Sei sulla pagina 1di 14

572 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 26, NO.

4, OCTOBER 2001

Signal Detection for Communications in the


Underwater Acoustic Environment
James C. Preisig and Mark P. Johnson

AbstractSignal detection is a critical stage in the implemen- comparison purposes are detailed in Section IV. Section V
tation of any effective communications system. The underwater presents the statistical framework necessary to analyze the
acoustic environment, particularly in the presence of underwater performance of the detection algorithms as well as the results
vehicles, presents significant challenges to reliable detection
without excessive false alarms. While there is often sufficient of the performance analysis. The application of the developed
signal-to-noise ratio with respect to stationary broad-band back- probability distributions to nonideal signal and channel sce-
ground noise to permit reliable operation, the presence of strong narios is described in Section VI. The application to analyzing
event-like interference signals such as narrow-band signals and multi-access performance is described in Section VII. Finally,
impulsive broad-band signals complicates the detection problem the mathematical details of the derivation of the probability
significantly. Frequency-hopped signals interleaved with quiescent
bands are proposed as the basis of a robust detection system. These distributions used in calculating the results in Sections V and
signals also make robust detection possible in a multi-access com- VII are presented in Appendices A and B.
munications system. Two new detection algorithms that exploit
the particular structure of these frequency-hopped signals are
developed and their performance is analyzed. This analysis uses a II. NOISE AND INTERFERENCE IN THE UNDERWATER ACOUSTIC
modification of the doubly noncentral -distribution that has not ENVIRONMENT
been used previously for such analysis. This distribution makes
possible the direct calculation of probabilities of detection and An acoustic communications system, particularly one func-
false-alarm under interference and signal scenarios that cannot tioning in an environment with an autonomous underwater ve-
be analyzed using the better known noncentral -distribution. hicle (AUV), must contend with a number of man-made inter-
With this analysis, the two developed detectors are shown to offer ference sources in addition to the more familiar ambient noise
superior performance to that of either the CFAR detector or the
binary data sequence detector. Experimental data confirms the sources such as wave-generated and biological noise. Signifi-
theoretically derived results. cant interference sources include acoustic navigation systems,
other acoustic communication sources, and thruster and actu-
Index TermsF-distribution, robust detection, underwater
acoustic communication. ator noise. Thruster and actuator noise often contains tonal com-
ponents and therefore introduces narrow-band interference to a
communications receiver on-board the vehicle. Navigation sig-
I. INTRODUCTION nals are usually impulsive in nature and, due to the high signal
levels involved, are often the most troubling of these interfering
T HE underwater acoustic environment poses significant
challenges for reliable signal detection. Significant
event-like interference in the form of both continuous
signals. This is especially true when the navigation signal pro-
jector and the communications receiver are co-located on a ve-
narrow-band and impulsive broad-band signals compromises hicle or mooring.
the performance of matched filter detectors. Two new detection Common underwater navigation systems used with vehicles
algorithms are presented herein that offer robust performance involve short duration (310 ms), high source level (188196 dB
in the presence of these types of interfering signals while re 1 Pa @ 1 m) signals usually transmitted in the 813-kHz
maintaining excellent performance in the presence of stationary or 1827-kHz frequency bands. Frequencies within the band
broad-band background noise. of 454 kHz, which encompasses the operating bands of most
The organization of this paper is as follows. Section II underwater acoustic communication systems, are within an oc-
contains a brief description of interference sources and channel tave of these two common navigation frequency bands. There-
characteristics in the underwater acoustic environment. The fore, these navigation signals are not strongly rejected by the
details of a frequency-hopped signaling format that is well analog input filters typically found in acoustic communication
matched to the proposed detectors is presented in Section III. receivers. In comparison, the received level of a communica-
The two proposed detection algorithms as well as two pre- tions signal at 5-km range may be on the order of 110 dB. As
viously published approaches that are analyzed here for a consequence, overload of the receiver analog-to-digital con-
verter is inevitable during navigation transmissions where the
communication and navigation functions are co-located. In the
Manuscript received May 15, 2000; revised July 19, 2001. This work was
supported by the Office of Naval Research. This paper is WHOI Contribution case where they are not co-located, the receive level of the nav-
Number 10193. igation signals may still be equal to or greater than that of the
The authors are with the Department of Applied Ocean Physics and Engi- communication signals. A solution to this problem is to develop
neering, Woods Hole Oceanographic Institution, Woods Hole, MA 02543 USA
(e-mail: jpreisig@whoi.edu). detection, synchronization, and data modulation methodologies
Publisher Item Identifier S 0364-9059(01)09912-5. that are robust to these types of interfering signals. This work
03649059/01$10.00 2001 IEEE
PREISIG AND JOHNSON: SIGNAL DETECTION FOR COMMUNICATIONS IN THE UNDERWATER ACOUSTIC ENVIRONMENT 573

focuses on the development and analysis of a robust detection


methodology.

III. FREQUENCY-HOPPED SIGNALING FOR DETECTION


The limit on realizable detection performance is fundamen-
tally set by the choice of the signal transmitted for detection pur-
poses. The performance of a linear matched-filter detector in the
presence of white noise is completely determined by the energy
in the transmitted signal and the spectral density of the noise
[1], [2]. However, other factors must be considered in more com-
plex interference environments or when there are additional per-
formance requirements. For example, signals employing either
linear frequency modulation (chirp) or phase-shift-keyed (psk)
Fig. 1. Receiver preprocessing.
signals modulated by data sequences with good auto-correla-
tion properties (e.g., the Barker code, the Turin code, maximum
length sequences) are often used for both detection and synchro- used for comparison with the measurements of the energy in the
nization. These signals offer good detection and synchroniza- th signal subband as and let the cardinality of this set equal
tion performance with matched-filter detectors. However, while . That is,
the matched-filter detector is optimal for detecting a known
signal in the presence of white Gaussian noise of known spec-
tral density, its performance can suffer when the interference vi- Note that it is possible that a single quiescent subband may be
olates this assumption [3]. The interfering signals described in used for estimating the background noise level for use in com-
Section II are examples of signals that can compromise the per- parison with more than one signal subband. The tradeoffs asso-
formance of a matched-filter detector operating on simple chirp ciated with the selection of as well as the frequency of re-use
or psk signals. Frequency-hopped signals are a class of signals of the quiescent subbands are discussed in Section V-B.
that make robust detection possible in complex propagation and The choice of the signal subbands as well as the selection of
interference environments. In addition, these signals effectively the order of transmission in these subbands provides a method
accommodate multiple users in a multi-access communications of differentiating the detection signals for different users in a
system as shown in Section VII. multi-access communications system. A method of analysis of
Frequency-hopped signalling makes use of the available fre- the performance of the detection algorithms in the multi-access
quency bandwidth by dividing it into subbands with the scenario is described in Section VII.
center frequency of the th band denoted by . The frequency-
hopped detection signal is made up of pulses where IV. SIGNAL DETECTION ALGORITHMS
. The index of the center frequency of the transmitted pulse
during the th symbol period is denoted by . Each pulse has The signal detection algorithms described herein share a
length . Thus, the frequency-hopped signal can be expressed common structure with most detection algorithms. That is,
as each uses the received signal to create a detection metric that is
compared with a threshold to select between the two possible
hypotheses, indicating no signal present and indicating
for (1) a signal present [1], [2]. The detection algorithms differ in the
manner in which the metric is calculated.
Here, is the transmit pulse shape resulting from user-se- It is assumed that signals are initially filtered and sampled
lected time-domain windowing as well as the effect of transmit at a rate . The sampled signal is then beat to baseband for
electronics. The function returns the real part of its argu- each of the subband carrier frequencies and passed through an
ment. The subbands in which the detection signal is transmitted appropriate lowpass filter, . The magnitude squared of
are referred to as the signal subbands. and are usually se- the complex baseband signals , denoted by for
lected so where is the delay spread of the , are the raw signals to be processed by the detection
channel impulse response. This selection ensures that succes- algorithms. The signal preprocessing is shown in Fig. 1 where
sive pulses transmitted in the same subband will not interfere . Without any loss of generality, it is assumed here
with one another. that the signals are downsampled following the filtering opera-
Note that , which implies that some of the subbands tion shown in Fig. 1 to a baseband rate, for some
do not contain a portion of the transmitted frequency-hopped integer .
signal. These signal-free subbands allow detection algorithms In a practical implementation, the multiband demodulation
to estimate background noise levels and/or provide standards and decimation processing could be implemented efficiently
for comparison with the detected energy in the signal subbands. using a multirate technique. However, the detailed design of
The signal-free subbands will be referred to as the quiescent the preprocessing algorithm is beyond the scope this paper. The
subbands. Denote the set of indices of the center frequencies performance analysis in Section V assumes that the filter
(i.e., the subband numbers) of the subset of quiescent subbands is constructed so that a signal transmitted in the th subband
574 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 26, NO. 4, OCTOBER 2001

does not affect the filtered outputs or not a signal is present, thus increasing the probability of false
for any . The effect of subband crosstalk resulting alarm for any given detector threshold. Finally, if the contami-
from either nonideal filter design or from environmental effects nated subband is one of the quiescent subbands, then ap-
is discussed in Section VI-A. proaches zero as the interfering signal gets large. This happens
whether or not a signal is present, thus reducing the probability
A. The Classical Approach of detection for any given detector threshold.
The censored mean-level detector was developed to eliminate
The classical approach to detection in the presence of un- the performance degradation associated with interfering signals
known noise levels is commonly referred to as the Constant in the context of target detection with radar [4], [5]. In the con-
False Alarm Rate (CFAR) detector [1]. This detector is also re- text of the frequency-hopped signaling considered herein, the
ferred to as the mean-level detector [4]. In the context of de- censored mean-level detector performs a rank ordering of the
tecting the frequency hopped signal described in Section III in energy measurements in the quiescent subbands and eliminates
the presence of white Gaussian noise, the detection metric is the largest measurements from the sum in the denominator
given by of (2) for a selected value of . While this technique would
be effective in eliminating the detector performance degrada-
tion in the third example given previously in this section (single
narrow-band interferer in a quiescent subband), it would do
(2) nothing to restore acceptable performance in the first two ex-
amples. In contrast, the detectors detailed in the Section V will
be shown to maintain effective performance in all of the inter-
While this detector is a CFAR detector only when the inter- ference scenarios described previously in this section.
ference is additive white Gaussian noise, it will be referred to
as the CFAR detector in the discussion of detector performance B. Robust Detectors for the Underwater Environment
for all interference scenarios considered herein. The shifting in
The approach for developing a robust detector is to calculate
time of the outputs for each of the frequency bands in com-
a ratio of signal band to quiescent band energy for each symbol
puting the metric is necessary to time-align the signals trans-
(tone) period, perform a nonlinear transformation on this ratio
mitted in each subband as described in (1). This shifting in time
to confine it to the range of zero to one, and then average the re-
is common to all of the detectors described herein.
sult over all symbol periods. Intuitively, by computing the ratio
Under the assumptions that the baseband background noise individually for each symbol period, the effect of an impulsive
is white complex Gaussian noise and that and or narrow-band interferer is restricted to a small subset of the
are independent random variables for or ratio calculations. The nonlinear transformation prevents a ratio
, the pdf of is the -distribution. Hence, the notation whose computation uses the energy estimate from one or more
. The division by 2 in both the numerator and denominator contaminated subbands from dominating the calculation of the
in (2) results from the fact that each variable is the sum of average ratio over all of the symbol periods.
the magnitude squared of the real and imaginary parts of the The starting point for computing the detection metrics is to es-
complex baseband signal in each subband. If the numerator is timate the energy in the signal and quiescent bands. Let
considered to be an estimate of the signal plus noise energy in and . The quantity
the signal subbands and the denominator is considered to be an is an estimate of the signal energy plus noise power in a signal
estimate of the noise energy common to all subbands, then subband while is an estimate of the noise power in each
is an estimate of the ratio of the signal plus noise energy to noise quiescent subband. The energy ratio for each symbol period is
energy. given by
In the presence of nonstationary impulsive or narrow-band
interference, the performance of the CFAR detector can de-
grade significantly. The following simplified examples, which (3)
are the limiting cases of the interference scenarios described
in Section II, illustrate this performance degradation. Consider The proposed detection metrics can each be expressed as
the case of an impulsive interfering signal contaminating all
subbands with an equal amount of energy at just one point in
time (i.e., is contaminated for all and one ). In the (4)
limit as the interfering energy gets large, will approach
one for . This limit will be
approached whether or not a signal is present, thus reducing the where
probability of detection for any given detector threshold greater
than one.
Next consider the case of a single narrow-band interferer con- (5)
taminating just one subband. If this subband is a signal subband
(i.e., for some ), then gets large as the and the function is the specific memoryless nonlinearity as-
energy in the interfering signal gets large. This happens whether sociated with each detector.
PREISIG AND JOHNSON: SIGNAL DETECTION FOR COMMUNICATIONS IN THE UNDERWATER ACOUSTIC ENVIRONMENT 575

(a) (b)
Fig. 2. Theoretical ROC curves with no impulsive or narrow-band interference for (a) C = 3, and (b) C = 1.

Three detectors are detailed here. The first detector is referred Thus, this detection metric is the ratio of the energy in the signal
to as the type detector with the detection metric denoted by subband to a weighted sum of the energy in the signal subband
. The memoryless nonlinearity for this detector is and the average energy in the quiescent subbands. The interpre-
tation of this detector is of particular interest for and
(6) . In the former case, (9) reduces to

is always nonnegative so this function is bounded between


zero and one. Substituting (3) into (6) yields
This function is referred to as in [6] where its pdf is derived
for the special case of no interfering signal in any of the quies-
cent subbands. For the case of , (9) reduces to
(7)

The intuition behind this nonlinearity is that, under the assump-


tion of equal noise power in each subband, the numerator is an
estimate of the signal energy in the signal subband. Thus, the re- This function is referred to as in [7] but its probability dis-
sult is an estimate of the signal to signal-plus-noise ratio (SNR). tribution is not derived therein.
Hardlimiting the ratio at zero prevents measurements that vio- The final detector examined here is the binary data detector.
late the assumption of equal noise power in each subband (e.g., It is presented here only for comparison purposes as the limiting
a large narrow-band interferer in one or more of the quiescent case of the memoryless nonlinearity taking on only the values
subbands) from overly influencing the result. Note that the pres- of 0 and 1. The detection metric in this case is denoted by
ence of a large narrow-band interferer in the signal band would and the memoryless nonlinearity is
drive the result to one while a large impulsive interferer that af-
fects all bands equally would drive the result to zero.
The second detector is referred to as the type detector with
the detection metric denoted by . The memoryless nonlin- (10)
earity for this detector is
Substituting (3) into (10) yields
(8)

where is a design parameter greater than zero. Note that, for


, this function is bounded between zero and one. Substi- (11)
tuting (3) into (8) yields
For the case of , (11) is a binary fsk decision device.
In Section V, the performance of each detector detailed in this
(9)
section is analyzed under different interference scenarios.
576 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 26, NO. 4, OCTOBER 2001

(a) (b)

(c) (d)
Fig. 3. Theoretical detector performance with impulsive interference for (a) P with C = 3, (b) P with C = 3, (c) P with C = 1, and (d) P with C = 1.
Note that the curves for the two type detectors in (a) and (b) are virtually identical and indistinguishable.

V. DETECTION PERFORMANCE ANALYSIS noise. The deterministic signals can be any one of three types.
These are the transmitted frequency-hopped detection signal,
Each of the detection algorithms is evaluated here using two
an impulsive interference signal, and a narrow-band interfer-
criteria. The first is the receiver operating characteristic (ROC)
ence signal. Thus, the complex baseband signals,
curve under nominal operating conditions. The second is the
, are complex Gaussian random variables with inde-
change in operating characteristics in the presence of impulsive
pendent real and imaginary parts. Let and denote the
and narrow-band interference. This section details the evalua-
mean and variance, respectively, of . Note that the means
tion of each algorithm and is divided into three parts. In the
will vary across the bands and with time according to the pres-
first, the statistical model upon which the theoretical perfor-
ence or absence of the deterministic signals. The variance is
mance analysis is based is detailed. In the second, the results
independent of band and time under the assumption of white
of the theoretical performance analysis using the above criteria
background noise. Then is a chi-squared or noncentral
are presented and discussed. These theoretical results are calcu-
chi-squared random variable with 2 degrees of freedom
lated using the probability distributions derived in Appendices
and a noncentrality parameter of . See Ap-
A and B. Finally, experimental data is presented supporting the
pendix A for a discussion of the parameters associated with the
theoretical results.
statistical characterization of chi-squared random variables.
It is assumed here that the effective bandwidth of the subband
A. The Statistical Model filter, , as shown in Fig. 1, is no greater than the spacing
The received signal is assumed to be the sum of de- between any two carrier frequencies, and . Under this as-
terministic signals plus zero-mean white Gaussian background sumption, and are independent random variables
PREISIG AND JOHNSON: SIGNAL DETECTION FOR COMMUNICATIONS IN THE UNDERWATER ACOUSTIC ENVIRONMENT 577

(a) (b)

(c) (d)
Fig. 4. Theoretical detector performance with narrow-band interference in a quiescent subband for (a) P with C = 3, (b) P with C = 3, (c) P with C = 1,
and (d) P with C = 1.

for and any and . A second assumption is that the a detection signal is present its contribution is the same in all
duration of the impulse response of is no greater than signal subbands (i.e., the channel has a flat frequency response).
samples (i.e., the number of samples per symbol period). Under A method for handling the effects of independent fading in the
this assumption, and are independent random vari- signal subbands is described in Section VI-B.
ables for .
Two further assumptions are made regarding the various de- B. Theoretical Performance Comparisons
terministic signals that contribute to the complex baseband sig- Theoretical performance measures for the detectors described
nals, . Let , , and de- in Section IV are presented for two signaling examples. In both
note the contribution to of the frequency-hopped detection cases, a system with 34 subbands of which one half are signal
signal, impulsive interfering signal, and narrow-band interfering subbands is considered. The signal parameters as defined in
signal, respectively. Then, it is assumed that Sections III and V-A are , ,
(i.e., the in-band SNR equals 3 dB with respect to stationary
(12) broad-band background noise), , and
. When the impulsive interfering signal is
That is, the contributions of these three signals to are present, it is assumed to be present for only one symbol period.
orthogonal. This assumption is common to the analysis of all However, the impulsive interfering signal contaminates all sub-
of the detectors so it does not affect the evaluation of their bands during this one symbol period. When the narrow-band
relative performance. Finally, it is assumed that interfering signal is present, it is assumed to contaminate only
. The implication of this is that when one frequency subband. However, the narrow-band interfering
578 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 26, NO. 4, OCTOBER 2001

(a) (b)

(c) (d)
Fig. 5. Theoretical detector performance with narrow-band interference in a signal subband for (a) P with C= 3, (b) P with C = 3, (c) P with C = 1,
and (d) P with C = 1. Note that the curves for the type I and type ( = 1) detectors in (b) and (d) are virtually identical and indistinguishable.

signal contaminates this frequency subband for all symbol signal in a quiescent subband (Fig. 4), and a narrow-band in-
periods. The two examples are distinguished by the number of terfering signal in a signal subband (Fig. 5). Parts (a) and (c) of
quiescent subbands used in computing the detection metrics: each of these figures show the actual probability of detection,
for the first example and for the second example. , as a function of the design and parts (b) and (d) show the
The results are shown in Figs. 25. Note that the probability of actual as a function of the design . The design is the
false alarm, , is plotted on a log scale in all of these figures probability of false alarm that the detector would have achieved
as well as in Fig. 7. if no interfering signal were present. The use of this quantity as
Fig. 2 shows the ROC curves for the two examples when no the independent axis in these figures gives a common indepen-
interfering signals are present. That is, the received signal con- dent axis with Fig. 2, allowing direct independent evaluation of
sists of the transmitted signal plus zero-mean white Gaussian the effect of the interfering signals on and .
background noise. The performance is comparable for all de- These results all show a significant decline in the performance
tectors except for the binary data detector, whose performance of the CFAR detector when compared to both the type and
is worse than that of the other detectors. The performance of all type detectors where such a decline is defined as a signifi-
detectors is better when than it is for , illustrating cant rise in the actual or a significant drop in the actual
the benefit of the statistical averaging in estimating the back- for a given design . Note that, in the case of the narrow-band
ground noise level. interferer in a signal subband, increases for all detectors at
The remaining figures show the performance results under any given threshold level and that the increase is most dramatic
different interference scenarios. The scenarios shown are an im- for the CFAR detector. However, the actual also increases
pulsive interfering signal (Fig. 3), a narrow-band interfering for all detectors and the increase is again most dramatic for the
PREISIG AND JOHNSON: SIGNAL DETECTION FOR COMMUNICATIONS IN THE UNDERWATER ACOUSTIC ENVIRONMENT 579

Fig. 7. ROC curve for experimental data.


Fig. 6. Memoryless nonlinearity G for different detectors.
used in computing for three different symbol periods, , in
CFAR detector. Therefore, while all detectors will maintain an (3). Thus, three energy ratios are contaminated by the interfering
adequate for a given design , the actual is unaccept- signal. However, at the SNR used here, the benefit of improved
ably high for the CFAR detector at the same design . statistical averaging for outweighs the disadvantage of
The performance decline of the binary data detector in the the contaminated quiescent subband being used in calculating
presence of interfering signals largely mirrors that of the type three of the energy ratios used in (4) and (5).
and type detectors. Thus, given the inferior performance of
the binary data detector when no interfering signal is present, C. Experimental Data
the performance of the binary data detector when interfering Experimental data was collected in Woods Hole harbor to
signals are present remains inferior to that of the type and type test the performance of each detector on a common set of re-
detectors. ceived signals. The frequency-hopped signals were transmitted
In general, the performance of the type detector is superior in the 1116-kHz frequency band with , ,
to that of the type detector, and, as noted previously, the per- and msec. Broad-band noise was transmitted simulta-
formance of the type and detectors is superior to that of the neously with the frequency-hopped signals at a level such that
binary data detector. Intuitively, this ordering of detector perfor- (i.e., the in-band SNR equaled 3 dB).
mance can be explained by the information lost in the evaluation The ambient noise in the harbor is quite variable and often
of by the type and binary data detectors. This information at a high level. Significant sources of noise in the harbor in-
loss is illustrated in Fig. 6 which shows the function for clude outboard motors, propulsion and machinery noise on large
each of the detectors. Note that for the functions and ships, and acoustic depth sounders. These sources combined to
both equal zero and are noninvertible. This represents a loss drop the SNR below 3 dB at times and added some impulsive
of information by the processor. For , equals one and is interference to the received signal. For approximately one half
therefore noninvertible over this range again representing a loss of the duration of the test transmissions, other acoustic equip-
of information. The binary data detector represents the limiting ment including broad-band tomography sources and subbottom
case of this information loss during the nonlinear transforma- profilers were also being used near the test site. These sources
tion in that only one bit of information ( or ) is added additional high-level interference including impulsive,
available after the transformation. This information loss results continuous broad-band, and chirp signals. Thus, the data col-
in the relatively poor performance of the binary data detector lected yields a good measure of detector performance under a
when compared to the other detectors. In contrast, the function variety of realistic interference conditions.
is invertible over the entire range of thus imposing no loss The ROC curves derived from the received data are shown in
information loss. It should be noted though that the practical Fig. 7. For these results, . is calculated using 3489
constraint of finite precision arithmetic results in an information data points and is calculated using 16 784 data points. The
loss for all the detectors. However, the evaluation of the effect results confirm the superior performance of the type and type
of such finite precision on the information loss of each function detectors when compared to both the CFAR and binary data
is beyond scope of this paper. detectors. The experimental results for the type , type , and bi-
It is instructive to observe that the detectors for which nary data detectors closely match the theoretically predicted re-
maintain superior performance compared to those for which sults for these detectors for the case of no interference, as shown
when there is a narrow-band interferer in a quiescent in Fig. 2(a). This indicates that these detectors are robust with
subband. Note that, when , each quiescent subband is respect to the interference sources encountered during the tests.
580 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 26, NO. 4, OCTOBER 2001

The performance of the CFAR detector falls short of that pre- Let denote the pdf of the noncentrality parameter
dicted in Fig. 2(a) confirming its performance degradation in the for each of the signal subbands. This pdf can be derived from
presence of the interference sources encountered during the test the Rayleigh or other statistical channel assumption. Then,
transmissions. assuming independent fading in each signal subband, the
effects of random channel fading on each of the detectors
analyzed here with the exception of the CFAR detector can be
VI. DEPARTURES FROM IDEAL SYSTEM ASSUMPTIONS
accommodated by replacing with
The theoretical performance analysis in Section V assumes
ideal system characteristics. The first of these assumptions is (13)
that the transmission of a signal in one subband does not af-
fect the filter output in any other subband. This corresponds
in the calculation of the pdf of each detector as detailed in
to subband orthogonality with no Doppler shift or spread in
Appendix B. Here, denotes the pdf of under the fading
the received signal. The second of these assumptions is that
channel assumption. In calculating the pdf of the CFAR
the acoustic channel is not delay spread. That is, the impulse
detector, the noncentrality parameter is the sum of the indepen-
response of the channel is an impulse. Fortunately, departures
dent noncentrality parameters for each signal subband. Thus,
from these assumptions can be analyzed theoretically using the
the pdf of the noncentrality parameter used in (13) will be the
doubly noncentral -distribution. The analysis exploits the fact
that the pdf is conditioned upon and (17) and is detailed convolution of the pdfs of the noncentrality parameter for each
in Section VI-A and VI-B. of the subbands.
Another issue that arises in the delay spread channel is the im-
pact of signal arrivals spread by a delay of greater than . For
A. Subband Crosstalk
the single-user case, these multiple arrivals increase since
Vehicle motion resulting in Doppler shifts of transmit signal the detector receives multiple noninterfering repetitions of the
bands as well as design choices in transmit pulse shaping , frequency hopped detection signal and may detect any one of
subband filtering , and subband spacing can all cause sig- them. This may impact proper synchronization depending on
nals transmitted in one subband to affect the filter outputs in the synchronization approach used by the receiver. However, the
other signal or quiescent subbands. In addition, these events can analysis of the synchronization issues is not considered here. In
result in a reduction in the received signal energy in the intended the multi-user case, these multiple arrivals can have the adverse
signal subband. The effect of either reduced subband signal en- impact of increasing or decreasing as described in Sec-
ergy or subband crosstalk on detection performance can be an- tion VII.
alyzed in a deterministic manner using the doubly noncentral
-distribution by appropriate adjustment of the noncentrality VII. MULTI-ACCESS DETECTION PERFORMANCE
parameter for the affected subbands. Consider the case of sub-
Frequency-hopped signaling can be used for detection pur-
band filtering that results in energy transmitted in one signal
poses in a multi-access communications system as well. Dif-
subband showing up in the output of a filter for a quiescent
ferent assigned signal subbands and order of transmission in the
subband or another signal subband. Such a case may occur if
subbands can be used to differentiate the signals from different
FFT-based processing restricts the receiver subband center fre-
users. The number of collisions between the signals from dif-
quencies and bandwidths to values that do not exactly match
ferent users is the factor that determines the effect of the signal
those of the transmitted signals. In this case, the noncentrality
from one user on the performance of an algorithm in detecting
parameter for the transmit subband is decreased by the amount
the signal from another user. Here, a collision is an incident
of energy that has leaked out of the band. In addition, the non-
of the signal from an interfering user occupying either a signal
centrality parameter for the signal or quiescent subbands into
or quiescent subband of a desired user during a symbol period
which the energy has leaked is increased be the amount of this
during which a detection algorithm would be using the energy
energy. This adjustment is made for only the affected subbands.
in that subband to calculate a detection metric. In this paper the
These modified noncentrality parameters are then used in the
impact of such collisions on detector performance quantified by
same manner as described in more detail in Section VII to cal-
a change in or is analyzed. The reader interested in de-
culate the impact on and .
termining how to minimize collisions by the choice of signal
and quiescent subbands or their order of transmission is referred
B. The Delay Spread Channel to [8].
The calculations in Section V assume that the magnitude The effect of interfering signals can be modeled deterministi-
squared of the received frequency-hopped detection signal is the cally by adjusting values of the noncentrality parameters associ-
same in each signal subband. However, subbands spaced fur- ated with one users received signals to account for the presence
ther apart in frequency than approximately Hz of interfering frequency-hopped signals in that users signal or
can exhibit independent fading behavior in realistic channels. quiescent subbands during some symbol periods. This is similar
Under the common assumption that the magnitude of the to the approach used in Section V to model narrow-band inter-
impulse response of the channel is either a Rayleigh or Ricean ference.
random variable, then is a chi-squared or noncentral As an example, consider the case where there are an equal
chi-squared random variable. number of signal and quiescent subbands and . Then
PREISIG AND JOHNSON: SIGNAL DETECTION FOR COMMUNICATIONS IN THE UNDERWATER ACOUSTIC ENVIRONMENT 581

(a) (b)

(c) (d)

(e) (f)
Fig. 8. Theoretical detector performance with multiaccess interference for: (a) P for case 1; (b) P for case 1; (c) P for case 2; (d) P for case 2; (e) P for
case 3; and (f) P for case 3. Note that the curves for the type I and type detectors in (d) are indistinguishable.
582 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 26, NO. 4, OCTOBER 2001

each signal subband is paired with a single quiescent subband. interfering signal being present in the users signal subband.
When is a prime number and there is a delay spread of The integration in (13) must also be carried out over to
symbol periods, it is possible to create sets of order- account for interfering signals in the desired users quiescent
ings of the signal and quiescent subbands pairs so that there are subbands.
only collisions between the signals assigned to any two
users. Let and as was the case for the anal- VIII. CONCLUSION
ysis in Section V-B. Assume that there are six collisions be-
The wide variety of interfering signals in the underwater
tween the signal from an interfering user with the signal from
acoustic environment results in poor detection performance
a desired user for which the detection performance is analyzed.
of both traditional matched filter and CFAR algorithms. A
This is a worst-case number of collisions for a channel in which
frequency-hopped signaling format is outlined here that uses
. Assume also that the received signal power is the
both signal and quiescent subbands. This signaling format is
same for each user in each signal subband and that the energy
exploited by two detection algorithms proposed here. The two
contributions of the signals from each user add orthogonally.
detection algorithms are the type and type detectors and
Three cases are considered here. In Case 1, the signal subband provide superior performance when significant interferers are
of each user is a quiescent subband for the other user for half both present and not present. The algorithms use independent
of the collisions while the signal subband for each user is also a calculation of energy ratios and nonlinear transformations of
signal subband for the other user for the remaining collisions. In the ratios followed by incoherent averaging to achieve robust
Case 2, the signal subband for each user is also a signal subband performance with respect to interfering signals.
for the other user for all collisions. In Case 3, the signal subband The performance of these two detectors is compared to that of
for each user is a quiescent subband for the other user for all the CFAR and binary data detectors. A new probability distri-
collisions. bution, the doubly noncentral -distribution, is presented and
Qualitatively, the collisions resulting in the signal from the used to calculate the probability distributions of the detectors
interfering user falling in the signal subband of the desired user and to theoretically analyze their performance under a number
will increase both and in a manner similar to the effect of of interference scenarios. This analysis predicts the superior
the narrow-band interference in a signal subband. The collisions performance of the type and type detectors. Experimental
resulting in the signal from the interfering user falling in the qui- data confirms this superior performance. The doubly noncen-
escent subband of the desired user will decrease both and tral -distribution presented here provides a powerful tool to
in a manner similar to the effect of the narrow-band inter- analyze the effect of many different environments and design
ference in a quiescent subband. A quantitative analysis requires tradeoffs on the performance of these detection algorithms.
modification of the probability distribution functions used in the
calculation of and . APPENDIX A
Under the assumptions detailed earlier in this section, the THE DOUBLY NONCENTRAL -DISTRIBUTION
noncentrality parameters of the pdfs of for six affected
symbol periods are modified. These pdfs are then used along The probability distribution required to calculate the ROC
with the pdfs of the 11 unaffected symbol periods to cal- curve for each of the detectors analyzed herein under the in-
culate the detector performance metric as described in the terference scenarios of interest is the doubly noncentral -dis-
introduction to Appendix B. For the symbol periods during tribution. The statistical assumptions that make this distribution
which the interfering signal falls in a signal subband, a term necessary are detailed in Section V. To relate the doubly noncen-
representing the energy contribution of the interfering signal to tral -distribution to the statistical characteristics of the random
the appropriate signal subband is added to the sum in (12). For variables of interest, let be a sequence of indepen-
the symbol periods during which the interfering signal falls in dent Gaussian variables with
a quiescent subband, a similar adjustment to the sum in (12)
is made for the appropriate quiescent subband. The changes
to and calculated using this technique for the signal Let
parameters detailed earlier are shown in Fig. 8. As expected,
Case 2 [Fig. 8(c) and (d)] shows an increase in both the actual (14)
and relative to their nominal values for a given decision
threshold [Fig. 2(b)]. Also as expected, Case 3 [Fig. 8(e) and
and
(f)] shows a decrease in both the actual and relative
to their nominal values for a given decision threshold. Case 1
[Fig. 8(a) and (b)], which has three collisions with the interferer
signal subband occupying a desired user quiescent subband and
three collisions with the interferer signal subband occupying a Then, for the case that , is a chi-squared random vari-
desired user signal subband, shows a small decrease in both able with degrees of freedom. If , then is a noncen-
and relative to their nominal values. tral chi-squared random variable with degrees of freedom and
The effect of the interfering signals can also be modeled a noncentrality parameter .
stochastically in the style of (13) where is modified If the nonzero means of the Gaussian random variables in the
to account for both channel fading and the probability of an sum in (14) represent the signal contribution to the random vari-
PREISIG AND JOHNSON: SIGNAL DETECTION FOR COMMUNICATIONS IN THE UNDERWATER ACOUSTIC ENVIRONMENT 583

able , then the noncentrality parameter has the interpretation Two identities [10] are useful in simplifying the computation
of a scaled signal to noise ratio. Note also that the sum of in- of The first is
dependent chi-squared random variables is also a chi-squared
random variable. Therefore, under the statistical assumptions (18)
outlined in Section V, the quantities and defined in
Section IV are themselves chi-squared random variables. In ad- where is the Beta function and is the Gamma
dition, the numerator and denominator used in the calculation function. The second useful identity is
of the CFAR detection metric (2) are both chi-squared random
variables. (19)
Assume that and are independent chi-squared random
variables with degrees of freedom and and noncentrality for nonnegative integer values of .
parameters and , respectively. Then the ratio Under the assumptions in Section V, is a complex
Gaussian random variable with independent real and imag-
inary parts. Therefore, is a chi-squared or noncentral
(15) chi-squared random variable with 2 degrees of freedom.
Furthermore, and are independent for . As
a result, the number of degrees of freedom of the chi-squared
is said to follow an -distribution for the case that and (or noncentral chi-squared) random variables and
. If and , in (15) is usually denoted by is always an even integer. Thus the coefficients of the Beta
and is said to follow a noncentral -distribution [9]. For the function in are always integers and the identity in (19) is
case of most general interest here, that is and , applicable.
in (15) will be denoted by . A term for the distribution of has At , (17) reduces to
not be found in the literature so herein it will be referred to as
the doubly noncentral -distribution. Note that if the variance
is the same for the Gaussian random variables included in the
sums comprising both and , then the division by can be
eliminated from the definition in (14) without affecting the ratio
in (15). Hence, the computation of the random variables
described in Section IV does not include a division by the noise The case of is not of concern here since, for the reasons
variance. stated in the preceding paragraph, is always an even integer.
The quantity is related to the quantity as defined in [6] by
the simple scaling . The pdf of is derived in [6] APPENDIX B
and will not be repeated herein. It is straight forward to show PROBABILITY DISTRIBUTIONS OF THE DETECTION METRICS
that for any two random variables, and , related by
for a positive constant , the pdfs of and are related by The probability distributions of the detection metrics detailed
in Section IV can all be represented as functions of the doubly
noncentral -distribution, Clearly, the CFAR detection
metric defined in (2) has as its pdf. The derivation of the
(16)
distributions for the remaining detection metrics is detailed
below.
Using (16) the pdf of in [6] can be manipulated to yield the Using (4) and (5) as the starting points, the probability distri-
pdf of bution of will be derived for each of the detectors. These distri-
butions can then be used to calculate the probability distribution
of for each detector. Under the assumptions in Section V, the
terms and are inde-
pendent random variables for . Therefore, the probability
distribution of the detection metric is the convolution of the
probability distribution of each term in the sum in (4). Appro-
priate values of the noncentrality parameters must be made for
each pdf used in the convolution. The noncentrality parameters
for each term in (4) will depend on the whether or not a signal,
impulsive interferer, or narrow-band interferer is present in the
signal band and symbol period in question.
In the derivations that follow, the dependence of the variable
(17) on and is suppressed for notational
convenience. In addition, the following notation will be used.
where is the Beta function. If and , then will denote the pmf of (i.e., the probability that the
(17) reduces to the -distribution. If and , then ) and denote the cdf of (i.e., the probability
(17) reduces to the noncentral -distribution. that ).
584 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 26, NO. 4, OCTOBER 2001

A. Detector B. Detector
Combining (5) and (6) yields Combining (5) and (8) yields

Differentiating with respect to yields


(20)

From (20), it is clear that

(21)

(24)
Note that for , and for , (20)
can be rearranged to yield
At , (24) reduces to

Differentiating with respect to yields

The case of is not of concern here since for the reasons


stated in Appendix A, is always an even integer.

(22) C. Detector
From (5) and (10), it is clear that
where is the delta function and the convention

(25)

REFERENCES
is assumed. Evaluating (22) and combining with (21) yields the
[1] C. W. Helstrom, Elements of Signal Detection and Estima-
desired result of tion. Englewood Cliffs, NJ: Prentice-Hall, 1995.
[2] H. L. Van Trees, Detection, Estimation, and Modulation Theory, Part
I. New York, NY: Wiley, 1968.
[3] S. A. Kassam and H. V. Poor, Robust techniques for signal processing:
A survey, Proc. IEEE, vol. 73, pp. 433477, Mar. 1985.
(23) [4] J. T. Rickard and G. M. Dillard, Adaptive detection algorithms for
multiple-target situations, IEEE Trans. Aerosp. Electron. Syst., vol.
AES-13, pp. 338343, July 1977.
where the dependence of on , , , and has been [5] J. A. Ritcey, Performance analysis of the censored mean-level de-
suppressed for notational convenience. tector, IEEE Trans. Aerosp. Electron. Syst., vol. AES-22, pp. 443454,
July 1986.
At , (23) reduces to [6] P. C. Tang, The power function of the analysis of variance tests with
tables and illustrations of their use, Statistic. Res. Memoirs, vol. 2, pp.
126149, 1938.
[7] P. B. Patnaik, The noncentral  - and F -distributions and their appli-
cations, Biometrika, vol. 36, pp. 202232, 1949.
[8] P. Fan and M. Darnell, Sequence Design for Communications Applica-
tions. Hertfordshire, U.K.: Research Studies Press, 1996.
[9] G. H. Robertson, Computation of the noncentral F distribution
(CFAR) detection, IEEE Trans. Aerosp. Electron. Syst., vol. AES-12,
pp. 568571, Sept. 1976.
[10] M. Abramowitz and I. Stegun, Handbook of Mathematical Functions
With Formulas, Graphs, and Mathematical Tables . Aliso Viejo, CA:
The case of is not of concern here since for the reasons National Bureau of Standards Applied Mathematics Series, 1964, vol.
stated in Appendix A, is always an even integer. 55.
PREISIG AND JOHNSON: SIGNAL DETECTION FOR COMMUNICATIONS IN THE UNDERWATER ACOUSTIC ENVIRONMENT 585

James C. Preisig received the B.S. degree in Mark P. Johnson received the Ph.D. degree in electrical engineering from the
electrical engineering from the United States Coast University of Auckland, Auckland, New Zealand, in 1992.
Guard Academy, New London, CT, in 1980, the S.M. Since 1993, he has been a Research Engineer at Woods Hole Oceanographic
and E.E. degrees in electrical engineering from the Institution, Woods Hole, MA, developing underwater acoustic communication
Massachusetts Institute of Technology, Cambridge, systems and acoustic tags for marine mammals. His research interests are in
in 1988, and the Ph.D. degree in electrical and ocean miniature signal processing instruments, bioacoustics, and algorithm design.
engineering from the Massachusetts Institute of
Technology/Woods Hole Oceanographic Institution
Joint Program in Oceanography and Oceanographic
Engineering, Woods Hole, MA, in 1992.
Since July 1997, he has been a Scientist in the De-
partment of Applied Ocean Physics and Engineering at the Woods Hole Oceano-
graphic Institution. His research interests are in the areas of adaptive signal pro-
cessing, underwater acoustic propagation modeling, and numerical optimiza-
tion.
Dr. Preisig is the recipient of the 1999 ONR Ocean Acoustics Young Faculty
Award. He is a member of the IEEE Sensor Array and Multichannel Signal
Processing Technical Committee and the ASA Underwater Acoustics and Signal
Processing Technical Committees.

Potrebbero piacerti anche