Sunteți pe pagina 1din 5

Signal Processing 84 (2004) 1423 1427

www.elsevier.com/locate/sigpro

Algorithms for estimating instantaneous frequency


Jaideva C. Goswami , Albert E. Hoefel
Schlumberger Technology Corporation, 110 Schlumberger Drive, Sugar Land, TX 77478, USA
Received 28 March 2003; received in revised form 22 April 2004

Abstract
Three algorithms to estimate instantaneous frequency of a frequency-modulated signal is discussed. These algorithms are
based on Hilbert transform, Haar wavelet, and generalized pencil of function (GPOF) methods. While GPOF-based frequency
detection method appears to be least sensitive to noise, wavelet-based method is easiest to implement. The latter method is
also computationally more e5cient and can be implemented in real-time. Results for both synthetic and experimental data
are shown.
? 2004 Elsevier B.V. All rights reserved.
Keywords: Frequency modulation; Phase-shift keying; Frequency-shift keying; Wavelet; Hilbert transform; Matrix pencil

1. Introduction
The problem of estimating instantaneous frequency
of a received signal is very important in many communication areas. In this paper, we discuss this problem in the context of wireless data acquisition in
oil;eld industry. In some oil-;eld exploration
applications, a small amount of measured data are
transferred between a measurement sensor and a
host unit where, both the sensor and host units are
located at a few kilometers underground. Binary
frequency-shift keying (BFSK) scheme is often used
for data communication. The major problem for such
data communication comes because of highly limited availability of space and power. Demodulation
algorithms should be easily implementable in hardware and ;rmware. With this application in mind,

Corresponding author.
E-mail addresses: jcgoswami@ieee.org (J.C. Goswami),
ahoefel@ieee.org (A.E. Hoefel).

we have studied three di@erent methods for estimating


instantaneous frequencies. These are based on Hilbert
transform [6], Haar wavelet [1], and generalized pencil of function (GPOF) methods [2,3,5]. In this section
we brieEy review frequency modulation scheme. Sections 24 describe algorithms for estimating instantaneous frequency by Hilbert transform, wavelet transform, and GPOF methods, respectively. And ;nally,
in Section 5, we discuss numerical and experimental
results.
In frequency modulation (FM), the instantaneous
frequency !i of the carrier varies linearly with the
modulating signal m(t) [4]. Thus !i can be written as
!i = !c + kf m(t);

(1)

where !c is the carrier frequency and kf is a constant.


The phase, (t), is given by
 t
(t) =
[!c + kf m()] d
(2)

 t
= !c t + kf
m() d:
(3)

0165-1684/$ - see front matter ? 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.sigpro.2004.05.016

1424

J.C. Goswami, A.E. Hoefel / Signal Processing 84 (2004) 1423 1427

Modulating
Signal

Phase
Modulator

Integrator

Modulating
Signal
Differentiator

Frequency
Modulator

FM
Signal

|s^()|

PM
Signal
c

Fig. 1. Relationship between phase and frequency modulation.

An FM signal, therefore, can be represented as




 t
s(t) = A cos !c t + kf
m() d :

(4)

(5)

and instantaneous frequency !i varies linearly with


the derivative of the modulating signal
d
!i =
= !c + kp m(t):
dt

Fig. 2. A typical magnitude spectrum of a signal.

by multiplying its spectrum s(!)

with a unit step function as

It is worth mentioning here that in a phase-modulated


(PM) signal, instead of frequency, the phase (t)
varies linearly with the modulating signal, namely
(t) = !c t + kp m(t)

u(!);

s+ (!) = s(!)

where u(!)

is the unit step function, de;ned in the


usual way as

1 !0
u(!)

=
(8)
0 otherwise:
From (7) we have

+ sgn(!)]
2s+ (!) = s(!)[1

(6)

It is, therefore, clear that PM and FM are intimately


related. One can be obtained from the other as shown
in Fig. 1. The commonly used digital modulation techniques, BPSK (binary phase-shift keying) and BFSK
are special cases of PM and FM, respectively, when
the modulating signal m(t) takes binary values. Although all the examples discussed in this paper pertain
to FM, the algorithm, with trivial modi;cation, can be
used for PM as well.
2. Hilbert transform algorithm
The Hilbert transform of a signal s(t) produces a
signal sh (t) that is orthogonal to s(t). The angle of the
complex signal s(t) + jsh (t) then gives the instantaneous phase, and its derivative, the instantaneous frequency of the signal.
Let s(!)

represent the Fourier transform of a


real-valued signal s(t). A typical magnitude spectrum
of s(t) is shown in Fig. 2. We can construct a signal
s+ (t) that contains only positive frequencies of s(t)

(7)

= s(!)

+ j[ j sgn(!)s(!)]

;




(9)

sh (!)

where sgn(!) is the signum function de;ned as



1
!0
sgn(!) =
(10)
1 ! 0:
In (9), sh (t) is the Hilbert transform of s(t), de;ned as

sh (t) = F1 {j sgn(!)s(!)}

s()
1
d;
=
 t 

(11)

where F1 represents inverse Fourier transform. It is


easy to verify that s(t); sh (t) = 0 sh (t) s(t).
Once we have the orthogonal signal sh (t) of s(t),
the instantaneous phase and frequency of s(t) can be
obtained by [6]


sh (t)
(12)
(t) = arctan
s(t)
and
!i (t) =

d
= !c + kf m(t):
dt

(13)

J.C. Goswami, A.E. Hoefel / Signal Processing 84 (2004) 1423 1427

1425

4. GPOF algorithm
Given a set of discrete data {fi : i = 0; 1; : : : ; N }
of a complex-valued function f(t), generalized pencil
of function method ;nds a set of complex coe5cients
{ck ; k : k = 1; 2; : : : ; M } such that

0.5
0

fi := f(ti ) =
(a)

Fig. 3. Haar (a) scaling function, and (b) wavelet.

3. Wavelet algorithm
The principle behind wavelet-based instantaneous
frequency estimation is the same as Hilbert transform
method, namely ;nding orthogonal signal of a given
frequency-modulated signal. From wavelet theory [1],
we can decompose a given function s(t) into orthogonal signals f(t) and g(t) such that

f(t) =
ck (at k);
(14)
k

g(t) =

dk (at k);

ck exp(k ti );

M N;

(19)

k=1

(b)

(15)

where parameter a depends on the sampling rate of


s(t). The bases (t) and (t) are called scaling function and wavelet, respectively. These two functions
are orthogonal to each other, i.e. (at); (at ) =
0; Z := {: : : ; 1; 0; 1; : : :}.
We will use the simplest scaling function and
wavelet due to Haar, as shown in Fig. 3. For this case
the coe5cients {ck } and {dk } with a = 2 are given as
2ck = s(k) + s(k + 1);

(16)

2dk = s(k) s(k + 1):

(17)

As with Hilbert transform case, once we have orthogonal signals, we can obtain the instantaneous phase
k := (tk ); tk = kMt

dk
;
(18)
k = A arctan
ck
where A is a constant depending upon the sampling
rate. Finite di@erence of {k } then gives the instantaneous frequency.

where ti = iMt and Mt is the discretization step. The


GPOF method has better numerical performance than
the conventional Pronys method [7] which involves
solving an ill-conditioned matrix equation, and ;nding roots of a polynomial. The GPOF method also
has better noise sensitivity [3]. A detailed discussion
on the subject may be found in [3]. The algorithm is
brieEy described below.
Consider vectors of discrete data {fi },
Fi := [fi fi+1 : : : fi+N L1 ]T ;

0 6 i 6 L;

(20)

where T stands for transpose, and form matrices of


size (N L) L,
A1 = [F0 F1 : : : FL1 ]

(21)

and
A2 = [F1 F2 : : : FL ]:

(22)

Then, with zk := exp(k Mt), it can be shown that {zk }


are the generalized eigen values of matrix pencil A2
zA1 . Once we have all the exponents, the coe5cients
{ck } can be easily computed from (19) since in {ck },
it is a linear equation.
In the present problem of instantaneous frequency
estimation, we choose a window size that covers at
least one cycle of the highest carrier frequency and
then extract two exponents. The imaginary parts of
these exponents k give instantaneous frequency. By
sliding the window, we can compute the instantaneous
frequency for the entire signal.
5. Results and discussions
As a ;rst step, we test and compare all three
algorithms discussed in this paper by applying them
to synthetic frequency-modulated data. The comparison is based on relative signal reconstruction error, ,

J.C. Goswami, A.E. Hoefel / Signal Processing 84 (2004) 1423 1427

10
20
30
40
50
60
70
80
90
100

Error 
Hilbert transform

Haar wavelet

GPOF

0.00004
0.00019
0.00042
0.00076
0.00123
0.02897
0.03840
0.04707
0.05133
0.05656

0.00024
0.00109
0.00273
0.00523
0.00850
0.02143
0.03500
0.06756
0.10211
0.11747

0.00004
0.00005
0.00010
0.00016
0.00022
0.00030
0.00042
0.00062
0.00099
0.01252

Frequency modulated signal


3

Exact

Computed)

30
20
10

Hilbert transform
0

0.5

1.5

2.5

Time (seconds)
40
30
20
10

Haar wavelet
0

0.5

1.5

2.5

Time (seconds)
40

Frequency (Hz)

Noise

Frequency estimation (
40

Frequency (Hz)

Table 1
Error () in frequency estimation by three algorithmsHilbert
transform, Haar wavelet, and generalized pencil of function
(GPOF)for di@erence noise level (as a percentage of signal
strength)

Frequency (Hz)

1426

30
20
10

GPOF
0

0.5

1.5

2.5

Time (seconds)

Fig. 5. Frequency demodulation of a BFSK signal shown in Fig.


4 by using three algorithms based on Hilbert transform, Haar
wavelet, and GPOF.

Frequency detection using Hilbert transform

1.2

Experimental modulating data


Computed normalized frequency

0.8
3

0.5

1.5

2.5

Time (seconds)

Fig. 4. A BFSK
(SNR = 3:0 dB).

signal

with

additive

0.6

Gaussian

noise

0.4
0.2

de;ned as

|Si; n Si; 0 |2
 = i
;
2
i |Si; 0 |

(23)

where summation is over all samples, and Si; n , and Si; 0


represent ith sample of reconstructed signal with and
without noise in the modulated data, respectively. Table 1, shows the results for error with di@erent levels
of additive Gaussian noise as a percentage of signal
strength. To visually see the e@ect of noise on frequency demodulation, we have plotted the signal in

0.2

0.5

1.5

2
2.5
3
Time (milli-seconds)

3.5

4.5

Fig. 6. Frequency demodulation of experimental BFSK data using


Hilbert transform.

Fig. 4 and the frequency demodulated signals in Fig. 5


for all three algorithms. Finally, we apply these algorithms to experimental data obtained by using a BFSK

J.C. Goswami, A.E. Hoefel / Signal Processing 84 (2004) 1423 1427

1427

Frequency detection using Haar wavelet


1.2
Experimental modulating data
Computed normalized frequency

1
0.8
0.6
0.4
0.2
0
0.2

0.5

1.5

2
2.5
3
Time (milli-seconds)

3.5

4.5

Fig. 7. Frequency demodulation of experimental BFSK data using Haar wavelet.

for real-time application, since it requires signal information only in the vicinity of the current time location. In Hilbert transform method, on the other hand,
we need the entire signal before estimating the instantaneous frequency because of the global nature of
Fourier bases. The GPOF method can be implemented
in real-time, however, matrix operations are di5cult
to be realized in hardware.

Frequency detection using GPOF

1.2

Experimental modulating data


Computed normalized frequency

1
0.8
0.6
0.4
0.2

References

0
0.2

0.5

1.5

2
2.5
3
3.5
Time (milli-seconds)

4.5

Fig. 8. Frequency demodulation of experimental BFSK data using


GPOF.

transmitter and a pick-up coil. The results are shown


in Figs. 68. For experimental data, all three methods
perform well because of low noise environment.
The results indicate that GPOF-based method is
least sensitive to noise. This is primarily due to the
fact that in GPOF, the frequency is determined by
a least squared algorithm which essentially minimizes the error. Wavelet-based method is the simplest to implement and fastest to compute. Furthermore, wavelet-based method can be implemented

[1] J.C. Goswami, A.K. Chan, Fundamentals of Wavelets: Theory,


Algorithms, and Applications, Wiley, New York, 1999.
[2] J.C. Goswami, R. Mittra, On the solution of a class of
large-body scattering problems via the extrapolation of FDTD
solutions, J. Electromagn. Waves Appl. 12 (1998) 229244.
[3] Y. Hua, T.K. Sarkar, Generalized pencil-of-function method
for extracting poles of an EM system for its transient response,
IEEE Trans. Antennas Propagat. 37 (1989) 229234.
[4] B.P. Lathi, Modern Digital and Analog Communication
Systems, Holt, Rinehart & Wilson, New York, 1983
(Chapter 4).
[5] A.J. Mackay, A. McCowen, An improved pencil-of-function
method and comparisons with traditional methods of pole
extraction, IEEE Trans. Antennas Propagat. 35 (1987)
435441.
[6] J.G. Proakis, Digital Communication, McGraw-Hill, New
York, 1995, pp. 152157.
[7] M.L. Van Blaricum, R. Mittra, A technique for extracting
the poles and residues of a system directly from its transient
response, IEEE Trans. Antennas Propagat. 23 (1975) 777781.

S-ar putea să vă placă și