Sunteți pe pagina 1din 7

The Analysis of Multiple Linear Chirp Signals

Chris Capus * Image Analysis Research Group Heriot-Watt University Yuri Rzhanov, Laurie Linnett t Center for Coastal and Ocean Mapping University of New Hampshire

Abstract
This paper presents a technique for the analysis and extraction of multiple linear chirps in a time signal. These signals occur in nature and are becoming increasingly important in sonar, geophysical, ultrasonic and radar applications. The work is based on and closely related to the fractional Fourier transform (FrFT), which was introduced in its current form by Namias in 1979,although the principles underlying the FrFT can be found in the work of Wiener and Weyl in the 1920s. The paper considers discrete analysis and synthesis of complex signals containing linear chirps with various characteristics. We show how individual chirps in a mixture of chirps (not necessarily linear) can be extracted and investigate the filtering and reconstruction of mixed chirp signals. Examples are presented to illustrate the concepts using both synthetic signals and real data. A visual interpretation of the magnitude and phase of the analytic results is introduced allowing a range of transform orders to be viewed simultaneously. The filtering of linear and 'near-linear' chirp signals is also discussed alongsidethe efficient pulse compression characteristics of the FrFT.

Introduction

Efficient and practical methods for the analysis and extraction of multiple linear and non-linear chirps are becoming increasingly important in sonar, geophysical, biomedical and radar applications. In recent years the fractional Fourier transform (FrFT), introduced by Namias [6] in 1979,has received significant attention, particularly in the optics literature. In this paper we explore the application of methods based on the FrFT to the analysis of real and synthetic signals containing multiple chirps with various characteristics.

The Analytic Signal and the FrFT

Equation (1) is the representation of the fractional Fourier transform, modified slightly from Namias' original [5], which has formed the basis for subsequent FrFT algorithms.

, __ -9t a) J exp
-ca

(-e +

ziy" cot a f(y) dy

where, 3ais the fractional transform of order a {a : 0.0 < a < ~ 1 2 ) .When a = 7 / x, y represent r2 frequency and time respectively and the equation is equivalent to the conventional Fourier transform. The equation is seen to be a multiplication by a chirp in one domain, followed by Fourier transformation, followed by multiplication by a chirp in the transformed domain and finally a complex scaling [l]. The FrFT code used in the followingexamples was derived from the digital implementations suggested by Ozaktas et a1 [7]. The benefits of using the analytic signal with the FrFT are illustrated with the simple example below. Fig. 1 shows a real signal and its FrFT magnitude plot at the order which produces a maximum spike output. Also shown is the spike alone produced by transforming the analytic signal through the same order FrFT.
*ceecgc@cee.hw.ac.uk tek +44 (0)131 451 3357
tel: +l 603 862 3755
~yrzhanov@cisunix.unh.edu, llinnett@cisunix.unh.edu

4/1

0 2000 The Institution of Electrical Engineers. % Printed and published by the IEE, Savoy Place, London WCPR 0

UK.

Figure 1: (i) Real chirp signal (ii) FrFT of real signal (iii) FrFT of analytic signal constructed from (i) using Hilbert transform The Hilbert transform has been used extensively in the following examples to convert real signals to an analytic form comprising the positive frequency components of the input signals only. Where synthetic data are used these too are generally complex signals.

3 Instantaneous Frequency
Central to detailed examination of chirp signals and fractional domains is the concept of instantaneous frequency. We have followed Cohen [3], taking instantaneous frequency to be the derivative of the phase of the analytic signal. :Signal :Phase :Instantaneous Frequency Symmetry in the visualisations given below is improved by defining t=O to be at the centre of the input signal. In conjunction with the conventional double sided frequency spectrum, f=O central in the transform space, this gives a central symmetry through all of the intervening fractional domains. Further definitions are: initial frequency - frequency of signal at t=O; initial phase - phase of signal at t=O. These are the parameters b and c, respectively, in the signal s ( t ) = expj(at2 bt + c ) .

Visualisation of Fractional Domains

An image representation is introduced to facilitate understanding of the changing concentration of energy through consecutive fractional domains.
inverse time signal

inverse frequency domain

time signal

frequency domain

invase time signal

Figure 2: FrFTs for a real linear chirp signal evaluated a t 400 equally spaced points in range

-H

< CY < H

The image in Fig. 2 is generated from the output of several hundred FrFTs calculated at orders ranging from -7 to H relative to the time domain (assumed sampling space of input signal). This is a magnitude image

412

in which each row represents a complete FrFT, the magnitude at each point given by the pixel intensity. Ordinarily we are interested in orders 0.0 (time domain) through 7r/2 (frequency domain). Fig. 2 is provided to show how the FrFT relates to the modulo 4 symmetry of the conventional Fourier transform and to illustrate the distribution of energy in the positive and negative frequency components of a real signal through successive FrFTs.

Synthetic Data- Chirp Separation

The example shown in Fig. 3 uses three synthetic analytic linear chirps constrained by Gaussian envelopes and illustrates the methods available for filtering in the fractional domains. Two of the chirps are identical in rate and frequency content but are offset in time. The third is a higher rate chirp.

Figure 3: (i) Time signal (ii) FFT (mag) of time signal (iii) FrFT (mag) of analytic signal: order 0.722 It is seen that the three chirps are overlapping in both the time and frequency domains but can be cleanly separated using an FrFT whose order corresponds with the chirp rate of one of the components. Calculation of this order is further detailed below. Extraction and reconstruction of just one of the chirps from the mixture can be achieved by performing an equivalent inverse FrFT (in this case of order -0.722) on one of the spike components in the complex FrFT output. The appropriate range of samples can be selected from the spikes seen in Fig. 3 (iii).

6
6.1

Samples and Resolution


Calculating the Transform Order and Chirp Rate

We have two methods for describing chirp rate. The first is the quadratic phase parameter, a, in the algebraic definition of the linear chirp (ej(=* +bt+c). This is the rate of change of frequency with time and values will typically be given in Hzs-'. The second is the a parameter in the FrFT definition.

'I

Figure 4: Origin of the FrFT a-values from the T-F Representation: {a : 0.0

< cr < r/2)

We can derive a simple geometric relationship from the time-frequency representation for the chirp, Fig. 4. Unsurprisingly the true relationship is dependent on the digital sampling scheme used and can be calculated from the time and frequency resolutions familiar from conventional Fourier analysis.

The geometric relationship,

cy

= tan-'

(k)
,
, where,
and,

is modified by the sampling scheme to,

a = tan-'

(y)

bf bt
fs

= frequency resolution = f , / N = time resolution = l/fs.


= sampling frequency

N
Hence,

= number ofsamples.

a = tan-'

(-ia>
f:/N

(4)

Equation (4) is used to calculate the optimum FrFT order (a) a given rate chirp (Hrs-') where the for sampling rate of the digitised signal and the number of samples are known.

62 .

Time and Frequency Resolution

For any given order of transform, specific for a particular rate of linear chirp, the individual chirps are distinguished on the basis of initial frequency. The linear relationship between time and frequency within these signals also allows us to determine time spacings between chirp components from these frequency values. However, we still cannot overcome time-frequency uncertainty. Whilst we can arbitrarily select an a-value to determine the order of the FrFT we wish to perform, that order will in turn determine the time and frequency resolutions attainable. We can ask the question what are the time and frequency shift equivalents of a one-sample shift in the FrFT output? For the definition given in equation (l), axis rescaling results in the following equivalences between the samples in the time ( N t ) ,frequency ( N j ) and fractional ( N , ) domains,

Nj Nt

= N,sin(a)
= Nu COS(Q)

(5)

So, a one sample shift in the fractional domain is equivalent to a 1/ sin(a) shift in frequency and a 1/ COS(Q) time shift. In practice we will require at least one intervening sample to separate peaks so that the theoretical resolution limits become 1/2 sin(a) . f s / N in the frequency domain and 1/2 cos(a) . l/Fs in the time domain. The non-intuitive apparent decrease in time resolution arising from a higher sampling frequency for a given signal is offset by the change in a which accompanies the resampling (4). In one FrFT implementation we have encountered [4] the sampling extension is built into the routine as a parameter.

7 R a Data el

- the Bat Signal

As an example using real data the FrFT image, Fig. 6, for a bat signal', Fig. 5, is constructed. Previous work by Bultan [2], which utilised the FrFT, indicated the presence of three strong linear chirp components in this signal. The F'rFT image clearly shows a fourth element and allows us to determine the most appropriate orders of FrFT for filtering and extraction of the individual chirps.
lThe authors wish to thank Curtis Condon, Ken White, and Al Feng of the Bedanan Institute of the University of Illinois for the bat data and for permission to use it i this paper. n

414

l a

3m

1Do

Figure 5: 250ms bat chirp

Figure 6: FrFT image of analytic bat signal

7.1

Reconstruction of Individual Chirps

Perfect linear chirps represented at an FrFT order precisely reflecting the chirp rate would produce impulses. This is clearly not the case here and indeed these chirps are not truly linear. The individual chirps can, however, be recovered from these FrFTs. For each component the RFT order is chosen which separates it most cleanly from the others. The filtering process consists of finding the minimum point either side of the chirp component to be extracted and zeroing all values outside this range in the complex FrFT output. An inverse FrFT of equivalent order is applied to the resultant and the real part of its output gives the reconstructed chirp component in the time domain. The reconstructions shown in Fig. 7 use all of the available energy separable from the other main components in the original signal. They are reconstructions from the optimum compression of each chirp component under the FrFT. There is clearly a need to demonstrate that the reconstructions are justifiable given that they have come from FrFTs of different orders. If they are an accurate reflection of the true components adding them together should give us back the original bat signal. Fig. 8 (i) shows the addition of these four components and Fig. 8 (ii) shows a zoomed section of this signal overlaid on the original. This comparison demonstrates the excellent phase coherence and amplitude approximation possible with the FrFT even when the input signals are not truly linear.

7.2

Time-frequency Representation

Following Bultans example we can create a good approximation to the constituents of the bat chirp by plotting the timefrequency representation of the linear chirps defined by the maxima of the four components seen in the FrFT image. In this example the intensity at each point in time is determined by the magnitude of the individually reconstructed chirps at that time.

Conclusions

The FrFT and associated techniques are proving to be valuable tools for analysis and filtering with signals containing linear and near-linear chirp components. In the future these techniques are likely to be particularly useful in active signal processing applications, where we have control over the design of the signals used.

415

O.?

'I

a '

Figure 7 Individually reconstructed chirp components :

Figure 8: (i) Reconstructed bat signal (ii) Zoomed section of reconstructed signal overlaid on original The image representations introduced here can be used to highlight chirp components within a signal and provide insight into optimum FrFT orders which might be used for filtering. Future work is likely to proceed with the development of transforms with a more general functional form and with piecewise FrFT fitting of non-linear signals.

References
[l] Luis B. Almeida. The Fractional Fourier Transform and Time-Frequency Representations. IEEE Transactions on Signal Processing, 42( 1l):pp. 3084-3091, November 1994. [2] Aykut Bultan. A Four-Parameter Atomic Decomposition of Chirplets. IEEE Transactions on Signal PTOcessing, 47(3):pp. 731-745, March 1999. [3] Leon Cohen. Time-Frequency Analysis. Prentice-Hall, New Jersey, 1995.
[4] Francisco J. Marinho and Luis M. Bernardo. Numerical Calculation of Fractional Fourier Transforms with a Single Fast-Fourier-Transform Algorithm. Joumal ofthe Optical Society of Amen" A , 15(8):pp. 2111-2116,

August 1998.

4/6

0.5

1.o

15

20

25

3.0

t (m)

Figure 9: T-F plot for four-component bat chirp [5] A.C. McBride. On Namias s Fractional Fourier Transforms. IMA Journal of Applied Mathematics, 39:pp. 159-175,1987.

[6] Victor Namias. The Fractional Order Fourier Transfrom and its Application in Quantum Mechanics. Journal
of the Institute of Mathematics Applications, 25(3):pp. 241-265, 1980.

[7] Haldun M . Ozaktas et al. Digital Computation of the Fractional Fourier Transform. IEEE Transactions on Signal Processing, 44(9):pp. 2141-2149, September 1996.

4/7

S-ar putea să vă placă și