Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

INTERACTION MEASURE OF AM-FM SIGNALS BY CROSS-ΨB -ENERGY OPERATOR

Abdel-Ouahab BOUDRAA∗ , Jean-Christophe CEXUS∗ , Karim ABED-MERAIM+ and Zineb SAIDI∗



IRENav, Ecole Navale, Lanvéoc Poulmic, BP600, 29240 Brest−Armées, France.
+
Telecom-Paris, Signal and Image Processing Department, 46 Rue Barrault, Paris 13, France.
{boudra,cexus,saidi}@ecole.navale.fr, abed@tsi.enst.fr

ABSTRACT x(t) as follows:


¶2
In this paper a new nonlinear similarity measure, called cross- d2 x(t)
µ
dx(t)
ΨB -energy operator, to quantify the interaction between AM- ΨR (x(t)) = − x(t) (1)
dt dt2
FM signals is introduced. We show that the time delay (TD)
estimation between two sensor measurements is an exam- or using the dot notation to denote differentiation with re-
ple of interaction measure by ΨB . Results on synthetic and spect to the independent variable, in this case time t
real signals are presented to illustrate the effectiveness of ΨR (x) = ẋ2 − xẍ (2)
the proposed method. ΨB is compared to cross-correction
2 2
for TD estimation. where x ≡ x(t), ẋ = dx(t)/dt and ẍ = d x(t)/dt . This
energy operator characterizes the very local property of the
1. INTRODUCTION signal and depends only on the signal and its first and second
Many signals possess similarities that can be exploited in time derivatives [4]. TKEO is an energy tracking method
signal processing methods such as time delay (TD) esti- that has low computational complexity, a very good time
mation between sensors [1]. The cross-correlation (CC) resolution and is very easy to implement efficiently. How-
is the simplest similarity measure used in signal process- ever, it is limited to monocomponent signal (bandpass sig-
ing although it is not sensitive to nonlinearity of the sig- nal). This operator has been successfully used in speech
nals or the sensors. Indeed, the CC is a linear similarity analysis [6],[7], signal processing [3] and image processing
measure. In practice, the interaction between signals may [8],[9].
be nonlinear. Furthermore, sensors characteristics may also To represent the interaction between two real time func-
cause nonlinear distortions. Consequently, the maximum of tions CTKEO has been defined [4]. This function may be
CC does not necessarily, corresponds to maximum of in- viewed as cross-energy between two real signals. CTKEO
teraction. In this case, the resulting TD may be erroneous. has been extended to complex-valued signals in [2]. The
To tackle this problem we present a similarity method that complex version of CTKEO [2], a symmetric energy op-
measures the nonlinear interaction between signals. This erator, is called cross-ΨB -energy operator. The CTKEO is
method is based on cross-ΨB -energy operator which is non- an energy like-function that measures the interactions be-
linear measure, recently proposed in [2]. Furthermore this tween two real signals x and y [5]. The CTKEO is given by
operator is well suited for nonstationary signals [3]. ΨB is [x, ẏ] the Lie bracket which measures the instantaneous dif-
derived from an energy-like function, called cross Teager- ferences in the relative rate of change between x and ẏ. In
Kaiser energy operator (CTKEO) [4]-[5], which measures the general case, if x and y represent displacements in some
the interactions between two real functions. ΨB may also generalized motions, [x, ẏ] has dimensions of energy (per
measure how much a signal is present in another one. Thus, unit mass), it is viewed as a cross-energy between x and y
the operator may be viewed as a nonlinear matched filter. [5]. More precisely, let x and y be two complex signals, ΨB
is defined as follows [2]:
1
2. TEAGER-KAISER ENERGY OPERATOR ΨB (x, y) = [ΨC (x, y) + ΨC (y, x)]
2
Teager-Kaiser energy operator (TKEO), a well known de- 1 1
modulator in signal and image processing, is usually defined = [x˙∗ ẏ + ẋy˙∗ ] − [xy¨∗+x∗ ÿ+y x¨∗+y ∗ ẍ] (3)
2 4
as a local energy measure for oscillating (simple harmonic)
where ΨC (x, y) is given by [2]:
signals. TKEO computes the energy of a real-valued signal
1 ˙∗ 1
ΨC (x, y) = [x ẏ + ẋy˙∗ ] − [xy¨∗ + x∗ ÿ] (4)
2 2

0-7803-9243-4/05/$20.00 ©2005 IEEE 775


ΨB (x, y) is symmetric bilinear form and ΨB (x, x) is the as- ΨB (x, y) = (ȧḃ + abφ̇1 φ̇2 ) cos(φ21 )
sociated quadratic form [2]. We show that the ΨB (x, y) of 1
complex signals x and y is equal to the sum of ΨB (x, y) of − (ȧbφ̇2 − aḃφ̇1 ) sin(φ21 ) − a(−bφ̇22 + b̈) cos(φ21 )
2
their real and imaginary parts [2]. 1 1
+ a(2φ̇2 ḃ + bφ̈2 ) sin(φ21 ) − b(−aφ̇21 + ä) cos(φ21 )
ΨB (x, y) = ΨB (xr , yr ) + ΨB (xi , yi ) (5) 2 2
1
− b(−2φ̇1 ȧ + aφ̈1 ) sin(φ21 ) (13)
where x(t) = xr (t) + jxi (t) and y(t) = yr (t) + jyi (t). 2
According to (5) the function ΨB (x, y) is a real quantity, as By factorizing (13), (6) follows immediately.
expected for an energy operator.
Proposition 2 Let x(t) = a(t)ejω1 t and y(t) = b(t)ejω2 t
3. ΨB -BASED SIMILARITY MEASURE be two AM signals. The maximum of energy of interaction
between these two signals is given by:
We provide here some properties of the ΨB measure and
show that one of them resolve the TD estimation problem. ab ab̈ + äb
Note, first, that | ΨB (x, y) | is a similarity measure. Indeed, | ΨB (x, y) |≤ Ω2 (| aḃ + ȧb | + | |)+ | ȧḃ − | (14)
2 2
| ΨB (x, y) |≥ 0 and according to equation (3) | ΨB (x, y) |=|
ΨB (y, x) |. where ω1 /2π and ω2 /2π are carrier frequencies of x and y
respectively and Ω = ω1 + ω2 .
Proposition 1 Let x(t) = a(t)ejφ1 (t) and y(t) = b(t)ejφ2 (t)
be two complex AM-FM signals. The energy of interaction Proof Equation (6) is reduced to
between these two signals is given by: · ¸
ab 2 ab̈ + äb
ΨB (x, y) = Ω + ȧḃ − cos ω21 t (15)
· ¸
ab 2 ab̈ + äb 2 2
ΨB (x, y) = φ̇ + (ȧḃ − ) cos(φ21 ) +
2 2 · ¸
+ (aḃ − ȧb)Ω2 sin ω21 t (16)
· ¸
ab
(φ̈21 ) + (aḃ − ȧb)φ̇2 sin(φ21 ) (6)
2
where ω21 = ω2 − ω1 . Taking the absolute value of both
where a ≡ a(t), b ≡ b(t), φ ≡ φ(t), φ1 ≡ φ1 (t), φ2 ≡ sides of (15), equation (14) holds.
φ2 (t), φ̇ = φ̇1 + φ̇2 , φ21 = φ2 − φ1 and φ̈21 = φ̈2 − φ̈1 .
Proposition 3 Let x(t) = Aejφ1 (t) and y(t) = Bejφ2 (t) be
Proof two complex FM signals. The maximum of ΨB is given by
ẋ(t) = ejφ1 [ȧ + jaφ̇1 ] (7) | AB |
jφ1 2 | ΨB (x, y) |≤ .((φ̇1 + φ̇2 )2 + | (φ̈1 − φ̈2 ) |) (17)
ẍ(t) = e [2j φ̇1 ȧ − aφ̇1 + ä + jaφ̈1 ] (8) 2
ẋ∗ (t)ẏ(t) = ejφ21 [ȧḃ + j ȧbφ̇2 − jaḃφ̇1 + abφ̇1 φ̇2 ] (9) Proof Since x and y are FM signals then A and B are con-
x(t)ÿ ∗ (t) = e−jφ21 [−2j ḃφ̇2 − bφ̇22 + b̈ − jbφ̈2 ]×a (10) stant values and equation (6) is reduced to
ẍ∗ (t)y(t) = ejφ21 [−2j ȧφ̇1 − aφ̇21 + ä − jaφ̈1 ]×b (11) · ¸
AB
ΨB (x, y)= (φ̇1+φ̇2 )2 cos(φ21 )+(φ̈1 − φ̈2 ) sin(φ21 ) (18)
In similar fashion we calculate ẏ(t), ÿ(t) and ẋ(t)ẏ ∗ (t). 2
Thus, substituting equations (7)-(11) in expression (3) leads
to: Taking the absolute value of both sides of (18), equation
1 (17) holds.
ΨB (x, y) = [(ȧḃ + abφ̇1 φ̇2 )(ejφ21 + e−jφ21 )]
2
1 4. ΨB DETECTION FRAMEWORK
+ [j(ȧbφ̇2 − aḃφ̇1 )(ejφ21 − e−jφ21 )]
2 Consider the two sensors TD estimation problem. These
1
− [a(−bφ̇22 + b̈)(ejφ21 + e−jφ21 )] sensors capture delayed and noisy versions of the same sig-
4 nal, s(t). Sensor measurements are given by:
1
− [a(2φ̇2 ḃ + bφ̈2 )(ejφ21 − e−jφ21 )] (
4 x(t) = s(t) + n1 (t)
1 (19)
− [b(−aφ̇21 + ä)(ejφ21 + e−jφ21 )] y(t) = α.As(t − τ ) + n2 (t)
4
1
− [b(−2φ̇1 ȧ − aφ̈1 )(ejφ21 − e−jφ21 )] (12) where s(t) = Aejωt is a complex FM signal. α is an at-
4 tenuation coefficient and τ denotes the TD between the two
Finally, sensor measurements. The processes {n1 (t)} and {n2 (t)}

776
are supposed zero-mean and independent from the signal where ΓXY (t1 , t2 ) = E[X(t1 ).Y ∗ (t2 )] is the cross corre-
s(t). Relation (19) can be written as follows: lation between two random processes X(t) and Y (t). Us-
( ing the same reasoning, as above, it is straightforward to
x(t) = AejΦ1 (t) + n1 (t) show that E[u̇.ṅ2 ] = 0, E[u̇∗ .ṅ2 ] = 0, E[ṅ1 .v̇] = 0 and
(20)
y(t) = α.AejΦ2 (t) + n2 (t) E[ṅ1 .ṅ2 ] = 0. Finally, E[ΨB (x, y)] = AαE[ΨB (u, v)] and
equation (21) holds.
where Φ1 (t) = ωt and Φ2 (t) = Φ1 (t − τ )
αA2
· ¸
Proposition 4 Suppose that n1 (t) and n2 (t), and x(t) and ΨB (x, y)= (φ̇1+φ̇2 )2 cos(φ21)+(φ̈1−φ̈2) sin(φ21) (32)
2
y(t) are mutually uncorrelated and mean square differen-
tiable. If n1 (t) and n2 (t) are zero-mean then, φ̇2 (t) = φ̇1 (t) = ω ⇒ φ̈2 (t) = φ̈1 (t) = 0. Finally, using
jΦ1 (t) jΦ2 (t) (32), equation (22) follows immediately. The argument τ
E[ΨB (x, y)] = E[ΨB (Ae , α.Ae )] (21)
that maximizes equation (22) provides an estimate of TD.
2 2
E[ΨB (x, y)] = 2αω A E[cos(φ21 )] (22) Because of finite observation time E[(cos(φ21 )] can only
be estimated. Thus,
Proof n1 (t) and n2 (t) have zero-mean and are uncorre-
lated. Then, E[n1 .n2 ] = 0 ⇒ E[ṅ1 .ṅ2 ] = −0.5E[n1 .n̈2 ]− 2α(ωA)2
Z T
0.5E[n2 .n̈1 ]. Similarly, E[u̇∗ .ṅ2 ] = −0.5, E[u∗ .n̈2 ]−0.5, E[ΨB (x, y)] ' cos(φ21 )dt = εB (τ ) (33)
(T − τ ) τ
E[n2 .ü∗ ]. The same result is obtained for E[v̇ ∗ .ṅ1 ], E[u̇.ṅ2 ]
and E[v̇.ṅ1 ]. Using these relations it is easy to see that: A simple estimate of the lag at which ΨB peaks is:
Aα ∗ Aα τ̃ = arg max [εB (τ )]
ΨB (x, y)= {u̇ v̇ + u̇v̇ ∗ }− {uv̈ ∗ + u∗ v̈ + ü∗ v + üv ∗ } (34)
2 4 τ ∈[0,T ]

+{A.u̇∗ ṅ2+α.ṅ1 v̇+2.ṅ1 ṅ2+A.u̇ṅ2+αṅ1 v̇ ∗ } (23)


Thus, the problem of TD estimation is a measure of interac-
where u ≡ e jΦ1 (t)
and v ≡ Ae jΦ2 (t)
. Taking the expecta- tion between two FM signals.
tion of (23) yields:
5. RESULTS
E[ΨB (x, y)] = AαE[ΨB (u, v)] (24)
Computer simulations have been carried out to illustrate the
+AE[u̇ .ṅ2 ]+αE[ṅ1.v̇]+2E[ṅ1.ṅ2 ]+AE[u̇.ṅ2 ]+αE[ṅ1.v̇ ∗]

ΨB -based similarity measure. Figure 1 shows an example of
(25)
two linear FM signals, x(t) and y(t) with the corresponding
instantaneous frequencies (IFs). The IF of x(t) increases
E[ṅ1 .v̇ ∗ ] = linearly with time while that of y(t) decreases with time.
Similarity measure is calculated using either equation (18)
v ∗ (t + h) − v ∗ (t) n1 (t + h0 ) − n1 (t)
E[ lim . lim ] (26) or (3). Figure 2 shows the energy of each signal and the en-
h→0 h h0 →0 h0 ergy of their interaction. The maximum of interaction corre-
sponds the instant where the two IFs intercept (Fig. 1) and
1 also where the energy of x(t) (ΨB (x, x)) and that of y(t)
= lim E[v ∗ (t + h).n1 (t + h0 ) − v ∗ (t + h).n1 (t)
0
h,h →0 hh0 (ΨB (y, y)) are equal. Maximums of interaction between x
− v ∗ (t).n1 (t + h0 ) + v ∗ (t).n1 (t)] (27) and y occur at t = 121 and t = 43 for ΨB and CC respec-
tively as shown in Figure (2). This result shows that the CC
measure is insensitive to nonlinear dependency between x
1 Γn1 v (t + h0 , t + h) − Γn1 v (t + h0 , t) and y. Away from the point the IFs cross, the amplitude
= lim [
h,h →0 h0
0 h of interaction decreases because there is less similarity be-
Γn1 v (t, t + h) − Γn1 v (t, t) tween signals. As the IFs converge from the time origin
− ] (28) to the their intersection, the interaction intensity of the sig-
h
nals increases and the maximum of interaction is achieved
at the intersection. ΨB is tested on real on real acoustic sig-
1 ∂Γn1 v (t + h0 , t)/∂t ∂Γn1 v (t, t)/∂t
= lim [ − ] (29) nals and results compared to that given by the CC and by
h →0 h0
0 h0 h0 the experimental (Exp) methods. Real data are measured
∂ 2 Γn1 v (t, t) in a tank with linear array sensors where air-full cylindri-
= (30)
∂t2 cal objects are buried under the sand. The interaction mea-
2 ∗
∂ E[n1 .v ] sure is calculated using equation (34) between the output of
= = 0 ⇒ E[ṅ1 .v̇ ∗ ] = 0 (31)
∂t2 the first sensor (reference) and each output of the remaining

777
Data 1 Data 1 Data 2 Data 2
1.8
RMSEΨB −Exp RMSECC−Exp RMSEΨB −Exp RMSECC−Exp
0.0526 0.180 0.217 0.250 1.6
ΨB(Y,Y) Ψ (X,X)
B
1.4
Table 1. RMSE calculated between ΨB , Exp and CC meth-
ods for Data 1 and Data 2 1.2
Intersection frequency

Amplitude
1

sensors. TD values are listed in Table 1 with the correspond- 0.8


ing Root mean square error (RMSE) values between pair of
0.6
sensors. Exp method is used as reference method. Data 1
and Data 2 correspond to echoes arriving from one and two 0.4
ΨB(X,Y)
targets respectively. Values reported in Table 1 show that 0.2
the RMSE of the CC method is 3.42 times higher than that
0
of ΨB method for Data 1. ΨB performs slightly better that 0 50 100 150 200
Times []
the CC. This may suggest, even partially, the nonlinear rela-
tionship between the signals that the CC cannot account for. Fig. 2. Interaction measure between x(t) and y(t).
The mismatch between expected TD values and the ΨB TD
ones may be due to the error in the estimation of the bearing
angles and the celerity in the medium. 7. REFERENCES
[1] T.K. Moon, ”Similarity methods in signal processing,”
IEEE Trans. Sig. Proc., vol. 44, no. 4, pp. 827-833,
1 0.25
1996.
Normalised frequency

Signal X
0.5 0.2
[2] J.C. Cexus and A.O. Boudraa, ”On the link between
Signal X

0 0.15 cross-Wigner distribution and cross-Teager energy op-


−0.5 0.1
erator,” IEE Electronics Lett., vol. 40, no. 12, pp. 778-
780, 2004.
−1 0.05
0 50 100 150
Times []
200 0 50 100 150 200 [3] A.O. Boudraa, J.C. Cexus, F. Salzenstein and L. Guil-
Times []
Intersection frequency lon, ”IF estimation using empirical mode decomposi-
1 0.25
tion and nonlinear Teager energy operator,” Proc. IEEE
Normalised frequency

0.5 0.2
ISCCSP, pp. 45-48, 2004, Hammamet, Tunisia.
Signal Y

0 0.15 [4] J.F. Kaiser, ”Some useful properties of Teager’s energy


−0.5 0.1 operators” Proc. ICASSP, vol. 3, pp. 149-152, 1993.
Signal Y
−1 0.05 [5] P. Maragos and A. Potamianos, ”Higher order differen-
0 50 100 150 200 0 50 100 150 200
Times [] Times [] tial energy operators,” IEEE Sig. Proc. Lett., vol. 2, no.
8, pp. 152-154, 1995.
Fig. 1. Linear chirp test signals (left) and the corresponding
IFs (right). [6] H.M. Teager and S.M. Teager, ”Evidence for nonlinear
speech production mechanisms in the vocal tract,” in
Proc. NATO Advanced Study Institute on Speech Pro-
6. CONCLUSION
duction and Speech Modeling, France, pp. 214-261,
In this paper a new similarity function, called cross-ΨB - 1989.
energy operator, to measure the interaction between AM- [7] P. Maragos, J.F. Kaiser and T. Quatieri, ”Energy sepa-
FM signals is presented. Some properties of ΨB are pro- ration in signal modulations with application to speech
vided. We have shown that the TD estimation problem be- analysis”, IEEE Trans. Sig. Proc., vol. 41, pp. 3025-
tween two signals is an example of interaction measure be- 3051, 1993.
tween two FM signals by ΨB . Presented results show that
ΨB is sensitive the nonlinear dependencies between signals [8] P. Maragos and A. C. Bovik, ”Image demodulation us-
compared the classical CC method. To confirm the effec- ing multidimensional energy separation,” J. Opt. Soc.
tiveness of ΨB as a nonlinear similarity measure, the method Amer. A, vol. 12, no. 9, pp. 1867-1876, 1995.
must be evaluated with a large class of AM-FM signals and [9] F. Salzenstein, P. Montgomery, A. Benatmane and A.O.
in different experimental conditions such as high noise lev- Boudraa, ”2D discrete high order energy operators for
els, sampling rates and samples sizes. We plan also to tackle surface profiling using white light interferometry,” Sev-
the TD problem in the case where signals and noises are cor- enth ISSPA, vol. 1, pp. 601-604, 2003, Paris, France.
related.

778

You might also like