Vous êtes sur la page 1sur 6

Mismatched Filters for Doppler-Sensitive Active Sonar Pulses

First Author #1, Second Author *2, Third Author #3


#

First-Third Department, First-Third University Address Including Country Name


1 3

first.author@first-third.edu third.author@first-third.edu

Second Company Address Including Country Name


2

second.author@second.com

AbstractA technique used for designing radar poly-phase codes, is used to design mismatched filters for long-range active sonar systems, where transmit waveforms are Doppler-sensitive. The long ranges and operational scenarios involved mean that only single-ping operations are considered. The waveforms under consideration are long-duration Cox combs and sinusoidally frequency modulated (SFM) pulses, arbitrarily selected to occupy the frequency range 1 to 2 kHz. When used in conjunction with a matched filter, these waveforms offer an attractive Doppler resolution, but suffer from significant range sidelobes. The proposed mismatched filters attempt to improve their range resolution or lower their sidelobes across the zeroDoppler region of the range-Doppler domain.

I. INTRODUCTION A typical active sonar system locates and possibly identifies underwater objects by transmitting an acoustical pulse of energy, then listens for any echoes returned from objects of interest. Unwanted signals include background noise - which is independent of the transmit waveform - and reverberation - which consists of degraded reflections of the transmit waveform on features such as the seabed or the sea surface. The receive processing chain of current systems typically consists of a beamformer followed by a matched filter. The beamformer processes the array outputs (one time series per hydrophone) to compute the directional component of those signals emanating from a pre-determined direction. The matched filter is a linear filter that performs a correlation between its input and a reference waveform, here the transmit waveform. It has a transfer function that amplifies only the frequencies containing the transmit waveform spectral components. In addition, it amplifies more high amplitude spectral components (less likely to be corrupted by background noise), than low-amplitude ones. As a result, the matched filter is the detector that yields the highest possible signal-to-(background) noise ratio (SNR), assuming a Gaussian distribution for this noise. Since the output SNR is proportional to the transmit waveform energy, it is possible to achieve long-range detectability by transmitting high energy waveforms. Because of the hardware and cavitation-related limits on transmit peak power [1], long-duration waveforms are transmitted, typically 1 or 2 seconds. However, reverberation noise and signal of interest share a number of

features, especially when the signal of interest emanates from a low velocity source. This leads to reverberation being significantly amplified by the matched filter. Therefore, such systems tend to suffer from false alarms caused by reverberation. A careful choice of the transmit waveform and its associated matched filter response can help alleviating this problem. Since the signal of interest and reverberation may emanate from moving scatterers, their spectra may be altered by the Doppler effect. As a result, the matched filter response or correlator output has to be studied in terms of two variables, range or delay, and another variable quantifying the amount of Doppler effect. For narrowband signals, such as in radar systems, where the ratio of the velocity of scatterers to the velocity of the transmit signal is very low, the Doppler effect approximates to a shift of the whole spectrum in frequency. For wideband signals, such as in sonar systems, the Doppler effect amounts to a scaling of the spectrum, quantified by a compression factor. The resulting theoretical matched filter response is defined by [2] as (, ) = () [( )]

where () is the waveform, is the range or delay, and is the compression factor, equal to = cv c+v

(1)

(2)

with c the sound velocity, and v the scatterer radial velocity (for the sake of simplicity, this compression factor will now be referred to as Doppler shift, keeping in mind that equals unity when the Doppler shift in Hertz is nul). The wideband ambiguity function is the modulus of this function. Its zeroDoppler cut is the modulus of the pulse autocorrelation function. Each waveform and associated matched filter has its own ambiguity function. Ideally, it should look like a thumb-tack. But since the volume under the ambiguity function modulus square is the energy of the transmit waveform, it is impossible

to design a waveform achieving such a perfectly localized response. What is actually achieved is a central peak located at (0, 0) and surrounded by sidelobes in the range-Doppler domain. The width of the main peak, as well as the sidelobes, may lead to uncertainties or ambiguities as to the true range and Doppler scatterer parameters. Hence, the name 'ambiguity function'. The ambiguity function shapes and properties of standard transmit waveforms such as continuous wave (CW) pulses, frequency modulated (FM) pulses, pseudo-random noise, and frequency-hopped pulse trains, to name but a few, have been widely reported in the sonar and radar literature ([3] and [6] are examples of comprehensive references). The width of the ambiguity function main lobe in the range direction (called range resolution) relates to that of the autocorrelation central peak. It can be made narrower if the signal power spectrum is widened. One method to achieve a wider spectrum is frequency modulation. Chirps, or frequency-modulated waveforms thus have an excellent range resolution. This comes at the cost of a reduced sensitivity to Doppler shifts. This is because the overlap between the waveform spectrum and its Doppler-shifted spectrum remains significant. To reduce the width of the main lobe in Doppler and thus improve the Doppler resolution to create a Dopplersensitive waveform, it is necessary to make sure that nonzero Doppler shifts yield a low correlator output. In other words, the overlap between the waveform spectrum and its Dopplershifted spectra must be small. This can be achieved if its spectrum consists of one spectral peak (such as in a CW pulse), or is comb-like. The price to pay for a spectrum consisting of narrow teeth is usually a wide autocorrelation main lobe and thus a poor range resolution. The properties of more advanced waveforms, such as Cox combs [4] and sinusoidally frequency modulated (SFM) pulses have been investigated in details in [5]. Cox combs generate a comb-like spectrum by adding together sinusoids of the required frequencies. SFM pulses achieve a comb-like spectrum by sweeping the required spectrum using periodic frequency modulation. The authors show that Cox comb waveforms and SFM can be employed as an alternative to CW pulses, providing useful reverberation suppression even at low target velocities. They also show that SFM pulses are better than Cox combs in terms of ambient noise processing. In this paper, a least-square technique used for designing low sidelobes radar poly-phase codes, is used to design mismatched filters for long-range active sonar systems, where the transmit waveforms are Cox combs and SFM pulses. The aim is to outperform the matched filter in terms of range resolution or sidelobes height across the zero-Doppler region of the range-Doppler domain, where reverberation usually lies. The long ranges and operational scenarios involved mean that only single-ping operations are to be considered.

II. DOPPLER-SENSITIVE PULSES The general form of a Cox comb is () = () exp(2 )


n=1 N

(3)

where is the series of frequencies that make up the comb spectrum, and () is the amplitude window of the pulse. An example is shown in Fig. 1, where the comb consists of 5 uniformly spaced tones, 12 Hz apart: 976, 988, 1000, 1012, and 1024 Hz.

Fig. 1. Spectrum of a Cox comb.

In this example, the waveform duration is one second. The pulse amplitude window is rectangular. The corresponding wideband ambiguity function is shown in Fig. 2. The compression factor has been converted into a Doppler shift to facilitate reading.

Fig. 2. Wideband ambiguity function envelope of a Cox comb.

The 'beating' of the autocorrelation results in a series of teeth, as shown in Fig. 3. It is actually of the form of a sequence of sinc functions spaced at intervals of 1/12 seconds, 12 Hz being the frequency spacing. The main lobe width of each function is approximately 1/Bandwidth = 1/48 seconds, 48 Hz being 4 times 12 Hz. The sequence is weighted by the autocorrelation of the window function, that is to say, a triangular function.

necessary to space unevenly its spectral components. Cox has proposed a geometric progression [4]. The general form of a sinusoidally frequency modulated (SFM) pulse is where () is the amplitude window of the pulse, 0 is the centre frequency, is the modulating frequency, and is the modulation index that controls the bandwidth of the pulse. Since the frequency modulation function of this waveform is periodic its spectrum consists of spectral lobes spaced at multiples of the repetition frequency of the frequency modulation function [5]. An example is shown in Fig. 5, for a one-second duration, rectangular-windowed, 1000 Hz carrier frequency pulse. The modulating frequency is is 12 Hz, and the modulation index is 2.75, leading to an approximate bandwidth of 90 Hz, calculated using Carson's rule (4) 2 (1 + ) (5) () = ()exp[20 + sin )] (2

The teeth of the comb spectrum are uniformly spaced, 12 Hz apart.


Fig. 3. Autocorrelation function of a Cox comb.

A cut at zero-delay along the Doppler axis is shown in Fig. 4. Because of the comb structure of the spectrum, the overlap between the waveform spectrum and any Doppler-shifted replica is small. However, first major Doppler sidelobes appear here at 12 Hz, which is the comb frequency spacing. This gives a range of resolvable velocities for slow targets. At these Doppler shifts, 4 teeth only are overlapping, resulting in a sidelobe height lower than that of the zero-Doppler main lobe.

Fig. 5. Spectrum of an SFM pulse.

The corresponding wideband ambiguity function is shown in Fig. 6. The autocorrelation consists of a series of teeth, spaced 1/12 seconds apart, as shown in Fig. 7, 12 Hz being the modulating frequency. The main lobe width of each function is approximately 1/Bandwidth = 1/90 seconds. The sequence is weighted by the autocorrelation of the window function, that is to say, a triangular function. Finally, a cut at zerodelay along the Doppler axis is shown in Fig. 8. As expected, the first major Doppler sidelobes appear here at 12 Hz.
Fig. 4. Zero-delay cut of the wideband ambiguity function of a Cox comb.

To lower the major Doppler sidelobes and thus avoid the realignment of the spectrum teeth cause by Doppler shifts, it is

III. MISMATCHED FILTER DESIGN When the receiver is no longer a matched filter, the receive filter two-dimensional response is called the cross-ambiguity function and is defined by , (, ) = () [( )]

Fig. 6. Wideband ambiguity function envelope of an SFM pulse.

Fig. 7. Autocorrelation function of an SFM pulse.

Fig. 8. Zero-delay cut of the wideband ambiguity function of an SFM pulse.

where () and () are the receive and transmit waveforms, respectively. This section succinctly reviews some of the ideas underlying mismatched filter design in the literature. Generally speaking, filter design does not have to be carried out in the canonical basis, whereby a filter impulse response is expressed in terms of its time samples. Some authors design filters or waveforms in other orthonormal bases, using the inner product defined over the space of functions as a tool. These bases are either built using ready-made and well-known families of orthogonal polynomials, such as the Hermite polynomials [12], or using the Gram-Schmidt orthonormalization procedure from an initial vector [11]. The use of Slepian sequences (also known as discrete prolate spheroidal sequences) has also been envisaged in [12]. Other authors have recourse to a decomposition of a waveform in terms of known sub-pulses [13]. In [7], mismatched filters are designed that are constrained to have a pre-defined correlation coefficient with the transmit signal. They are designed so that the total energy over a desired region of the range-Doppler domain is minimized. The narrowband cross-ambiguity function is the measure used to quantify the mismatched filter response. In [8] mismatched filters are designed, that minimize the total energy in their range sidelobes, while offering good pulse compression performances. The design concept is based on the minimization of an objective function representing the integrated sidelobe level (ISL), subject to a linear constraint setting the peak response to a pre-defined constant. A weighting function allows a degree of control over the relative heights of the sidelobes. Improved results are obtained by using longer mismatched filters, for instance three times longer than the matched filter. The authors also propose a method to minimize the peak sidelobe level (PSL) in which they iteratively improve the solution of the ISL minimization problem by modifying the weight function at each step (iteratively reweighted least squares, IRLS). However, the mismatched filter response, optimized for minimum ISL or PSL, produces better pulse compression performances when the signal matched response is initially favourable [9]. In addition, filters designed using this method exhibit a high degree of sensitivity to ambient white noise. This is because the design method yields mismatched filters whose spectrum is generally close to the inverse of the signal spectrum. Spectrally, these filters do the opposite of what a matched filter does, and as a result, in presence of white noise, their performance is severely degraded. A necessary - but not sufficient - condition for these filters to achieve a low SNR loss is that the signal auto-correlation function should already achieve a good ISL (or "merit factor"). The topic of which

(6)

signals lend themselves to good minimum ISL mismatched filters with low SNR loss, is presently an active research topic. In [10], sidelobe suppression is addressed using a pair of 'contrasting' mismatched filters constructed using the procedure described in [8]. The sidelobes of the first filter peak at different locations than those of the second filter. A simple blanking logic is used to cancel out those sidelobes. In [11], mismatched filters are designed under the constraint of a controlled loss between transmit waveform and receive filter (matched filtering loss). The author decomposes the unknown receive filter over an orthonormal basis whose first vector is the transmit waveform. The rest of the basis vectors are constructed using the Gram-Schmidt orthonormalization procedure. This decomposition allows the constructed mismatched filter to satisfy by design a predetermined matched filtering loss. The sidelobe energy is a quadratic form. The only constraint in the minimization problem is that the mismatched filter energy which must be set to a pre-determined constant. The solution is not sought numerically but by employing algebraic operations on data matrices such as diagonalization. It is noteworthy that the designed mismatched filter is guaranteed to be close to the matched filter via the controlled matched filtering loss, as well as the pre-defined mismatched filter energy. As a result, the mismatched filter response can exhibit a high signal-to-noise ratio (SNR). In this work the method proposed in [8] has been followed. The underlying mathematical framework is given in details in [8] and [9]. IV. NUMERICAL RESULTS A. Example 1: Cox Comb The filter was chosen to be 50% longer than the original Cox comb. The designed filter achieves an SNR loss of 18.5 dB. The corresponding cross-ambiguity function is shown in Fig. 9. Zero-Doppler and zero-delay cuts are given in Fig. 10 and 11.

Fig. 10. Cross-correlation function of a Cox comb and the optimum IRLS filter.

Fig. 11. Zero-delay cut of the wideband cross-ambiguity function of a Cox comb and the optimum IRLS filter.

The mismatched filter has the desirable impulse-like response along the time axis, with lower sidelobes. In addition, it succeeds, to some extent, to clear an area around the zero-Doppler ridge, usually occupied by reverberation. However, there are undesirable Doppler sidelobes at zero delay. B. Example 2: Sinusoidally Frequency Modulated (SFM) Pulse The filter was also chosen to be 50% longer than the original SFM. The designed filter achieves an SNR loss of 20.5 dB. The corresponding cross-ambiguity function, zeroDoppler cut, and zero-delay cut are shown in Fig. 12, 13, and 14.

Fig. 9. Wideband cross-ambiguity function envelope of a Cox comb and the optimum IRLS filter.

The mismatched filter designed for the SFM pulse also exhibits a desirable impulse-like response along the time axis (with low sidelobes), but is less successful in achieving low Doppler sidelobes. V. CONCLUSION Mismatched filters have been designed using the IRLS method, for use with Doppler-sensitive active sonar waveforms. They were chosen to be 50% longer than the original waveforms. Preliminary results are encouraging as the filters have at zero-Doppler, an impulse-like response with low sidelobes. They also manage to clear partially from sidelobes the region of the delay-Doppler domain around zero-Doppler. Further work includes the generalization of the IRLS objective function to sidelobes of the filter response in the delay-Doppler domain, as opposed to delay only for the original method. In addition, a constraint on the SNR loss will be added. REFERENCES
[1] A. D. Waite, Sonar for Practising Engineers, Thomson Marconi Sonar Ltd, Second Ed, 1998. [2] Z. B. Lin, Wideband ambiguity function of broadband signals, Journal of Acoust. Soc. Am., vol. 83, no. 6, pp. 21082116, 1988. [3] S. Pecknold, Ambiguity and cross-ambiguity properties of some reverberation suppressing waveforms, Defence Research and Development Canada, DRDC Atlantic TM 2002-129, 2002. [4] H. Cox and H. Lai, Geometric comb waveforms for reverberation suppression, in Proc. 1994 Asilomar Conference on Signals, System, and Computers, vol. 2, pp. 11851189. [5] T. Collins T, P. Atkins, Doppler-sensitive active sonar pulse designs for reverberation processing, IEE proc-radar, sonar Navig, vol. 145, no.6, pp. 347-353, 1998. [6] N. Levanon and E. Mozeson, Radar Signals, NY: John Wiley & Sons, 2004. [7] C. A. Stutt and L. J. Spafford, A 'best' mismatched filter response for radar clutter discrimination, IEEE Trans. Info. Theory, vol. 14, no.2, pp. 280287, Mar. 1968. [8] K. R. Griep, J. A. Ritcey, and J. J. Burlingame, Poly-phase codes and optimal filters for multiple user ranging, IEEE Trans. Aerospace and Electronic Systems, vol. 31, no. 2, pp. 752767, Apr. 1995. [9] N. Levanon, Cross-correlation of long binary signals with longer mismatched filters, IEE Proc., Radar Sonar Navig., vol. 152, no. 6, pp. 372382, 2005. [10] N. Levanon and A. Scharf, Range sidelobes blanking by comparing outputs of contrasting mismatched filters, IET Radar Sonar Navig., vol. 3, iss. 3, pp. 265277, 2009. [11] C. Candan, On the design of mismatched filters with an adjustable matched filtering loss, in IEEE 2010 Radar Conference, pp. 1311 1316. [12] I. Gladkova and D. Chebanov, On a new extension of Wilcox's method, in WSEAS Trans. on Mathematics, vol. 3, pp. 244249, 2004. [13] H. He, P. Stoica, and J. Li, On synthesizing cross ambiguity functions, in ICASSP 2011 (International Conference on Acoustics, Speech and Signal Processing).

Fig. 12. Wideband cross-ambiguity function envelope of an SFM pulse and the optimum IRLS filter.

Fig. 13. Cross-correlation function of an SFM pulse and the optimum IRLS filter.

Fig. 14. Zero-delay cut of the wideband cross-ambiguity function of an SFM pulse and the optimum IRLS filter.

Vous aimerez peut-être aussi