Instantaneous frequency estimation of FM signals by CB-energy operator

A.-O. Boudraa
The CB energy operator is an extension of the cross-Teager-Kaiser energy operator which is a nonlinear energy tracking operator to deal with complex signals and its usefulness for non-stationary signals analysis has been demonstrated. Two new properties of CB are established. The rst property is the link between CB and the dynamic signal which is a generalisation of the instantaneous frequency (IF). The second property obtained for frequency modulated (FM) signals is a simple way to estimate the IF. These properties conrm the interest of the CB operator to track the non-stationarity of a signal. Results of IF estimation in a noisy environment of a nonlinear FM signal are presented and comparison to the Wigner-Ville distribution and the Hilbert transform-based method is provided.

If one substitutes x(t) (4) in (2), then one obtains CB (x(t)) = |x(t)|2 (b (t) b(t))2 + (b (t) + b(t)) 2 (8)

which shows that CB can be written in terms of a dynamic signal. Furthermore, (8) can also be written as CB (x(t)) = |x(t)|2 8p2 fi2 (t) + ib2 (t) a(t) a(t) (9)

where there is a relationship between CB , the IF and the instantaneous bandwidth. Finally, links given by relations (8) and (9) show that CB conveys precious local information about the instantaneous behaviour, over the time, of a signal. For a FM signal (a(t) A), using (9) the IF is given by fi (t) = CB (x(t)) 2p A 2 (10)

Introduction: The CB-energy operator is an extension of the crossTeager-Kaiser operator which is a nonlinear tracking energy operator [1] to deal with complex signals [2]. Based on the Lie bracket, CB is well suited to measure the instantaneous interaction between two complex signals [3]. The output of CB is related to the instantaneous cross-correlation (CC) of complex signals x(t) and y(t), Rxy(t, t), as follows [2]: 2 Rxy (t, t) CB (x(t), y(t)) = t2 2 R (t, t) xy t2 t=0 (1)
t=0

where A is a constant. Relation (10) is a simple way to estimate the IF of a FM signal. This IF estimation is obtained without involving integral transform and with no a priori knowledge about the phase f(t) of the signal. An advantage of the CB operator is its localisation property. This localisation is the consequence of the differentiation based method. Results: Consider a noisy polynomial phase signal (PPS): y(t) = z(t) + n(t) = Aejf(t) + n(t)
P

Relation (1) shows that CB is a cross-energy function of two signals and thus links to transforms using the concept of instantaneous CC, such as time-frequency representations (TFRs), can be found. In fact, relationships between the CB operator and some TFRs such as cross-ambiguity function or cross-Wigner-Ville Distribution (WVD) show that CB is well suited to study non-stationary signals [2 4]. CB has found applications in both signal and image processing such as time series analysis, gene time series expression data clustering, transient detection or time delay estimation [5 8]. The CB operator is a symmetric bilinear form dened as follows [2]: CB (x, y) = 0.5[CC (x, y) + CC (y, x)] (2)

= Ae

j
k=0

ak t k

(11) + n(t), t [ [0, T ]

where the associated quadratic form CC(x, x) CB(x, x) is given by x y CC (x, y) = 0.5[ y + xy ] 0.5[x + x y] (3)

In this Letter new properties of CB are established. We show the link between CB and the dynamic signal which describes the rate of the log-magnitude and phase [9]. The second relationship is a simple way to estimate the IF of a frequency modulated (FM) signal. Thus, it is natural to use CB as the local operator, based of the signal time derivatives, as a basis for tracking instantaneous features such as the instantaneous frequency (IF) function. In the following, for x(t) y(t) the notation CB(x(t), y(t)) ; CB(x(t)) is used. Dynamic signal and CB operator: The dynamic signal is similar to the denition of complex cepstrum in the frequency domain and can be viewed as a generalisation of the IF of a signal with varying magnitude [9]. Consider an AM-FM signal x(t) = a(t)ejf(t) (4)

a(t) is instantaneous amplitude and f(t) instantaneous phase. The IF is given by fi (t) = 1 f(t) 2p (5)

The dynamic signal is dened by [9]

b(t) =

d a(t) log x(t) = + jf(t) dt a(t)

(6)

and the instantaneous bandwidth is given by ib(t) = a(t) a(t) (7)

where z(t) is a noise-free signal, n(t) is a complex white Gaussian noise, f(t) is signal phase, P is order of polynomial and ak are corresponding polynomial coefcients. T is signal duration. A fourth-order PPS, with parameters A 2 and (a0 1, a1 10, a2 10, a3 2, a4 25), is selected and Monte-Carlo simulations implemented to show the effectiveness of CB (10) as an IF estimator. Estimation is evaluated for an input signal-to-noise ratio (SNR) ranging from 26 to 40 dB with an interval of 2 dB. One hundred realisations are performed for each SNR and mean square error (MSE) between the estimated value and desired value chosen as performance criteria. Since both CB (10) and the unwrapped phase of the analytical signal (Hilbert transform (HT)) (5) involve derivatives, numerical differentiations are used. For very noisy signals, traditional Euler derivatives are useless. A better option is to use derivatives lters such as the Savitzky-Golay (SG) differentiation lter [10]. This lter uses a polynomial t across a moving window that preserves higher-order moments of the signal. Derivatives are calculated with a fourth-order SG lter and 33 point windows. In Fig. 1a we display the IF (ideal) of z(t). Figs. 1b d show the results of IF tracking using HT, WVD and CB-based methods for SNR 30 dB. Compared to HT and VWD, the best extraction is given by the CB approach. Fig. 1b shows an agreement of VWD estimate with true IF but with uctuations of high magnitude due essentially to cross-terms corruption in the presence of noise [11]. HT shows a good match but with ripples (modulations errors) (Fig. 1c) compared to CB (Fig. 1d). Fig. 2 shows MSE against input SNR for the HT, WVD and CB-based methods. This result shows the ability of CB to resolve a nonlinear FM component. As seen in Fig. 2, apart from different SNRs, the CB approach provides a signicant performance improvement over the HT and VWD approaches. Of particular interest are the results at low SNRs for which the WVD approach performs less well than CB and HT. This is expected, as pointed out in [12]. As the SNR decreases, WVD produces biased representation of signal energy as well as spurious components (artefacts). Except for higher SNRs (30 dB), HT outperforms the WVD approach mainly due to use of an integral transform which does implicit smoothing (lowpass ltering) [11]. For SNR 36 dB, CB and DWV have similar performance. It is important to mention that CB , like most local approaches or those based on signal derivatives, is sensitive to a very noisy environment. For moderately noisy data, classical Euler discretisation schemes are robust to noise. For low SNRs, the use of derivatives lters such as SG is necessary for efcient tracking of the IF by CB or HT. It was found that for different polynomial orders 3 the obtained results are similar but the

ELECTRONICS LETTERS 12th May 2011 Vol. 47 No. 10

precision of estimates both by CB and HT are dependent on the size of the SG sliding window used.
22 20
frequency, Hz

true IF

WVD

18 16 14 12 10 8

knowledge about the phase of the signal. Preliminary results show that CB is effective for estimating the IF of a nonlinear FM signal compared to HT and WVD based approaches. Further, the computational cost of CB is much lower. As future work we plan to apply CB to a wide range of synthetic and real signals to conrm the obtained results and to explore new properties of CB. # The Institution of Engineering and Technology 2011 4 March 2011 doi: 10.1049/el.2011.0586 One or more of the Figures in this Letter are available in colour online.

a
22 20
frequency, Hz

b
HT PsiBSG

A.-O. Boudraa (Ecole Navale, IRENav, BCRM Brest CC 600, Brest Cedex 9 29240, France) E-mail: boudra@ecole.navale.fr References

18 16 14 12 10 8 0.2 0.4 0.6 time, s 0.8 1 0.2 0.4 0.6 time, s 0.8 1

Fig. 1 IF estimate for fourth-order PPS

a True b WVD c HT d CB
90 80 70 60 MSE, dB 50 40 30 20 10 0 10 5 0 5 10 15 20 SNR, dB 25 30 35 40 Psi_B WVD HT

Fig. 2 MSE in polynomial IF estimation for HT, WVD, CB operator

Conclusion: In this Letter, two new properties of the CB energy operator are presented. The link between CB and the dynamic signal shows that CB conveys precious local information about the instantaneous behaviour, over time, of a signal. The second property obtained for FM signals is a simple way to estimate the IF with no a priori

1 Kaiser, J.F.: Some useful properties of Teagers energy operators. Proc. ICASSP, Minneapolis, MN, USA, 1993, Vol. 3, pp. 149152 2 Cexus, J.C., and Boudraa, A.O.: Link between cross-Wigner distribution and cross-Teager energy operator, Electron. Lett., 2004, 40, (12), pp. 778780 3 Boudraa, A.O., Benramdane, S., Cexus, J.C., and Chonavel, T.: Some useful properties of cross-CB-energy operator, Int. J. Electron. Commun., 2009, 63, (9), pp. 728735 4 Boudraa, A.O.: Relationships between CB-energy operator and some timefrequency representations, IEEE Signal Process Lett., 2010, 17, (6), pp. 527530 5 Boudraa, A.O., Cexus, J.C., and Zaidi, H.: Functional segmentation of dynamic nuclear medicine images by cross-CB-energy operator, Comput. Methods Programs Biomed., 2006, 84, (2 3), pp. 146152 6 Boudraa, A.O., Cexus, J.C., and Abed-Meraim, K.: Cross-CB-energy operatorbased signal detection, J. Acoust. Soc. Am., 2008, 123, (6), pp. 42834289 7 Zhang, W.F., Liu, C.C., and Yan, H.: Clustering of temporal gene expression data by regularized spline regression an energy based similarity measure, Pattern Recognit., 2010, 43, pp. 3969 3976 8 Boudraa, A.O., Cexus, J.C., Groussat, M., and Brunagel, P.: An energy-based similarity measure for time series, Adv. Signal Process, 2008, ID 135892, 8 pages 9 Poletti, M.A.: Instantaneous frequency and conditional moments in the time-frequency plane, IEEE Trans. Signal Process, 1991, 39, (3), pp. 755756 10 Savitzky, A., and Golay, M.J.E.: Smoothing and differentiation of data by simplied least squares procedures, Anal. Chem., 1964, 36, (8), pp. 16271639 11 Cohen, L.: Time-frequency analysis (Prentice-Hall, 1995) 12 Rao, P., and Taylor, F.: Estimation of instantaneous frequency using the discrete Wigner distribution, Electron. Lett., 1990, 26, (4), pp. 246248

ELECTRONICS LETTERS 12th May 2011 Vol. 47 No. 10