Patent Grant
7106250
This invention relates to a method and system for predictive deconvolution, which is otherwise known as pulse compression in radar applications. More particularly, the invention relates to robust predictive deconvolution using minimum mean-square error reiteration.
In many sensing applications, such as radar pulse compression, sonar, ultrasonic non-destructive evaluation for structural integrity, biomedical imaging, and seismic estimation, it is desirable to estimate the impulse response of an unknown system by driving the system with a known signal having a finite temporal extent. The process of separating the known signal from the received output of the system in order to estimate the unknown impulse response is known as predictive deconvolution.
Predictive deconvolution provides a means to obtain the high spatial resolution of a short, high bandwidth pulse without the need for very high peak transmit power, which may not be feasible. This is accomplished by transmitting a longer pulse that is phase or frequency modulated to generate a wideband waveform. The transmission of the wideband waveform into the unknown system results in a received return signal at the sensor that is the convolution of the waveform and the system impulse response, which possesses large coefficient values at sample delays corresponding to the round-trip travel time of the transmitted waveform from the sensor to a significant reflecting object (also called a scatterer) and back to the sensor. The purpose of predictive deconvolution is to accurately estimate the unknown system impulse response from the received return signal based upon the known transmitted waveform.
A well-known approach to predictive deconvolution, used extensively in radar and biomedical imaging applications, is known as matched filtering, e.g. as described in M. I. Skolnik, Introduction to Radar Systems, McGraw-Hill, New York, 1980, pp. 420–434; T. X. Misaridis, K. Gammelmark, C. H. Jorgensen, N. Lindberg, A. H. Thomsen, M. H. Pedersen, and J. A. Jensen, “Potential of coded excitation in medical ultrasound imaging,” Ultrasonics, Vol. 38, pp. 183–189, 2000. Matched filtering has been shown to maximize the received signal-to-noise ratio (SNR) in the presence of white Gaussian noise by convolving the transmitted signal with the received radar return signal. One can represent matched filtering in the digital domain as the filtering operation
{circumflex over (x)}MF(l)=sH{tilde over (y)}(l), (1)
where {circumflex over (x)}MF(l) for l=0, . . . , L−1, is the estimate of the lth delayed sample of the system impulse response, s=[s1 s2 . . . sN]T is the length-N sampled version of the transmitted waveform, {tilde over (y)}(l)=[y(l) y(l+1) . . . y(l+N−1)]T is a vector of N contiguous samples of the received return signal, and (●)H and (●)T are the conjugate transpose (or Hermitian) and transpose operations, respectively. Each individual sample of the return signal can be expressed as
y(l)={tilde over (x)}T(l) s+v(l), (2)
where {tilde over (x)}(l)=[x(l) x(l−1) . . . x(l−N+1)]T consists of samples of the true system impulse response and v(l) is additive noise. The matched filter output can therefore be written as
{circumflex over (x)}MF(l)=sHAT(l) s+sHv(l), (3)
where v(l)=[v(l) v(l+1) . . . v(l+N−1)]T and
is a collection of sample-shifted snapshots (in the columns) of the impulse response.
From (4), it is obvious that estimation via matched filtering will suffer from spatial ambiguities (also known as range sidelobes in the radar vernacular) due to the influence from neighboring impulse response coefficients (i.e. the off-diagonal elements of A(l)). To alleviate this effect, Least-Squares (LS) solutions have been proposed, e.g. in M. H. Ackroyd and F. Ghani, “Optimum mismatched filters for sidelobe suppression,” IEEE Trans. Aerospace and Electronic Systems, Vol. AES-9, pp. 214–218, March 1973; T. Felhauer, “Digital Signal Processing for Optimum Wideband Channel Estimation in the Presence of Noise,” IEE Proceedings-F, Vol. 140, No. 3, pp. 179–186, June 1993; S. M. Song, W. M. Kim, D. Park, and Y. Kim, “Estimation theoretic approach for radar pulse compression processing and its optimal codes,” Electronic Letters, Vol. 36, No. 3, pp. 250–252, February 2000; B. Zmic, A. Zejak, A. Petrovic, and I. Simic, “Range sidelobe suppression for pulse compression radars utilizing modified RLS algorithm,” Proc. IEEE Int. Symp. Spread Spectrum Techniques and Applications, Vol. 3, pp. 1008–1011, September 1998; and T. K. Sarkar and R. D. Brown, “An ultra-low sidelobe pulse compression technique for high performance radar systems,” in Proc. IEEE National Radar Conf., pp. 111–114, May 1997. LS solutions decouple neighboring impulse response coefficients which have been smeared together due to the temporal (and hence spatial) extent of the transmitted waveform. The LS solution models the length-(L+N−1) received return signal as
y=Sx+v, (5)
where x=[x(0) x(1) . . . x(L−1)]T are the L true impulse response coefficients that fall within the data window, v=[v(0) v(1) . . . v(L+N−2)]T are additive noise samples, and the convolution of the transmitted waveform with the system impulse response is approximated as the matrix multiplication
The LS model of (6) is employed extensively in radar pulse compression, seismic estimation, e.g. as described in R. Yarlaggadda, J. B. Bednar, and T. L. Watt, “Fast algorithm for lp deconvolution,” IEEE Trans. Acoustics, Speech, and Signal Processing, Vol. ASSP-33, No. 1, pp. 174–182, February 1985; and ultrasonic non-destructive evaluation, e.g. as in M. S. O'Brien, A. N. Sinclair, and S. M. Kramer, “High resolution deconvolution using least-absolute-values minimization,” Proc. Ultrasonics Symposium, pp. 1151–1156, December 1990; and D.-M. Suh, W.-W. Kim, and J.-G. Chung, “Ultrasonic inspection of studs (bolts) using dynamic predictive deconvolution and wave shaping,” IEEE Trans. Ultrasonics, Ferroelectrics, and Frequency Control, Vol. 46, No. 2, pp. 457–463, March 1999. The general form of the LS solution is
{circumflex over (x)}LS=(SHS)−1SHy. (7)
For the received signal model of (5), it can be shown that the LS solution of (7) is optimal in the mean-square error (MSE) sense when the additive noise is white. However, upon further inspection one finds that the LS received signal model does not completely characterize the received return signal because it does not account for the convolution of the transmitted waveform with impulse response coefficients x(l) prior to l=0. The result is that the presence of a significant impulse response coefficient within N−1 samples prior to x(0) can cause severe mis-estimation of the desired coefficients within the data window.
There is, therefore, a need for a predictive deconvolution system with improved robustness, accuracy, and resolution.
According to the invention, a method for processing a received, modulated pulse that requires predictive deconvolution to resolve a scatterer from noise and other scatterers includes (a) receiving a return signal; (b) obtaining L+(2M−1)(N−1) samples y of the return signal, where y(l)={tilde over (x)}T(l)s+v(l); (c) applying RMMSE estimation to each set of N contiguous samples to obtain initial impulse response estimates [{circumflex over (x)}1{−(M−1)(N−1)}, . . . , {circumflex over (x)}1{−1}, {circumflex over (x)}1{0}, . . . , {circumflex over (x)}1{L−1}, {circumflex over (x)}1{L}, . . . , {circumflex over (x)}1{L−1+(M−1)(N−1)}]; (d) computing power estimates {circumflex over (ρ)}1(l)=|{circumflex over (x)}1(l)|2 for l=−(M−1)(N−1), . . . , L−1+(M−1)(N−1); (e) computing MMSE filters according to w(l)=ρ(l) (C(l)+R)−1 s, where ρ(l)=E[|x(l)|α] is the power of x(l), for 0<α≦2, and R=E[v(l) vH(l)] is the noise covariance matrix; (f) applying the MMSE filters to y to obtain [{circumflex over (x)}2{−(M−2)(N−1)}), . . . , {circumflex over (x)}2{−1}, {circumflex over (x)}2{0}, . . . , {circumflex over (x)}2{L−1}, {circumflex over (x)}2{L}, . . . , {circumflex over (x)}{L−1(M−2)(N−1)}]; and repeating (d)-(f) for subsequent reiterative stages until a desired length-L range window is reached, thereby resolving the scatterer from noise and other scatterers.
Also according to the invention, a radar receiver system includes a receiver, a processor including a RMMSE radar pulse compression algorithm, and a target detector.
Also according to the invention, a method for processing a received, modulated radar pulse to resolve a radar target from noise or other targets includes (a) receiving a radar return signal; (b) obtaining L+(2M−1)(N−1) samples y of the radar return signal, where y(l)={tilde over (x)}T(l)s+v(l); (c) applying RMMSE pulse compression to each set of N contiguous samples to obtain initial radar impulse response estimates [{circumflex over (x)}1{−(M−1)(N−1)}, . . . , {circumflex over (x)}1{−1}, {circumflex over (x)}1{0}, . . . , {circumflex over (x)}1{L−1}, {circumflex over (x)}1{L}, . . . , {circumflex over (x)}1{L−1+(M−1)(N−1)}]; (d) computing power estimates {circumflex over (ρ)}1(l)=|{circumflex over (x)}1(l)|2 for l=−(M−1)(N−1), . . . , L−1+(M−1)(N−1); (e) computing range-dependent filters according to w(l)=ρ(l) (C(l)+R)−1 s, where ρ(l)=E[|x(l)|α] is the power of x(l), for 0<α≦2, and R=E[v(l) vH(l)] is the noise covariance matrix; (f) applying the range-dependent filters to y to obtain [{circumflex over (x)}2{−(M−2)(N−1)}, . . . , {circumflex over (x)}2{−1}, {circumflex over (x)}2{0}, . . . , {circumflex over (x)}2{L−1}, {circumflex over (x)}2{L}, . . . , {circumflex over (x)}2{L−1+(M−2)(N−1)}]; and repeating (d)–(f) for subsequent reiterative stages until a desired length-L range window is reached, thereby resolving the radar target from noise or other targets.
The Reiterative Minimum Mean-Square Error (RMMSE) predictive deconvolution system and method of the invention provide high-fidelity impulse response estimation compared to other systems such as matched filtering and LS. In one embodiment, the RMMSE estimator reiteratively estimates the MMSE filter for each specific impulse response coefficient by mitigating the interference from neighboring coefficients that is a result of the temporal (i.e. spatial) extent of the transmitted waveform. The result is a robust estimator that adaptively eliminates the spatial ambiguities that occur when a fixed receiver filter is used.
The invention has obvious applications in military radar and sonar, and it is also useful for civilian radar, e.g. in airport radar systems and in weather radar systems, e.g. Doppler radar, and other environmental radar applications. It may also find use in range profiling, image recognition for Synthetic Aperture Radar (SAR) and Inverse SAR (ISAR), remote sensing, ultrasonic non-destructive evaluation for structural integrity, seismic estimation, biomedical imaging, inverse filtering of optical images, or any other application requiring robust deconvolution of a known waveform (or filter) from a desired unknown impulse response.
Additional features and advantages of the present invention will be set forth in, or be apparent from, the detailed description of preferred embodiments which follows.
Definitions: The term “convolution” means the process that yields the output response of an input to a linear time-invariant system, such as is described and defined in J. G. Proakis and D. G. Manolakis, Digital Signal Processing: Principles, Algorithms, and Apiplications, 3rd Ed., pp. 75–82, Prentice Hall: Upper Saddle River, N.J. (1996), incorporated herein by reference. The term “deconvolution” as used herein means the process that given the output of a system determines an unknown input signal to the system. See Id. at p. 355, incorporated herein by reference. The term “scatterer” means something in the path of a transmitted waveform that causes a significant reflection (relative to the noise) back to the receiver of the sensor.
Referring now to
Minimum Mean-Square Error (MMSE) estimation is a Bayesian estimation approach that employs prior information in order to improve estimation accuracy. The exact form that the prior information will take will be addressed shortly. First, however, the signal model must be constructed. From (3), we see that the collection of N samples of the received return signal can be expressed as
{tilde over (y)}(l)=AT(l)s+v(l). (8)
This is the received signal model used by the matched filter formulation and takes into account all the necessary impulse response coefficients for a given received return sample. To develop the MMSE filter, the matched filter sH in (3) is replaced with the MMSE filter, denoted wH(l), in which the form of the MMSE filter is dependent upon the length-N sample swath of the impulse response under consideration. Thereafter, the MMSE cost function
J(l)=E[|x(l)−wH(l){tilde over (y)}(l)|2] (9)
is solved for each impulse response coefficient l=0, . . . , L−1, where E[●] denotes expectation. This is done by differentiating with respect to wH(l) and then setting the result to zero. The MMSE filter is found to take the form
w(l)=(E[{tilde over (y)}(l){tilde over (y)}H(l)])−1E[{tilde over (y)}(l) x*(l)], (10)
where (●)* is the complex conjugate operation. After substituting for {tilde over (y)}(l) from (8) and assuming that the impulse response coefficients are, in general, uncorrelated with one another and are also uncorrelated with the noise, one obtains
w(l)=ρ(l)(C(l)+R)−1s, (11)
where ρ(l)=E[|x(l)|α] is the expected power of x(l), for 0<α≦2, and R=E[v(l) vH(l)] is the noise covariance matrix. Preferably, α is in the range from about 0.9 to about 1.7 which has been shown by experimentation to work well. Furthermore, the value of a may vary from one stage to the next. Accordingly, α may be set at a relatively large value initially, e.g. on the order of 1.4–1.7, allowing for fast convergence, and subsequently decreased, e.g. to 0.9–1.1, in order to resolve small details (the first-stage (initial) filter described below in equations 13 and 14 treats all range cells estimates as equal, which is essentially the same as setting α0 to 0).
On assuming neighboring coefficients are uncorrelated, the (i, j)th element of the matrix C(l) is
in which κL=max{0, i−j} is the summation lower bound and κU=min{N−1, N−1+i−j} is the upper bound. Also, any prior information regarding the noise can be employed via the noise covariance matrix R. For instance, for a white noise assumption R is diagonal.
In its current state the MMSE filter is a function of the powers of the surrounding impulse response coefficients, which in practice are unavailable. This lack of prior knowledge can be taken into account by setting all the initial coefficient estimates equal. Therefore, the initial MMSE filter reduces to the form
{tilde over (w)}≅{tilde over (C)}−1s (13)
where the noise term is assumed negligible and
is invariant to the sample delay l. The initial MMSE filter can therefore be pre-computed and then implemented in the same way as the traditional matched filter. The inclusion of the matrix {tilde over (C)} serves to provide a “local” LS initial estimate.
The initial estimate of the impulse response found by applying the MMSE filter is used as a priori information to reiterate the MMSE filter and improve performance. This is done by employing the MMSE filter formulation from (11) in which the respective powers of the impulse response coefficients are taken from the current estimate. It has been found that two or three reiteration steps allow the RMMSE filter to exclude the effects of scatterers prior to the data window, as well as to suppress the spatial ambiguities very close to the level of the noise floor. The RMMSE filter does especially well when the impulse response is somewhat sparsely parameterized (i.e. highly spiky), as is the case for system designed to have high resolution. Note that each reiteration step will reduce the number of coefficient estimates by 2(N−1). To counteract this, it is necessary to increase the size of the data window by 2M(N−1) samples, where M is the number of reiteration steps. Typically, however, L>>N so that this reduction in data window size is negligible.
An important factor regarding practical implementation is the non-singularity of the N×N matrix ([C(l)]+[R]). This can be addressed by instituting a nominal level for which the estimated coefficients are not allowed to go below. An alternative to this would be to re-estimate only those coefficients that are above some threshold since the small-valued coefficients do not contain a detectable scatterer.
To demonstrate the performance of the RMMSE algorithm we examine three cases. The first case is typical of the scenario often addressed for LS estimation techniques and consists of an impulse response with a single large scatterer in noise and low-level clutter. The second case we address is when there is a second large scatterer that resides just prior to the data window. In this case it is expected that the LS estimator will substantially degrade since this important region is not accounted for in the received signal model. The final case is a more realistic scenario in which there are several spatially dispersed scatterers of varying magnitude.
For the first two cases, the waveform selected is the length N=30 polyphase modulated Lewis-Kretschmer P3 code, which on receive (after down-conversion to baseband in the receiver front-end) is defined as
The noise and clutter are modeled as zero-mean Gaussian with powers set to 70 dB and 50 dB below the signal power, respectively. For both cases we perform two iterations of the RMMSE algorithm and then compare the RMMSE-estimated impulse response with the true impulse response, as well as with the results obtained from using LS and the matched filter. The impulse response for the first case is depicted in
For the scenario just described, the LS and RMMSE estimators perform almost identically. However, when there is a significant scatterer present just prior to the range window, as depicted in
Table 1 presents the Mean Squared-Error (MSE) for both cases discussed using LS, RMMSE, and normalized matched filtering. The MSE is averaged for all 200 coefficients in the data window and over 100 runs using the same scatterer(s), with the clutter and noise distributed according to a zero-mean Gaussian distribution for each run independently.
For case 1 (single scatterer), the normalized matched filter has a large MSE due to spatial ambiguities, while LS and RMMSE perform nearly the same with LS just marginally better. However, for case 2 (scatterer just prior to the data window), the MSE attained by LS degrades to nearly that of the normalized matched filter, while the RMMSE maintains roughly the same MSE as in the previous case.
For the third case, we examine the performance for the three estimation techniques when the impulse response contains several dispersed scatterers of varying magnitude (both within and outside of the data window).
Obviously many modifications and variations of the present invention are possible in the light of the above teachings. It is therefore to be understood that the scope of the invention should be determined by referring to the following appended claims.
This application is a continuation-in-part of U.S. application Ser. No. 10/673,343 entitled “ROBUST PREDICTIVE DECONVOLUTION SYSTEM AND METHOD”, filed on Sep. 30, 2003, now U.S. Pat. No. 6,940,450 which claims priority from U.S. Provisional Patent Application Ser. No. 60/499,372, filed on Sep. 3, 2003, the disclosures of both of which are hereby incorporated by reference.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4594593 | Nohmi | Jun 1986 | A |
| 4771287 | Mims | Sep 1988 | A |
| 4985704 | Smith | Jan 1991 | A |
| 5059966 | Fujisaka et al. | Oct 1991 | A |
| 5172118 | Peregrim et al. | Dec 1992 | A |
| 5281971 | Moulton | Jan 1994 | A |
| 5323162 | Fujisaka et al. | Jun 1994 | A |
| 5777573 | Klausing et al. | Jul 1998 | A |
| 5805107 | Schruth et al. | Sep 1998 | A |
| 5815111 | Gouenard et al. | Sep 1998 | A |
| 6072419 | Normant | Jun 2000 | A |
| 6603424 | Abatzoglou | Aug 2003 | B1 |
| Number | Date | Country | |
|---|---|---|---|
| 20050057392 A1 | Mar 2005 | US |
| Number | Date | Country | |
|---|---|---|---|
| 60499372 | Sep 2003 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10673343 | Sep 2003 | US |
| Child | 10947784 | US |
