adap-
tive Control and Signal Processing, 22:306–317.
An alternative approach to on-line frequency estima-
tion has been proposed. This approach was obtained Hsu, L., Ortega, R., and Damm, G. (1999). A globally con-
vergent frequency estimator. IEEE Tran. on Automatic
by examining the problem from a novel viewpoint and Control, 44(4):698–713.
invoking Lyapunov theory. We examine the case of an
unbiased signal input. Further research is required to Spong, M. and Vidyasagar, M. (1989). Robot dynamics and
Control. John Willey & Sons, Inc., USA.
solve the problem of a biased signal input.
Xia, X. (2002). Global frequency estimation using adap-
tive identifiers. IEEE Tran. on Automatic Control,
47(7):1188–1193.

You might also like