The history of beamforming in sonar applications goes back many years.Perhaps Collodon and Sturm’s use of a horn receiver in their 1826 measurement of the speed of sound in water is a first example.Their 13-km range certainly required a receiver with a good beam pattern to increase the signal to a measureable level.Although receivers have changed over the years, beamforming is still an active area of interest and research.
Modern sonars often utilize multiple sensors in an array configuration.More sensors yield more signal information and can help suppress noise and interference.The key to extracting information from the array is to exploit the fact that the signals arrive at the different sensors at different times due to the array geometry and signal angle.An example of the time delay for a signal incident on two sensors is shown below.
Modelling the signal interaction with the array can be conveniently done using matrices.The measured data,
The mapping involves a series of time delays, which are functionally represented by Δ.The signal is referenced to the first array sensor, and the time delays are also referenced to the first array sensor.It is convenient to call this mapping vector,
The mathematical interpretation of (3) corresponds to what is commonly called delay-and-sum beamforming.This equation forms the foundation for beamforming methods.
An important observation needs to be noted.The model in (1) was developed for only one signal.It is the optimal solution for the case that includes only one signal. However, the one-signal model is often used in multiple-signal environments.Much work has gone into mitigating the effects of other signals on (3) by introducing weights on the sensors that reduce the leakage from other directions or can sometimes put a null in the direction of the interfering signal.However, these methods have their limitations.It is best to use a multiple-signal model in a multiple-signal environment.
2. Narrowband signals
The required time shifts in the mapping matrix have a particularly simple representation for narrowband signals.This is based on the observation that shifting the phase of a narrowband signal approximates a time shift.The resulting equations allow for high resolution beamforming and direction-of-arrival estimation.
2.1. Phase shift approximation
The extra distance, ∆, that the signal,
For narrowband signals it is convenient to express this distance as the radian measure of the fraction of a wavelength extra distance that the signal travels.
This phase angle can then be used to simulate advancing or delaying a narrowband signal by simple multiplication of the phase term.
It should be noted that this is only an approximation in the sense that it does not actually shift the signal in time.Instead it only changes the phase to match a signal shifted in time.This phase-shifted signal still starts and stops at the same time samples as the original signal.So, it is not actually time shifted, but it does approximate a time-shifted signal over part of its interval.This can be problematic for short signal pulses and large time shifts.
2.2 Narrowband beamforming
The general sonar problem involves multiple sensors and multiple signals.This can be expressed as a mapping,
The narrowband mapping or steering matrix,
Other array geometries can be easily accommodated with this approach by correctly modeling the various phase delays associated with the various time delays.
Beamforming requires inverting (8).As before, the Moore-Penrose inverse should be used.This yields the following representation for the signals.
It is important to note that in this representation the
It is worthwhile to understand the mathematical details in (10).The term
A simulation example is shown in Fig. 2 for two overlapping narrowband +20 dB signals arriving on a two-sensor array with half-wavelength spacing at incident angles of ±3º.This example simulates a direct signal plus a multipath signal that reflects off the water’s surface and arrives at a slightly later time.The top plot shows the output using simple beam steering.The two signals are not separated since the beamwidth of this approach cannot differentiate signals with small angular separation.The next two plots show the outputs of the Moore-Penrose beamformer directed at +3º and −3º. The two signals are seen to be clearly separated with this approach.
Sonar data was collected with a synthetic aperture rail system at the NSWC test pool.A short 20-kHz narrow beam signal was projected and two vertical wide beam elements received the signal.Fig. 3 shows the results from simple beam steering and from the Moore-Penrose beamformer output.The vertical axis corresponds to the ping number along the rail length, and the horizontal axis corresponds to the relative time sample.The bottom target is clearly seen between time samples 1100 and 1500 in both images.A multipath signal can also be seen between time samples 2500 and 2900 using simple beam steering, since it is within the array’s mainlobe.However, the output from the two-signal Moore-Penrose beamformer steered towards the bottom target effectively eliminates or cancels the multipath signal.
The separation of the direct path and the multipath signals in Fig. 2 and 3 is interesting since both signals exist deep within the classical mainlobe of the steered beam.The ability to separate the signals comes from using a multiple-signal model.Null beam steering is somewhat similar in the sense that by placing a null at a second signal effectively leads to decoupling. However, null beam steering is limited by its fundamental dependency on a one-signal model, which has no analog to the important (
2.3. Likelihood function
The least-squares method offers a robust high resolution approach for direction-of-arrival estimation.The likelihood function,
It is useful to insert the Moore-Penrose representation for
It is convenient to drop the constant term
This representation can be interpreted as an
It is worthwhile to investigate the properties of this metric tensor for the two-signal and
where.The inverse is then
It can be seen that when
The off-diagonal terms in (15) are particularly interesting.These terms are a measure of the coupling between the signals.Interestingly, as the angle between the two signals becomes small, the off-diagonal terms approach the magnitude of the diagonal terms.
Since the off-diagonal terms can grow to be nearly as large as the diagonal terms when the separation angle is small, they clearly cannot be simply ignored.
There are special cases when the off-diagonal terms can be safely ignored.These special cases occur when the off-diagonal terms are zero, which occur when
Because the off-diagonal terms are zero at these values, the problem naturally decouples and the signals can be completely separated.These zeros correspond to the zeros commonly seen in conventional sidelobe structures.This is the goal of null steering, which adaptively adds weights to the steering matrix to produce nulls in the beam pattern in the direction of an unwanted signal by exploiting the condition in (17) where the off-diagonal terms go to zero.
The likelihood function is commonly expressed using a projection operator and a sample covariance matrix representation.This may be derived by taking the trace of (13) and rotating the vector
The least-squares function for the one-signal case reduces tothe familiar periodogram representation,since (
2.4. Direction-of-arrival estimation
Direction-of-arrival estimation of signals is an important function of array processing.This information can be used for localization applications or used as input for high resolution beamforming applications.The likelihood function approachcan be used to solve this problem by finding the parameters that maximize the function.
A simulation example is shown in Fig. 4 for the same two overlapping narrowband +20 dB signals arriving on an arrayat incident angles of ±3º as seen in Fig. 2.The likelihood function is calculated over the 250 time sample interval.Two cases are considered; the first one with an array of four sensors and the second with eight sensors with half-wavelength spacing.
The peak in the likelihood function corresponds to the best estimate of the directions of arrival.This approach has been shown to approach the Cramer-Rao bound in applications [Stoica 1989].The Cramer-Rao bound has been derived in the literature and is given as
So, the ability to estimate the directions of arrival depends on two terms.The bound is inversely proportional to the signal-to-noise ratio.The bound is also inversely related to the number of sensors in a nonlinear way.Four sensors yield ten-fold better estimates than two sensors.Likewise, eight sensors yield an eight-fold better estimate than four sensors.This increase in resolution with more sensors is seen in the above figure.
An alternative and popular approach to direction-of-arrival estimation is the signal subspace method [Schmidt 1986].This alternative method depends on the eigenvalue structure of the sample covariance matrix.There are a couple of problems with this method.Noise in the sample covariance matrix can lead to poor estimation of the true eigenvalues.Additionally, coherent signals such as seen in multipath lead to problems in the structure of the sample covariance matrix.A number of signal subspace techniques have been developed to address these problems [Wang 1985], [Di Claudio 2001].
2.5. Number of signals resolvable by an array
The number of signals resolvable by an array is somewhat dependent on the array processing methods used.Signal subspace methods are strictly limited to
Both the Moore-Penrose inverse and the likelihood function require the inversion of the term
This resultthat an
3. Time-shift operators
Proper array processing requires shifting the signals in time.Time-shift operators, Δ, are useful to translate a digitized signal by an arbitrary amount of time, τ.These can be functionally expressed as
Unitary matrix operators can be constructed for this task [Piper 2009].Although the specific time-shift matrices developed in this chapter are very useful for solving the general broadband problem, they have numerous applications in other signal processing problems [Laakso 1996].
3.1. Unit time shifts
The simplest time-shift operator is the identity matrix, which shifts the time-series signal by a zero amount.Using matrices with ones along other diagonals, results in time shifts by a discrete number of time samples.It is worthwhile to consider the unit time-shift operator, Δ(1).This may be written in a Toeplitz matrix form as
The effect of this operator is to shift vector elements up or backwards in time by one unit.It should be noted that this simple operator is quite independent of the shape of the waveform it operates on.It is convenient to require these operators to be unitary so that inverses are simply complex transposes.
3.2 Fractional time shifts
Construction of the fractional time-shift operator, Δ(τ), for an arbitrary time shift, τ, is best done in the frequency domain where the fractional time shift can be effected by a phase shift.This operation can be accomplished by multiplying three matrices; the Fourier matrix, F, the phase shift matrix, Δ(
The Fourier, phase shift, and inverse Fourier matrices are listed below
Multiplication of these matrices yields the following exact analytical expression for the fractional time-shift operator
This matrix has a Toeplitz structure.The diagonal terms, Δ
An example of the fractional time-shift operator advancing a unit impulse signal by 2.2 time samples is shown below in Fig. 5.
4. Broadband signals
Broadband signals are much more useful in sonar systems than narrowband signals.The increased bandwidth leads to increased range resolution, and the longer signal duration can lead to higher signal-to-noise ratios.This type of signal is important, and considerable research effort has gone into array processing of broadband data.The key to solving this problem is to use the same functional approach used in the narrowband case but use time-shift operators to perform the necessary time shifting.
4.1. One-signal beamformer
The one-signal model for an
Inversion of this equation in a least-squares sense can be done with the Moore-Penrose inverse.This representation can be interpreted as a simple delay-and-sum approach.
It is possible and somewhat popular to solve the above problem in the frequency domain.This can be realized by inserting Fourier transforms and their inverses into (30). Performing the multiplications reveals that the signal can be constructed in the frequency domain using simple phase shift operators, Δ(
4.2 Two-signal beamformer
For two signals and
The resultant Moore-Penrose inverse solution for the two-signal broadband problem can be functionally written as
Using the above equation with time-shift operators developed in the previous section leads to the matrix to be inverted in (33) being rank deficient.This is a problem.It is therefore necessary to use an
A simulation example of this approach is shown in Fig. 6 for two overlapping +20 dBchirp signals arriving on a two-element arrayat incident angles of ±3º.This example is similar to the one seen in Fig. 2 in that it also simulates a direct signal plus a delayed multipath signal.The top plot shows the measured signal of sensor 1.The next two plots show the two outputs of the Moore-Penrose beamformer directed at +3º and −3º.The two chirp signals can be seen to be clearly separated with this approach.
The two-signal case may be generalized to the multiple-signal case by expanding the signal vector and steering matrix to account for more signals.The inner product metric term,
It may be computationally advantageous to solve this problem in the frequecy domain instead of the time domain.This can be realized by inserting Fourier transforms and their inverses into (34) and multiplying to yield the following frequency domain representation.The Δ(
4.3. Broadband direction-of-arrival estimation
As in the narrowband case, high resolution direction-of-arrival estimates can be obtained using a likelihood function approach.The likelihood function is derived from the least-squares method and is given as
The motivation for this chapter has been the
I would like to thank Prof. Henry Unruh for his advice to “make matrices do what you want them to do.”I would also like to thank Dr. Quyen Huynh for his suggestion to use an approximate inverse.I am indebted to Prof. John Bird for pointing out an error in my work.In addition, I would like to acknowledge partial support for this work by the In-Laboratory Independent Research program funded by the Office of Naval Research.
Stoica P. Nehorai A. 1989MUSIC, Maximum Likelihood, and the Cramer-Rao Bound, IEEE Transactions on Acoustics, Speech, and Signal Processing, 37 5May 1989) 720 741 0096-3518
Schmidt R. 1986Multiple Emitter Location and Signal Estimation, IEEE Transactions on Antennas and Propagation, 34 3March 1986) 276 280 0001-8926X
Wang H. Kaveh M. 1985Coherent Signal-Subspace Processing for the Detection and Estimation of Angles of Arrival of Multiple Wide-band Sources, IEEE Transactions on Acoustics, Speech, and Signal Processing, 33 5August 1985) 823 831 0096-3518
Di Cladio E. Parisi R. 2001WAVES: Weighted Average of Signal Subspaces for Robust Wideband Direction Finding, IEEE Transactions on Signal Processing, 49 10October 2001) 2179 2191 0105-3587X
Friedlander B. Weiss A. 1991On the Number of Signals Whose Directions can be Estimated by an Array, IEEE Transactions on Signal Processing, 39 7July 1991) 1686 1689 0105-3587X
Piper J. 2009Exact and Approximate Time-Shift Operators, Proceedings of SPIE Algorithms for Synthetic Aperture Radar Imagery, 0027-7786X, 7337Orlando, Florida, USA,April 16-17, 2009
Laafso T. Valimaki V. Karjalainen M. Laine U. 1996Splitting the Unit Delay, IEEE Signal Processing Magazine, 13 1January 1996) 30 60 1053-5888