I.

INTRODUCTION

Modern electronic interception systems must

Digital Channelized Receiver perform the tasks of detection, classification, and
identification in a difficult environment consisting
Based on Time-Frequency of noise, interference, and multiple nonstationary
signals. Moreover, some waveforms are intentionally
Analysis for Signal Interception designed to reduce the probability of interception (low
probability of interception (LPI) signals [1]). This
environment demands advanced signal processing
algorithms running on digital receivers, which
have been attracting considerable attention over

GUSTAVO LOPEZ-RISUE
NO, Member, IEEE the past years [24]. We propose an advanced
GRAJAL
JESUS digital channelized receiver (ADCRx), which is a

ALVARO SANZ-OSORIO
particular case of a more general structure known
Universidad Polite cnica Madrid as time-frequency receiver (TFRx). The TFRx main
Spain feature is the use of time-frequency analysis before
detection and encoding (Fig. 1). Time-frequency
analysis [5, 6] allows simultaneous description of a
signal in time and frequency, so that the temporal
A digital channelized receiver is presented for the interception evolution of the signal spectrum can be analyzed. It
of a wide variety of signals of complex structure, including those has become essential for nonstationary signal analysis;
with low probability of interception. The receiver is designed therefore, it is an appealing tool for the interception
from the perspective of the time-frequency analysis. It uses an of many radar and communications signals. The TFRx
extended time-frequency representation based on the noncoherent
detector works on the time-frequency representation
and builds feature vectors containing the information
integration of the short-time Fourier transform (STFT) on which
relative to the time and frequency where each
the detection system and the encoder work. The encoder includes
detection occurred. The encoder clusters all the feature
robust frequency estimation, automatic modulation recognition,
vectors belonging to the same signal and estimates the
and clustering, to handle broadband and simultaneous signals pulse descriptor word (PDW).
and to prevent out-of-channel detections (a typical phenomenon
in channelized receivers). The receiver has been evaluated for a
wide range of signals and shows a good performance in terms
of detection, estimation, and processing of simultaneous signals.
Signals collected from real-life systems and synthetic signals have
been utilized.

Manuscript received October 20, 2003; revised March 30 and Fig. 1. Architecture of time-frequency receiver.
October 4, 2004, and January 11 and January 25, 2005; released
for publication March 1, 2005.

IEEE Log No. T-AES/41/3/856439. The ADCRx is based on an extension of

the short-time Fourier transform (STFT) or
Refereeing of this contribution was handled by J. P. Y. Lee.
sliding-window Fourier transform [5, 6]. The
This work was supported in party by project receiver is mainly intended for radar signal
TIC2002-04569-C02-01 of the Science and Technology Ministry,
interception; nevertheless, its performance on
and the Navy R&D Center (CIDA).
some communications signals is also analyzed.
This work was presented in part at the International Conference on The encoders PDW comprises the following
Radar, Adelaide, Australia, September 2003.
parameters [3]: time of arrival (TOA), pulsewidth
Authors addresses: G. Lo pez-Risuen o, European Space Research (PW), carrier (or mean) frequency (f), pulse
and Technology Center, European Space Agency, Noordwijk, amplitude (PA), and intrapulse modulation. The
The Netherlands; J. Grajal and A. Sanz-Osorio, Grupo de
modulation is classified into one out of four
Microondas y Radar, Departamento de Sen ales, Sistemas y
Radiocomunicaciones, ETSI de Telecomunicacio n, Universidad categories: no modulation, linear frequency
Polite cnica Madrid, Ciudad Universitaria s/n, 28040 Madrid, Spain. modulation (LFM), phase-shift keying (PSK), and
E-mail: ([email protected]). frequency-shift keying (FSK).
Digital channelized receivers based on the
0018-9251/05/$17.00
c 2005 IEEE STFT were first proposed by Fields et al. [7], and

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005 879
Zahirniak et al. [8],1 as an attempt to benefit A. Time-Frequency Representations
from the advantages of digital technology and the
properties of channelized receivers [9], such as: Time-frequency analysis aims to provide
simultaneous-signal interception, high sensitivity for a physically-meaningful representation of the
narrowband signals, and high dynamic range. Those time-varying spectrum of a signal. A time-frequency
receivers were designed as mere digital replicas of representation is a mapping of the signal under
the previous analog channelized receivers, and were analysis onto the two-dimensional space of time
intended for conventional radar pulses of moderate and frequency. There are many ways of performing
signal-to-noise ratio (SNR around 10 dB) rather such a mapping. Among them, three important
than for the wide variety of signals appearing in categories arise [5, 6]: linear, quadratic and adaptive
current interception scenarios. These new signals have representations. The most important linear techniques
motivated this new time-frequency approach leading are the STFT and the wavelet transform. The
to the ADCRx. most important quadratic representation is the
The ADCRx uses a new enriched STFT-based Wigner-Ville distribution, which gives rise to Cohens
representation including noncoherent integration of class by the convolution with different kernels.
different lengths; thus the integration length becomes Adaptive representations are adaptive versions of the
the third dimension of this extended time-frequency previous categories. The most important ones are the
representation. This improves the performance signal-dependent time-frequency representation [6, 11]
in detection, since the receiver adapts itself to (adaptive version of Cohens class) and the adaptive
signals with different lengths; it also improves the approximation techniques [6, 12, 13], such as the
performance in encoding, since signals are described atomic decomposition2 or the basis pursuit. Adaptive
in a more highly-dimensional space. The ADCRx approximations perform an adaptive linear expansion
provides a more complete PDW by means of a new of the signal using a redundant set of elementary
encoder design featuring a clustering algorithm functions.
adapted to the channelized architecture, a novel robust The most suitable time-frequency techniques
frequency estimation technique, and an automatic for signal interception are the linear and the
modulation classifier based on the instantaneous adaptive-approximation ones, since they deal with
frequency estimation. simultaneous signals without the appearance of the
The receiver works on a block-by-block basis, cross-terms inherent to the quadratic distributions [5].
namely the signal is divided into blocks of samples Moreover, within the linear techniques, the STFT
separately analyzed by the receiver. Afterwards, the is particularly more adequate than the Wavelet
PDWs of every block are passed to a data processor Transform: The STFT can be interpreted as a
that performs deinterleaving and user-interfacing [3], uniform filter bank with constant noise power at
which is beyond the scope of this paper. The the output of each STFT channel. However, the
block-by-block operation is motivated by the practical Wavelet Transform is a filter bank with constant
implementation as subsequently clarified. The paper relative bandwidth [5], so that higher frequency
is organized as follows. Sections II, III, and IV channels hold a higher noise power and a worse
describe the ADCRx time-frequency processor, detection performance. Regarding the computational
detector, and encoder, respectively. Section V shows burden, the STFT is more efficient than the adaptive
the performance for a wide range of signals, including approximations [12, 13]. Consequently, it becomes
high time-bandwidth product signals, such as the LPI the most appropriate time-frequency technique for fast
ones. Section V also discusses the implementation signal interception.
aspects of the ADCRx. The advantages of ADCRx
become apparent by the comparison with more B. STFT Implementation
traditional approaches. Finally, the conclusions are
drawn in Section VI. Part of the content of this paper The discrete version of the STFT is defined
was previously presented in [10]. as [5, 6]
n+(L1)
X
II. TIME-FREQUENCY PROCESSOR STFTx(w) (n, k) = x(m)w(m n)ej(2k=L)m ,
m=n
We first present a brief review of the
k = 0, : : : , L 1 (1)
time-frequency representations and the suitability
of the STFT for interception. Then, the discrete where w(m) is the analysis window defined from 0
STFT and the extension used in the time-frequency to L 1, and x(m) the signal under analysis. Index
processor are described.
2 Also known as matching pursuit [13] or adaptive gabor
1 These systems are also described in [3]. representation [6].

880 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
 Fig. 2. Analysis window for 64-channel receiver. 256 samples, 1 dB passband ripple and 60 dB losses within attenuation band.
(a) Window. (b) Window frequency response.

n refers to the discrete time and k to the channel of C. Extension of STFT

normalized center frequency k=L.
As in [3], [7], and [8], the design of the analysis The STFT holds a fixed time-frequency resolution
window is based on a filtering approach: The that depends on the analysis window. Furthermore,
Parks-McClellan method [14], which computes the time and frequency resolutions are related by the
minimax optimum linear-phase filter that meets a uncertainty principle [5, 6]: A good time resolution
proposed filter mask. To achieve a high dynamic implies poor frequency resolution and vice versa.
range, the filter must possess abrupt transition bands This lack of flexibility, as a consequence of the
and low sidelobes leading to a window much longer STFT nonadaptive characteristic, has a strong impact
than the required number of channels. For instance, on the detection performance [15]: If a signal has a
for a 64-channel receiver, with a 1 dB ripple wider bandwidth than the frequency resolution, the
passband and a 60 dB attenuation band starting at signal energy is spread out over several coefficients
the center of the adjacent filters, a 256-tap window in frequency, so that the probability of detection
is required. This window is shown in Fig. 2. based on individual coefficients reduces. Similarly, a
Let K be the number of desired channels, and L signal with a greater duration than the time resolution
the window length, several techniques can be applied spreads out its energy over several coefficients in time.
to compute discrete Fourier transforms (DFTs) of K The ADCRx implements a more flexible
out of L coefficients, e.g. time data folding [3] or time-frequency representation which is an extension
polyphase filters [8]. Actually, only channels k = 1 of the STFT. This extension relies on different-length
to k = K=2 1 are used since the signal under noncoherent integration and maintains the low
analysis, x(m), is real-valued. Channels k = 0 and k = computational burden of the STFT. The motivation of
K=2 hold different statistical properties and are not this extension is addressed in the following. The STFT
used; thus the number of effective channels is processing gain for a narrowband signal becomes [3]3 :
K=2 1. Gp = K=(2Lins Bn ), where Lins is the channel insertion
The filter-bank interpretation allows the STFT losses at the signal frequency, and Bn the relative
decimation in time. The decimation factor M is noise bandwidth with respect to a K-tap rectangular
upper-bounded by window. This gain exclusively comes from the
M K=2 (2) channelization, and does not benefit from the fact
that a signal with a greater duration than the analysis
due to both the Nyquist sampling criterion and the window stays longer in the same channel, as usually
ambiguity bandwidth of the instantaneous frequency happens to the signals with a high time-bandwidth
estimate [8]. This estimation is performed by the product, such as the LPI ones. In other words, we are
digital instantaneous frequency measurement method losing processing gain whereas LPI signal interception
(DIFM) (treated in Section III). Decimation in time paradoxically demands a high processing gain.
alleviates the STFT computation since the analysis
window is shifted M samples instead of one as in (1). 3 Theprocessing gain of an algorithm is the ratio between the
For an N-sample block, the size of the decimated SNR required to detect a signal without and with it at the same
STFT matrix becomes (1 + (N L)=M) (K=2 1). conditions of probability of false alarm and detection.

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 881
 To increase the processing gain by noncoherent
integration, we define the smoothed spectrogram as
Li m
X
Ii (m, k) = jSTFT(rM, k)j2 (3)
r=1+Li (m1)

where Li is the integration length, and M the

decimation factor.4 A finite set of smoothed
spectrograms with different integration lengths must
be used to account for signals with different durations.
The definition (3) assumes that the integrators work
in an integration and dump mode for the sake of
computational efficiency. The integrator lengths Li are
assumed to be divisible by one another.
With regard to the number of integrators, at Fig. 3. Noncoherent integration processing gain versus signal
least two integration types should be considered. duration.
First, the one-sample integrator (I1 ), which holds the
highest resolution in time; and second, the full-length
integrator (IF ), which achieves the highest processing vector per detection. The detection problem is
gain for long-duration signals. The full-length addressed as a hypothesis testing problem:
integrator is the longest integrator possible within the
H0 : x=n (4)
block of samples, i.e., LF = 1 + (N L)=M.
Intermediate-length integrators are also needed X
H1 : x= sr + n (5)
in order to better match the intermediate-duration
r
signals and prevent the collapsing losses of the longest
integrators [16]. This is illustrated in Fig. 3, where where n is a real-valued zero-mean white Gaussian
the processing gain of integrators with different noise of known power. The noise power is known
lengths is drawn. For each integrator, the collapsing since it mainly comes from the receiver internal
losses appear when the signal is shorter than the noise [3]. Every vector sr corresponds to a signal
integrator length. For longer signals, the processing present in the analyzed N-sample block.
gain is constant and equal to the gain for a signal As the signals are unknown, the detection is
matching the integrator length. For each signal length, carried out by a local procedure, i.e., a detection test
an optimum processing gain exists that corresponds is performed on each point of the time-frequency
to the integrator matching the signal length. This representation, namely each coefficient of the
gain has been approximated in Fig. 3 as a straight
smoothed spectrograms. The local tests turn out to
line.5 According to Fig. 3, the difference in the
be
processing gain of the noncoherent integrators
H1
used in the ADCRx defines an upper bound for the Ii (m, k) ? thi , i = 1, 2, : : : , F (6)
mismatching losses. Mismatching losses are defined H0
as the difference in detecting a signal by means of the
matched integrator instead of the ADCRx. In practice, where the thresholds are set to meet a desired value of
they can be set to 3 dB. To summarize, the ADCRx global false alarm probability (PFAg ):
time-frequency processor outputs the following 0 1
extended representation: fSTFT, I1 , I2 , : : : , IF g. [
PFAg = PH0 @ fIi (m, k) > thi gA : (7)
i,m,k
III. DETECTION AND FEATURE EXTRACTION
That is, the global PFA is the probability of having
A. Detection at least a local detection under the hypothesis H0 . A
local PFA can also be defined for every local detection.
The ADCRx detection stage works on the previous For the time-frequency points of the ith smoothed
time-frequency representation and builds a feature spectrogram, it becomes

4 The classic spectrogram of a signal is the squared absolute value PFAi (m, k) = PH0 (Ii (m, k) > thi ),
of the STFT (jSTFT(n, k)j2 ) [5, 6].
5 The processing gain is approximately proportional to the i = 1, 2, : : : , F: (8)
squared root of the integrator length [17], i.e., it is a straight
line in logarithmic units. In the context of the ADCRx, this As the noise is white, the local PFA is constant for
approximation is valid for Li M=K 1 [17]; nevertheless, we use the local decisions in the same smoothed spectrogram.
the approximation in a general way for a didactic purpose. To find the thresholds fthi g meeting the global PFA

882 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
 Fig. 4. Detection stage of ADCRx.

(PFAg ), we assume for simplicity that the local PFA TABLE I

is the same for all the smoothed spectrograms. For a Example of ADCRx Time-Frequency Processor. Analysis Window
of Fig. 2
given local PFA , the thresholds thi can be analytically
computed as in [18]. The relationship between the Parameter Symbol Value
local PFA and PFAg has to be obtained by simulation
Block length N 1024
(see Appendix C).
(Total) number of channels K 64
Reciprocally, the global probability of detection Effective number of channels K=2 1 31
(Pdg ) can be defined as the probability of having Decimation M 32
at least a local detection under the hypothesis H1 . Integration lengths L1 1
As in (8), a local probability of detection (Pdi (m, k)) for I1 , I2 , I3 L2 5
can be defined for each coefficient of the smoothed L3 25
spectrograms. Parks-McClellan Analysis Window

Length L 256
B. Feature Vector Construction Pass-band ripple Rp 1 dB
Attenuation Ra 60 dB
As mentioned above, there is a corresponding
feature vector per local detection. Apart from the
time-frequency location of the local detection, the integration length (Li ) for channel k, i.e.,
feature vector includes the instantaneous frequency
estimation by using the DIFM. For the channel k, the = [i, Ii (m, k), m, k, DIFM(r1 , k),
DIFM is defined as DIFM(r1 + 1, k), : : : , DIFM(r2 , k)]T (10)
DIFM(r, k)
where indices r1 and r2 are
argfSTFT((r + 1)M, k)g argfSTFT(rM, k)g
= r1 = 1 + (m 1) Li
2M (11)
(9) r2 = m Li :
where M is again the STFT decimation factor, and Table I shows an example of a time-frequency
argfg is the phase angle. The DIFM was used in processor used in subsequent sections. It includes
former digital channelized receivers [3, 7, 8] to three integrators of 1, 5, and 25 samples, so that the
estimate the carrier frequency. It is used here to maximum loss in the noncoherent processing gain
1) estimate the carrier frequency in a more robust using the ADCRx will be around 5 log10 L3 =L2 =
way, and 2) recognize the signal modulation (both 5 log10 L2 =L1 = 3:5 dB, according to the discussion in
tasks treated in Section IV). the previous section. A complete diagram of detection
The exact structure of the feature vector is detailed stage of this system is also shown in Fig. 4. The
as follows. If a local detection occurs at Ii (m, k), the diagram follows the filter-bank perspective for the
feature vector () contains the smoothed spectrogram sake of clarity. At the output of every channel, the
index (i), its numerical value (Ii (m, k)), time location integrators work in parallel, in integration and dump
(m) and channel (k), and the DIFMs within the mode.

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 883
IV. ENCODER to the integration length of I2 , etc. The classification is
carried out by successive hypothesis tests:
From the feature vectors built in the detection
stage, the encoder estimates the number of signals in c
PA Shorter than IF
F1
? th ,
the block under analysis, and their PDWs. Depending c Duration as IF PAF1,F
PA F
on its SNR, bandwidth, and duration, a signal can
be simultaneously detected in different points of c
PA Shorter than IF1
F2
? thPAF2,F1 , : : :
different smoothed spectrograms, so that several c
PA Duration as IF1
F1
feature vectors can correspond to the same signal.
Therefore, clustering is required to group all the c Duration
PA as I1
1
? thPA1,2 (13)
feature vectors of the same signal. The PDW of every c Duration
PA as I2
2
signal is estimated from the feature vectors of the
corresponding cluster. The encoder takes advantage of that rely upon the fact that, when the signal is shorter
the STFT filter-bank interpretation by working in two than the integrator length, PAc tends to be smaller
i
steps: in-channel and in-block. First, the in-channel than the true value due to the additional integration
processing clusters the feature vectors from the of noise-only samples. Each threshold thPAi1,i is
same channel, and computes an in-channel PDW by obtained by simulating radar pulses with a similar
assuming only one signal per channel. Second, the length to the integrators Ii1 and Ii and by minimizing
in-block processing clusters the in-channel PDWs of c =PA
the probability of error in the test PA c (see
i1 i
the same signal, fuses those PDWs into an in-block Appendix C). For illustration purposes, Fig. 5
PDW, and removes the out-of-channel detections. shows the probability of misclassification for several
The one-signal-per-channel assumption is valid conventional radar pulses. We use the system
for short- or medium-duration blocks, i.e., blocks configuration of Table I, which has three noncoherent
of samples lasting microseconds or hundreds of integrators and discriminates among three types of
microseconds. For instance, in a dense environment of duration: short (as the integrator I1 ), medium-length
105 signals per second in the band from 9 to 9:5 GHz, (as the integrator I2 ) and long signals (as the integrator
the mean number of signals within a 100 s interval I3 ). As can be noticed, the pulses are properly
is 10; thus, on average, there would be a signal every classified for SNRs above 0 dB.
50 MHz-bandwidth channel.6 In practice, this kind of 2) PA Estimate: If the signal is classified into the
blocks is required to achieve low latency times. category of signals of length similar to the integrator
c is defined as
Ii , the in-channel PA estimate (PA)
A. In-Channel PDW c
the one corresponding to the integrator Ii , i.e., PAi
2
(eqn. (12)). As the noise power is known ( ), the
In-channel PDW construction is divided into SNR estimate becomes
several steps which are repeated for each channel
c2
where, at least, a local detection occurred. d = PA :
SNR (14)
1) Signal Duration: Let k(i) be the set of 22
feature vectors with smoothed spectrogram index i 3) TOA and PW Estimates: Temporal parameters
and channel k, the in-channel PA estimate for each (TOA and PW) are estimated according to the prior
smoothed spectrogram becomes classification. That is, if the signal has a duration
p
maxf2 I()=Li : 2 k(i) g, if k(i) 6= similar to the integrator Ii , the feature vectors coming
c
PAi = from the smoothed spectrogram Ii are used. In order
0, otherwise:
to compute TOAd and PW, d the initial (n ) and final
init
(12) (nend ) decimated samples of the signal at channel k are
I() is the amplitude of the smoothed spectrogram required:
corresponding to the feature vector . The estimates
c s are unbiased and consistent for sinusoidal signals ninit = minf(m() 1) Li + 1 : 2 k(i) g (15)
PA i
at a high SNR [19, sect. 6.1.2].7 nend = maxfm() Li : 2 k(i) g (16)
Using these estimates, the signal is classified into
a number of categories corresponding to the different where Ii is the smoothed spectrogram corresponding
smoothed spectrograms: signals of duration similar to to the signal duration, and m() is the time location
the integration length of I1 , signals of duration similar of each feature vector . Definitions (15) and (16) are
appropriate for low SNR. For high SNR, the smoothed
6 A uniform distribution of the signals in time and frequency is
spectrogram I1 is used, instead of the Ii , in order
assumed.
to provide a more accurate estimation. Smoothed
7 No insertion losses in the channel are assumed. In the context spectrogram I1 is used when SNR d is over a certain
of [19, sect. 6.1.2.], Ii (n, k) can be viewed as the signal smoothed threshold (SNRth ) that assures a high probability of
periodogram. detection (in practice, this threshold can be defined

884 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
 Fig. 5. Probability of misclassification in terms of duration for several conventional radar pulses. System configuration of Table I;
thPA12 = 4 and thPA23 = 2; block of 1024 samples.

as the SNR to detect a sinusoid with a probability of with the same sign, so that a bias is introduced. We
99% by means of the I1 ). have checked that the bias is below the 10% of the
For the sake of clarity, TOA and PW estimators channel bandwidth for different PSKs (with smaller
are defined in Appendix A, where the effect of the bandwidth than the channel) and configurations
analysis window transient is also considered. (channel bandwidth and decimation factors). If
4) Frequency Estimate: The frequency estimate a smaller error is required, PSK-oriented carrier
is computed by a weighted average of the DIFM frequency estimators [21, ch. 3] can be used after the
samples from the time sample ninit (eqn. (15)) to modulation classification. The modulation classifier is
nend 1 (eqn. (16)) at the considered channel. The frequency-shift invariant; hence, it is not affected by
weights are the values of the smoothed spectrogram I1 the frequency estimation bias (see Section IVB).
and the expression of the frequency estimate becomes For sinusoids, Kays estimate is better and reaches
Pnend 1 the Cramer-Rao bound for high SNR [20]. This is
f = r=ninit I1 (r, k)DIFM(r, k) also shown in Fig. 6. The weighted and nonweighted
Pnend 1 : (17)
r=ninit I1 (r, k)
DIFM average exhibit the same performance, which
is slightly worse than the Cramer-Rao bound. The
This expression also turns out to be the mean maximum likelihood frequency estimate is not
frequency estimate in the framework of the considered since it does not work with DIFM samples
time-frequency analysis [6]. but with the signal Fourier transform, which is more
The weighted DIFM average has been compared computationally demanding. Moreover, both Kays
with other methods working with instantaneous and the maximum likelihood estimate have the same
frequency estimates as well, such as Kays performance for high SNR [20]. On the whole, the
estimate [20], the single-DIFM estimate, and the weighted DIFM average is the most robust estimate
nonweighted DIFM average [8]. The weighted for both nonstationary and stationary signals, and
DIFM average has the best performance for becomes very suitable for the encoder implementation.
nonstationary signals as illustrated in Fig. 6, where
the root-mean-square error (RMSE) versus the SNR is
B. Modulation Recognition
shown for a BPSK signal intercepted by the system
of Table I. The RMSE of those estimates does not The in-channel automatic modulation classifier
decrease as the SNR increases since they are biased. (AMC) follows a decision-theoretic approach inspired
The bias is due to the DIFM ambiguity bandwidth [8], by [22] and [23]. It distinguishes among four
which amounts to 1=(2M) for a given decimation categories: no modulation, LFM, PSK, and FSK,
factor (M). Assuming ideal symbol-to-symbol which are the typical modulations encountered in
transitions, a 180 phase transition results in a radar and digital communications systems.
frequency equal to half the ambiguity bandwidth, i.e., AMC uses the DIFM outputs of the considered
1=(2M). Thus these transitions are always computed channel within the signal duration, i.e., between

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 885
 Fig. 6. RMSE of frequency estimated by several techniques using DIFM samples for BPSK with 62 samples/symbol and a sinusoid.
Table I system configuration. BPSK and channel bandwidths are approximately the same (for 250 MHz sampling rate, 4 and 3.9 MHz,
respectively).

the decimated samples ninit and nend 1 (eqns. (15)

and (16)). In the preprocessing steps, a number
of initial and final samples are removed to avoid
peaks from the signal transients, and the signal
length is verified to be greater than a minimum
value. If lower, the signal is considered to be
nonmodulated. Then, AMC proceeds as depicted in
Fig. 7. First, a frequency linear model is obtained
by least squares. The model error (") helps separate
LFM and nonmodulated signals from PSK- and
FSK-modulated ones. The discrimination between
LFM and nonmodulated signals is made by the
magnitude of the estimated chirp-rate (a ). AMC
discriminates between PSK and FSK by the maximum
of the 1st-order difference of the DIFM sequence
(f), i.e., Fig. 7. In-channel AMC flow chart.

f(r, k) = DIFM(r + 1, k) DIFM(r, k) (18) The thresholds for the estimators " and a hold the
following general structure:
since PSK phase transitions are expected to be greater
than FSK frequency transitions. This relies upon d PW)
thT (SNR, d PW)
d = (SNR, d PW)
d + c (SNR, d
T T T
the fact that, for PSKs with ideal symbol-to-symbol
transitions, the amplitude of the frequency impulse (19)
corresponding to a phase step is: f = =2M where T and T are the mean and standard deviation
(M is the decimation factor). We restrict the FSK of the considered statistic T. To compute thLFM ,
symbol separation to the channel bandwidth in order " and " come from the distribution of the model
that the AMC succeeds in the classification. For error " for both nonmodulated and LFM signals.
broader band FSKs, the correct recognition should Regarding tha , a and a come from the distribution
be done by the in-block processing. However, the of the chirp-rate estimate a for nonmodulated signals
variety of FSK signals makes the design of rules for only. Thresholds tha and thLFM can be analytically
in-block recognition rather difficult, and the only computed as indicated in Appendix B. Constant cT
modulation-related task performed by the in-block in (19) minimizes the probability of error among
clustering is the recognition of fast LFM signals (see the different modulation classes. This probability
Section IVC). is computed by simulation as the sum of the

886 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
probability of wrong classification for several signals Otherwise, the chirp-rate estimation is as described in
representative of all the modulation categories (see the previous paragraph.
Appendix C). For the sake of simplicity, the threshold To illustrate the benefits of the in-block PDW
thf is defined as a constant and also computed by estimation, a chirped pulse sweeping 40 MHz in
simulation to minimize the misclassification error 4 s have been analyzed by the system of Table I
(Appendix C). at a 250 MHz sampling rate. The signal features a
fast LFM modulation, and sweeps 10:24 channels
C. In-Block PDW in 1000 samples. Fig. 8 shows the reduction in the
TOA RMSE by means of in-block clustering with
The in-channel PDWs are clustered according respect to the main in-channel PDW (on the left). The
to a number of heuristic rules. First of all, the in-block probability of right modulation recognition
in-channel PDW with the greatest SNR,d which is also depicted (on the right). It becomes 90% for
we call main PDW, is selected and its adjacent 4 dB SNR. However, the in-channel probability of
channels are grouped into the cluster according to recognition is zero since the minimum number of
d and main PDW modulation.
time overlapping, SNR, samples to run the AMC is not reached.
For each cluster, the final PDW is the corresponding 2) Cancellation of Out-of-Channel Detections:
Once a cluster has been formed, other nongrouped
main PDW except for clusters considered to come
channels may also be removed to prevent
from LFM signals. In that case, a reestimation of the
out-of-channel detections. Two types of out-of-channel
parameters is performed. The remaining nongrouped
detections are considered: rabbit-ear effect and signal
in-channel PDWs undergo the same process by
sidelobe detections. The cancellation algorithm uses
selecting a new main PDW.
two thresholds determined by simulation and proceeds
d is required in order to consider
The estimate SNR as follows.
the overlapping in time, since temporal parameter
estimation degrades for low SNR. A threshold SNRf a) The lowest threshold is related to the rabbit-ear
is used, so that overlapping in time is not considered effect. If the number of channels in the cluster
d < SNR . The detailed description of the
for SNR are greater than the threshold, the nongrouped
f
in-block rules is beyond the scope of the paper. overlapped-in-time channels with short duration8
are removed. This is the case of high-SNR pulses,
Nevertheless, there are two aspects that deserve
since their leading and trailing edges usually cause
further consideration: The improvement of PDW
short-duration detections in adjacent channels
estimation for LFMs by in-block clustering, and the
(high-SNR pulses are used to compute this threshold).
cancellation of out-of-channel detections.
b) The greatest threshold eliminates signal
1) PDW Estimation for LFM Signals: In the
sidelobe detections. If the number of channels in the
in-block processing, a cluster is considered as LFM
cluster are greater than this, the signal has a very
if the main PDW is found to be LFM modulated. broad band and all nongrouped, overlapped-in-time
However, for fast LFM signals, there could be channels are removed (independently of their
not enough samples to carry out the in-channel duration). This is the case of high-SNR very short
modulation analysis. In those cases, the LFM pulses and high-SNR very broadband PSKs. For these
character becomes apparent after examining the cases, the signal sidelobes cause detections in channels
time-frequency arrangement of the in-channel so far away from the signal frequency that they are
PDWs forming the cluster. In this paper, the rule for not clustered with the corresponding main PDW
automatic recognition of fast LFMs is as simple as (high-SNR broadband PSKs are used to compute this
checking that there is a significant shift in time and threshold).
frequency between the main PDW and the PDW of
an adjacent channel. The chirp-rate estimate is then In [7] and [8], another cancellation method
obtained as the frequency separation-time separation is proposed based on the DIFM. It is suitable for
nonmodulated narrowband pulses with high SNR.
ratio between these two PDWs.
For low SNR or modulated signals, the DIFM errors
For LFM signals, the final PDW takes into account
become so high, as was shown in Fig. 6, that this
the in-channel PDWs of the cluster. The in-block
method cannot apply.
TOA is reestimated as the minimum in-channel
TOA (similarly for the PW) and the in-block
mean frequency is the weighted average of the V. RESULTS AND DISCUSSION
in-channel frequency estimates. The weights are the
c The in-block PA
corresponding in-channel PAs. c is the This section describes the performance of the
c Regarding the chirp-rate estimation, ADCRx for a wide variety of signals related to the
main PDWs PA.
if the main PDW is found to be LFM modulated, the
8n ninit + 1 2.
chirp-rate estimate is simply the main PDW chirp-rate. end

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 887
Fig. 8. In-block clustering for fast chirped pulse sweeping 40 MHz in 4 s: TOA RMSE of in-block and main PDW are compared on
the left. On the right, in-block probability of modulation recognition is shown. (System configuration in Table I, sampling rate
250 MHz.)

radar field: conventional pulses, (linear) chirped TABLE II

pulses, CW signals and phase-coded signals. To Example of ADCRx Encoder (for Time-Frequency Processor in
Table I)
complete the analysis, some digital communications
signals have been evaluated as well. Among the Parameter Symbol Value
phase-coded waveforms, this section studies Barker
Threshold for thPA2,3 2
pulses and long-duration BPSK signals with
long/medium-length signal
random modulation. Long-duration BPSKs with Threshold for thPA1,2 4
random modulation are a generalization of the medium-length/short signal
pseudorandom-code radars [24, ch. 10] and the BPSK SNR threshold to compute SNRth 3 dB
communications signals, since neither the code in TOA and PW by I1
radar nor the information in communications are Guard samples for AMC L=2M 1 = 3
(left and right)
a priori known by the interception receiver. Several d < SNRth
Minimum length 5, if SNR
minimum shift keying (MSK) signals [25] are also
(after guard sample removal) 3, otherwise
utilized as representatives of FSK communications Constant for threshold tha ca 9
signals. Constant for threshold thLFM cLFM 20
Regarding the system, a 250 MHz sampling Threshold for PSK/FSK thf 6:94 104 (80 )
rate has been selected. The parameters of the SNR threshold for in-block SNRf 0 dB
time-frequency representation are the same as in clustering
Table I: a 3.9 MHz channelization and a 4.1 s Number of channels for 4
signal block. For the encoder, the main parameters rabbit-ear cancellation
Number of channels for 20
are listed in Table II. In the subsequent sections,
broadband signals
the receiver is evaluated in terms of sensitivity and (sidelobe cancellation)
estimation error of the in-block PDW. Some of the
signals in this section were collected by an X-band
receiver at 250 Msamples/s and quantized with additional computer-generated noise has been added
8 bits. These real-life signals correspond to existing to meet the desired SNR values.
radar and communications systems operating in
real environments. Hence, the data are affected by A. Sensitivity
the linear and nonlinear distortions of the sampling
receiver. These distortions have turned out to be ADCRx is compared with other typical systems
unimportant since analyses with purely synthetic in terms of sensitivity for a wide variety of signals,
signals with the same characteristics as the real-life some of them exhibiting the LPI features. The other
ones yielded the same results. The real-life signals receivers are: A channelized receiver based solely
were collected at a high SNR condition, so that on the STFT like the approaches of Fields et al. [7]

888 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
 TABLE III
Sensitivity (dB)

Signal ADCRx STFT (only) DFT ED MF

Conventional pulse 500 ns 1:6 0:4 3.9 3.6 5

Conventional pulse 1 s 4 1:5 1:4 0.6 8
CWLFM 100 MHz/1 ms 8:5 1:5 11:5 5:3 14
CWLFM 500 MHz/1 ms 8:5 2 6 5:3 14
Chirped pulse 1 MHz/1 s 4 1:1 1 0.6 8
Chirped pulse 2 MHz/2 s 5:5 1:6 2:5 2:2 11
Chirped pulse 40 MHz/4 s() 4 3:4 4.5 5:3 14
BPSK 2 MHz 8 3:2 6:5 5:3 14
BPSK 4 MHz 8 3:2 6:5 5:3 14
Barker-13 pulse 4.8 s() 6 3:7 5:5 5:3 14
BPSK 10 MHz 5:5 2:2 2 5:3 14
MSK 1 MHz (freq. sep.)/500 ns (bit duration) 6 2 9:5 5:3 14
MSK 3.3 MHz (freq. sep.)/150 ns (bit duration)() 5:5 2:9 6 5:3 14

Note: SNR at PFAg = 106 , Pdg = 90%, 250 MHz sampling rate, 1024-sample block. Signals with asterisk () are from real-life
systems in their typical operating environments. Pulsed signals are time centered in block and, if nonmodulated, its frequency changes
from a realization to another according to a uniform distribution within the channel bandwidth. For other synthetic signals, carrier
frequency coincides with center of 16th STFT channel.

and Zahirniak et al. [8]; a receiver based on the (TOA, PW and modulation). Furthermore, the ED is
DFT with a quadratic detector after every bin (it not suited to handle simultaneous signals.
can be considered as a channelized receiver with an
analysis window as long as the block length); the B. Parameter Estimation Performance
energy detector (ED); and, finally, the matched filter
(MF), which is the optimum detector when the signal Relative estimation errors of the in-block PDW
parameters are known, and the signal phase offset is parameters are in general below 20%, and, in most
uniformly distributed over [0, 2). cases, below 10% for a SNR greater than 0 dB.
The comparison is shown in Table III, where c
This is illustrated in Fig. 9, where the RMSE of PA,
the sensitivity is defined as the minimum SNR for PW,
d f, and the chirp-rate estimate are depicted for
PFAg = 106 and Pdg = 90%. CWLFM refers to the signals of Table III. Although not shown, TOAd
the continuous-wave linear frequency modulated d
performance is quite similar to that of PW.
signals. Signals with an asterisk () come from the The exceptions to the general good performance
real-life collecting system. The overall ADCRx are the broadband signals, i.e., broadband PSKs/FSKs
performance is better than the others (apart from and fast LFMs. For instance, PA c RMSE becomes
the MF). Its processing gain with respect to the relatively higher for the 10 MHz BPSK, since the
receiver based on the single STFT is due to the estimator is not intended for signals occupying several
noncoherent integration. The improvement ranges channels. The 40 MHz/4 s chirped pulse has such
from 1 to 7 dB. The DFT receiver is only suitable a fast modulation that the in-channel modulation
for very narrowband signals due to its narrow analysis is not possible and the chirp-rate estimation
channelization. For instance, the 100 MHz/1 ms must be estimated in the in-block analysis, which
CWLFM sweeps a very small bandwidth in the presents a greater error than the in-channel one.
analyzed interval (around 400 KHz). It should be This chirped pulse was already used to illustrate the
noted that the DFT has the best performance for the in-block processing (Section IVC, Fig. 8).
MSKs herein used as well, since MSKs typically Despite the overall good performance, the
feature a very concentrated spectrum [25]. Obviously, estimation errors do not always tend to zero as the
the DFT will perform worse than the ADCRx for SNR increases. For PA, c this is due to the channel
higher MSK symbol rates due to the increase in insertion losses, which are not considered in the
bandwidth. estimator because of the difficulty to be computed
For signals with very broad bandwidths, the for broadband signals. The estimates TOA d and PW d
ADCRx can become worse than the ED since are biased due to the decimation uncertainty and the
channelization no longer provides an improvement window transient (see Appendix A). Concerning the
and the only improvement factor comes from chirp-rate estimation, a bias appears that increases
the noncoherent integration (note that ED is a as chirp-rate does. It is caused by the channelization,
noncoherent integrator without channelization). since the filtering of a linear chirp by the STFT adds
Finally, note that neither the DFT nor the ED are a nonlinear frequency modulation to the linear chirp.
able to extract temporal information about the signals Obviously, this additional term becomes zero for

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 889
 Fig. 9. Parameter estimation for signals in Table III versus SNR. (a) RMSE of P c c in
A normalized to true value. (b) RMSE of PW
samples (note that block length is 1024). PSKs and FSKs have same RMSE as CWLFMs. (c) RMSE of f relative to channel
bandwidth. (d) RMSE of chirp-rate relative to true value (only for LFM signals).

nonmodulated signals. For the frequency estimation, The classifier thresholds have been set to
this issue has been already discussed in Section IVA; favor the recognition of nonmodulated pulses.
nevertheless, the errors reported herein can be Thus, the required SNR for the 90% of correct
slightly greater than those presented in Section IVA classification appears very low for conventional
due to the errors in the previous estimation of the pulses. Obviously, for LFM signals sweeping a narrow
beginning (ninit ) and ending (nend ) of the signal (ninit bandwidth within the analyzed time interval (e.g., the
and nend were known a priori in the simulations of 100 MHz/1 ms CWLFM) the correct classification
Section IVA). occurs at higher SNRs than for broader band LFMs.
When misclassified, they are usually classified as
nonmodulated. The case of short chirped pulses, like
C. Modulation Recognition Performance the 1 MHz/1 s one, is discussed later in this section.
Concerning the PSKs and FSKs, a degradation
In general, the probability of correct classification in performance occurs for those occupying several
approaches 100% as the signal SNR increases. The channels, so that 100% of correct classification could
SNR value reaching the 100% of correct classification not be achieved. In those cases, the in-channel AMC
depends on the signal and the system architecture. The cannot properly recognize the modulation since
classifier performance for the signals in Table III is filtering destroys the modulation information carried
shown in Table IV by means of the confusion matrix in the phase. This is the case of the 3.3 MHz/150 ns
at 10 dB SNR. The SNR for 90% of probability of MSK, whose 3 dB bandwidth is 7 MHz. Although not
correct classification is provided as well. illustrated, another drawback caused by channelization

890 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
 Fig. 10. Modulation recognition for multicomponent signal composed of chirped pulse 40 MHz/4 s and CW signal without
modulation. Separation: 1 channel. (Probability conditioned to detection). (a) Multicomponent signal with several spectral separations.
(b) Number of estimated PDW versus SNR.

TABLE IV
Confusion Matrix for Modulation Recognition (Conditioned to Detection) at 10 dB SNR

Signal No Modulation LFM PSK FSK SNR (dB) 90%

Pulse 500 ns 98% 2% 0 0 5

Pulse 1 s 97% 3% 0 0 1
Chirped pulse 1 MHz/1 s 89% 11% 0 0
Chirped pulse 2 MHz/2 s 0 100% 0 0 4
Chirped pulse 40 MHz/4 s() 0 100% 0 0 4
CWLFM 100 MHz/1 ms 62% 38% 0 0 14
CWLFM 500 MHz/1 ms 0 100% 0 0 1
BPSK 2 MHz 0 0 100% 0 0
BPSK 4 MHz 0 0 100% 0 0
Barker-13 pulse 4.8 s() 0 0 100% 0 0
BPSK 10 MHz 16% 0 82% 2% 15
MSK 1 MHz/500 ns 0 0 0 100% 8
MSK 3.3 MHz/150 ns() 2% 15% 48% 35%
Note: Input SNR for 90% of right modulation classification is also provided. Signals with asterisk () were collected from real-life
systems in their typical operating environments. Symbol means that the 90% probability cannot be achieved.

is the degradation in the modulation recognition recognition due to the transient of the analysis
when the PSK or FSK carrier frequency exhibits an window.
important offset from the channel center frequency. Typical modulation recognition approaches, like
Both problems can be overcome by reconstructing Azzouz and Nandis [22] and Liedtkes [26], are
the signal using the Gabor expansion [6] of the STFT intended for isolated communications signals and
coefficients from the channels where it was detected. require long records of them. Moreover, due to the
However, this operation can be very time consuming processing gain of channelization, the use of the
for fast signal interception. ADCRx results in a higher recognition performance
The performance for short chirped pulses, such in terms of SNR (except for the above-referred case of
as the 1 MHz/1 s one, is very poor as well: there very broadband PSKs and FSKs).
are not enough samples due to their relatively short
duration and the use of the maximum decimation D. Multicomponent Example and Estimation of the
factor (M = 32). This problem may be overcome by Number of PDWs
the use of a lower decimation factor. For example,
if M = 16 were used, a 90% of correct classification The ADCRx has also been tested for simultaneous
would be achieved at 20 dB SNR. Nevertheless, signals. In this section two simultaneous signals
this reduction in the decimation approximately are considered: a 40 MHz/4 s chirped pulse
doubles the computational burden. Experiments also and a nonmodulated CW signal in three relative
shows that, below the decimation factor K=4, no arrangements with a 1-, 2- and 3-channel separation,
further improvement is achieved in the modulation respectively. Fig. 10(a) shows the probability of right

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 891
modulation recognition for both signals. Only the can be implemented in one FPGA since, due to
results for 1-channel separation are displayed since the decimation factor (M = 32), we require a time
the others are similar. Both signals have the same between consecutive FFTs of less than 128 ns. The
SNR. The modulation recognition is similar to the rest of the FPGA area can be utilized to implement
case of the signals separately considered except for the smoothed spectrograms, the detection, the feature
high SNRs, at which both signals tend to occupy building, and even the heaviest part of the in-channel
their adjacent channels, so that in-block clustering PDW estimation. To assure real-time operation, a
groups both signals in the same cluster. Due to the few digital signal processors (DSP) can work in
arrangement of the in-channel PDWs, the cluster parallel afterwards to complete the in-channel PDWs
tends to be classified as LFM. Thus, the probability and construct the in-block PDWs of each block of
of recognizing the modulation of the CW signal samples. This multiprocessor architecture represents
decreases. Likewise, the probability of estimating a feasible real-time implementation of an ADCRx
two PDWs, i.e., two signals, decreases with the SNR, with an instantaneous bandwidth of 125 MHz. To
whereas the probability of finding only one PDW facilitate the implementation, there are already
increases (Fig. 10(b)). FPGAs including a PowerPC processor, such as the
Regarding the estimation of the number of signals, Xilinxs Virtex Pro family, and many manufacturers
it is just the number of in-block PDWs. Fig. 10(b) provide multiprocessor boards (see for example
shows the probability of estimating one or two signals [28]).
and the probability of detecting false signals (more
than two PDWs). There are some important peaks
caused by in-block clustering errors when processing VI. CONCLUSIONS
fast LFMs at particular SNR values. For other signals,
the in-block clustering performs properly, and shows This paper describes an ADCRx for automatic
a very low probability of estimating false signals. signal interception, whose design is based on
Nevertheless, small peaks in the error (usually below time-frequency analysis. It uses an extension of the
10%) may appear at certain SNR values. STFT consisting of a set of smoothed spectrograms
obtained by noncoherent integration of the
representation generated by the STFT. Noncoherent
E. Computational Load and Practical Implementation integration improves the processing gain and is
carried out along the time axis. Different integration
The novel features of the ADCRx slightly increase lengths are used in order to adapt to different signal
its computational load in comparison to the previous durations.
digital channelized architectures [3, 7, 8]. Simulations The ADCRx includes a novel encoder with the
using MATLAB on a 1.2 GHz, 256 MB RAM, following notable aspects: 1) the use of two levels
Pentium IV showed that the smoothed spectrograms of processing (in-channel and in-block), 2) the
increase the computational load only by 5%. It should fusion of the information from the various smoothed
also be noted that the in-channel processing of the spectrograms for a particular channel, 3) the use
encoder can be computationally as demanding as of weighted DIFM averaging to obtain a robust
the previous ADCRx stages. This depends on the frequency estimate independently of the modulation,
number of local detections in the detection stage 4) the AMC, and 5) the cancellation of out-of-channel
and is similar to the results obtained in the detection at the in-block level.
simulation of other channelized architectures The ADCRx is an efficient algorithm, which is
[3, 7, 8].9 technologically feasible for fast signal interception
Concerning the practical implementation of the (using a multiprocessor architecture, an instantaneous
ADCRx, a recent work [27] on the implementation bandwidth of several hundreds of MHz can be
of the fast Fourier transform (FFT) in a commercial analyzed in real time). It combines both the
field programmable gate array (FPGA) has shown advantages of digital technology and channelized
pipelined architectures working up to 960 Msamples/s receivers, and outperforms previous approaches in
for different FFT sizes, and performing consecutive terms of sensitivity and variety of signals that can
windowed FFTs of 64 samples every 70 ns with a be intercepted. Additionally, it is able to handle
cost in area lesser than 15%. The time-frequency simultaneous signals with very different properties.
processor of Table I, at a sampling rate of 250 MHz, Nevertheless, there are a number of limitations
inherent to the channelization and the two-level
9 Nevertheless, the encoders computational load is expected to encoder architecture: Modulation recognition of
decrease when implemented in a real-time oriented computing broadband PSKs and FSKs, estimation of extra
language, since most of the time in the current version is devoted to PDWs for fast LFMs occupying many channels, and
transferring data among the modules forming the encoder software. estimation errors for broadband signals.

892 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
 d can be written
estimator PW
d = maxfM, PW0 g
PW
8
> (nend ninit + 1)M K,
>
>
>
> ninit 6= 1, nend 6= LF
>
>
>
>
>
> L=2 + (nend 1)M K=2,
>
>
>
< ninit = 1, nend = LF
PW0 = :
>
> (L 3K)=4 + (nend ninit + 1)M M=2,
>
>
>
>
>
> ninit 6= 1, nend = LF
>
>
Fig. 11. Possible values of TOA for decimation factor M >
> (L 3K)=4 + nend M M=2,
>
>
(M K=2). :
ninit = 1, nend 6= LF
(22)
APPENDIX A. IN-CHANNEL TEMPORAL PARAMETER
ESTIMATION
B. Analysis Window Transient
The in-channel TOA and PW are estimated by d and PW d from (21) and (22)
The use of TOA
using the parameters ninit and nend (eqns. (15) and
results in an increase in the error for both parameters
(16)). In doing so, two issues should be taken into
as SNR increases. This is due to the fact that the
account: the uncertainty due to the decimation, and the
analysis window length (L) is greater than the number
transient due to the analysis window. Both problems
of channels (K). Figs. 12(a) and (b) show this effect
and the solutions adopted in the paper are described as
for the configuration in Table I and a nonmodulated
below.
pulse of 500 samples. Note the staircase behavior
of the TOA and PW RMSE (without correction)
A. Decimation Uncertainty caused by the bias in the estimation of the TOA and
the PW. The bias is due to the transients induced
Using the approximate relation between time and by the analysis window: The TOA estimate tends to
bandwidth (time 1=bandwidth), the effective length occur earlier than the true value and the PW estimate
of the STFT analysis window can be assumed to be becomes greater than the true one.
equal to the total number of channels (K). Given a The behavior of the TOA and PW estimates
decimation factor of M samples, a typical situation suggests the use of a staircase-like correction term for
for the pulse leading edge is shown in Fig. 11. both ninit and nend before TOA and PW computation
Considering ninit (the first decimated sample of the (by (21) and (22)). This correction depends on the
STFT where the signal was detected), there is a range SNR, ranges from 0 to L=2M (half of the maximum
of possible TOAs depending on the window length transient duration), and has to be adjusted for every
and the decimation factor. This uncertainty is modeled system. For the system in Table I, this correction
here as a discrete uniform distribution. In particular, becomes
TOA becomes a random variable, defined as 8 d < 15
>
> 0, if SNR
>
>
L=2 + (ninit 2)M + K=2 + Ud (1, M) (20) >
< 1, d < 20
if 15 SNR
: (23)
where L is the (true) window length, and Ud (1, M) >
> 2, if 20 SNRd < 40
>
>
stands for a discrete uniform random variable ranging >
:
3, d 40
if SNR
from 1 to M. For ninit = 1, the expression is slightly
different (Ud (1, (K + M)=2)). As can be noticed in Figs. 12(a) and (b), the
d utilizes the mean of the
The TOA estimator (TOA) correction results in a significant improvement in
corresponding discrete uniform distribution, i.e., the estimation of both in-channel TOA and PW.
8 2+L+K Both figures also show the greater importance of the
< , ninit = 1 estimate bias over its variance, also caused by the two
d = 4
TOA issues discussed above, i.e., decimation and window
: L+K M 1
+ (ninit 1)M , otherwise. transient.
2 2
(21)
APPENDIX B. AMC THRESHOLD COMPUTATION
d = TEND
The PW estimate is defined as: PW d
d + 1, where TEND
TOA d is the pulse trailing edge, This section summarizes the deduction of the
d The
which is computed in a similar way as TOA. analytical expressions for thresholds tha and thLFM

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 893
 Fig. 12. Mean error and RMSE of TOA and PW estimation with and without transient correction. ADCRx of Table I and
nonmodulated pulse of 500 samples. (a) Error in TOA estimation. (b) Error in PW estimation.

of the in-channel AMC (Section IVB). Let x(r) be R (R + 1) conversion matrix

a narrowband signal corrupted by a real, zero-mean 0 1
1 1 0 0
white Gaussian noise n(r), i.e.,
B 0 1 1 0C
1 BB
C
C
x(r) = A cos(2f0 r + (r)) + n(r) (24) = . .. C : (29)
2M B@ .. .A
where f0 is the carrier frequency, and (r) stands
0 1 1
for the frequency/phase modulation. For the channel
where the signal is, the phase of the decimated STFT In the following, the mathematical expressions of both
can be approximated by thresholds are treated.

argfSTFT(w)
x (rM, k0 )g
A. Threshold tha
2(f0 k0 =K)rM + (rM) + nk0 (rM)
A least squares linear model for the instantaneous
(25)
frequency vector of (27) is given by
for high SNR [20]. This result is obtained by a
a
first-order approximation. nk0 is the phase error, = A+ fi (30)
which follows a real, zero-mean Gaussian distribution. b
Its covariance depends on the aperiodic correlation of where A+ is the Penrose pseudoinverse of
the analysis window, i.e., 0 1
0 1
L1
Lf0 X B M 1C
Cn (mM) = w(r) w(r + mM) (26) B C
SNR A=B B .. .. C
C (31)
r=0 @ . .A
where Lf0 accounts for the insertion losses of channel (R 1) M 1
k0 at frequency f0 . The SNR is defined at the input of
the STFT, i.e., SNR = A2 =22 . and a , b are the model parameters: fi (rM) = a rM + b.

Due to the relationship between the STFT phase Provided that fi follows a linear model plus a
and the instantaneous frequency estimate (DIFM, Gaussian noise of covariance (28), the estimate a also
Section III), the phase error gives rise to a real, exhibits a Gaussian distribution with variance
zero-mean Gaussian frequency error. Assuming an
a2 = A+ (1)Cnf (A+ (1))T (32)
R-dimensional vector of instantaneous frequency
samples, where A+ (1) means the first row of A+ .
If the signal does not have modulation, the mean
fi = [fi (0), fi (M), : : : , fi ((R 1)M)]T (27) of a becomes zero. Therefore, the threshold tha has
its covariance matrix turns out to be the following expression
q
Cnf = Cn T (28) tha = ca A+ (1)Cnf (A+ (1))T (33)

where Cn is the (R + 1)st-order covariance matrix and ca is set to minimize the error probability
of the corresponding phase error vector, and is the (Appendix C).

894 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
B. Threshold thLFM A. PFA Thresholds

From the least squares equation (30), the linear For signal detection, there is a threshold thi for
model error " becomes every smoothed spectrogram Ii . The set of thresholds
fthi gFi=1 is defined to meet a specified global PFA or
1 PFAg (eqn. (7)). On the one hand, this requirement
"= (f AA+ fi )T (fi AA+ fi ): (34)
R i is not fulfilled for a unique set of thresholds; hence,
It can be equivalently written as a quadratic form of the additional condition that the local PFA (eqn. (8))
is the same for all the smoothed spectrograms is
the frequency error vector (nf ), i.e.,
imposed. On the other hand, an analytical expression
1 T relating the local PFA and PFAg is rather difficult to
"= n (I AA+ )T (I AA+ )nf : (35) find, since the points of the smoothed spectrograms
R f
are correlated, and this correlation depends on the
Therefore, the mean of " becomes time-frequency processor configuration. Therefore, the
correspondence between the local PFA and the PFAg is
1
" = Tr((I AA+ )T (I AA+ )Cnf ) (36) determined by simulation. Let the local PFA value be
R
denoted p.
where Tr means the trace operator.
1) The local PFA axis is explored for different
Diagonalizing Cnf , Cnf = VVT , and considering
values of p.
the unitary transformation nf = VT vf , " can be 2) For each value p, the thresholds fthi gFi=1
expressed as corresponding to every smoothed spectrogram are
1 T analytically obtained by Marandas method [18].
"= v V(I AA+ )T (I AA+ )VT vf : (37) 3) The ADCRx is simulated with an adequate
R f
number of Monte Carlo realizations by using only
Thus, the variance of " becomes noise at the input (only the time-frequency processor
and the detection system must be simulated). By
2
"2 = E["2 ] 2" = f(diag())T Bu Bu diag()g means of those Monte Carlo simulations, the PFAg
R2 corresponding to this value of p can be estimated.
(38)
where Bu is the unitary-transformed matrix Once the curve PFAg versus p is constructed, the
required value of p for a specified PFAg is obtained,
Bu = V(I AA+ )T (I AA+ )VT (39) and so the required thresholds fthi gFi=1 .
For low-sidelobe analysis windows and decimation
Bu Bu means element-by-element multiplication, and factors (M) close to the maximum (K=2), it has been
diag() is the diagonal of the eigenvalue matrix. found by simulation that the smoothed spectrogram
Using (36) and (38), the threshold becomes coefficients are nearly independent; a low-sidelobe
window approximately guarantees independence
thLFM = " + cLFM " : (40) among frequency channels, and a high decimation
factor approximately provides independence in time.
Constant cLFM is set to minimize the error probability Thus the following analytical correspondence between
(see Appendix C). Both tha and thLFM depend on the PFAg and the local PFA value (p) holds:
d
signal SNR and duration, so that the estimates SNR
d F
X
and PW will be used in practice. K LF
PFAg 1 p (41)
2 Li
i=1
APPENDIX C. SIMULATION-BASED THRESHOLD
and the above-described simulation procedure is
COMPUTATION
avoided in this case.
This appendix briefly describes the computation
of some of the most important thresholds utilized B. Thresholds for Duration-Based Signal Classification
by the ADCRx. In particular, it is explained the
The thresholds treated in this and the following
procedure to set the thresholds for signal detection on
section have been obtained following a common
the different smoothed spectrograms (Section IIIA) so
methodology. Every threshold is used to distinguish
that a given global PFA can be met; the thresholds for
between two hypotheses (hypotheses Ha and Hb ) in
the classification of signals according to their duration the framework of a hypothesis test:
(Section IVA); and the constants ca (of threshold tha )
Hb
and cLFM (of threshold thLFM ), and the threshold thf (x) ? th (42)
of the AMC (Section IVB and Appendix B). Ha

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 895
where (x) is a given statistic of the input signal x to For the thresholds fthPAi1,i gFi=2 , we have used
be tested and th is the threshold to be determined. nonmodulated pulses as representatives 1) for the sake
To obtain its value, we follow the Bayes criterion, of simplicity, and 2) due to the fact that the errors in
i.e., minimize the probability of error. Since the the PA estimation for modulated signals do not differ
probability of error depends on the signal and the much from the errors for nonmodulated pulses if their
SNR considered, several signals representing each bandwidths are limited to a channel (see Fig. 9). Thus,
hypothesis (Ha and Hb ) are evaluated for different they would give rise to similar probabilities of error
SNR values. For each signal, the probability of error and thresholds.
is defined as the mean of the probability of error
computed at different SNR values. The probability
of error at a given SNR is the probability that the C. AMC Constant Determination
signal is misclassified when it holds this particular
SNR. This probability is computed by Monte Carlo The computation of the constants cLFM and ca , and
analysis. Once the probability of error for every the threshold thf is similar to the method presented
selected signal is obtained (after averaging for the in the previous section.
designated SNR values), the error probability for each The constant cLFM is part of the threshold to
hypothesis is the mean of the probability of error of distinguish LFM-modulated and nonmodulated signals
all the signals representing this hypothesis. That is, from PSK- and FSK-modulated ones (Section IVB).
the signals representing a hypothesis are assumed This threshold is defined as: thLFM = " + cLFM " ; the
to be equiprobable. Finally, both hypotheses are values for " , " were obtained in Appendix B. The
assumed to be equiprobable as well; therefore, the corresponding hypothesis test is
total probability of error is computed as the mean of Hb : PSK or FSK
the probability of error of both hypotheses. To obtain " ? thLFM (44)
the adequate threshold value, the total probability Ha : LFM or No Modulation

of error is obtained for a range th , and the one where " is the error of the least-squares linear
corresponding to a minimum probability of error is fit of the frequency estimate (Section IVB). The
selected as the desired threshold. representatives of the hypothesis Ha are nonmodulated
In the case of the thresholds fthPAi1,i gFi=2 , they and chirped pulses with different durations; in the
are used to separate signals of similar duration to the case of the chirped pulses, their bandwidth is equal
length of the smoothed spectrogram Ii from signals of to the channel bandwidth. For the ADCRx in Table I
shorter duration according to the hypothesis test the nonmodulated and chirped pulses have 1024
c (the block length) and 500 samples duration. For
PA H : Shorter than Ii
i1 b
? th : (43) lower durations, the signal length estimation would
c Ha : Duration as Ii PAi1,i
PA be smaller than the minimum length to perform the
i
modulation analysis (see in Table II the minimum
For each threshold thPAi1,i , the signal representing d > SNR =
the hypothesis Ha is a nonmodulated pulse lasting length plus the guard samples when SNR th
K + M(Li 1) samples; K, M, and Li are the total 3 dB); the signal would then be classified as a
number of channels, the decimation factor and the nonmodulated pulse. The pulses are centered in
smoothed spectrogram integration length, respectively. time in the block, and their mean frequency is the
The parameter K accounts for the effective length of channel center frequency. We have selected four SNR
the analysis window (the definition and motivation values: 5, 10, 15, and 20 dB over the SNR at which
of the effective length can be found in Appendix A). a single DIFM features a 5% error with respect to
The representative of the hypothesis Hb is also a the channel bandwidth for a sinusoid (in this case,
nonmodulated pulse of duration: K + M(Li1 1) this SNR value is 0 dB approximately according
samples; Li1 is the integration length of the smoothed to Fig. 6). As the hypothesis Hb representatives,
spectrogram Ii1 . Note that Li is greater than Li1 , BPSKs and MSKs with smaller bandwidth than the
according to the definition of the time-frequency channel are chosen. The BPSKs symbol periods are
processor. The frequency of both pulses is centered in 64 samples (3 dB bandwidth equal to the channel)
a channel and their respective probabilities of error are and 1024 samples (the full block length; but at least
computed for only one SNR value: the one assuring one bit transitions must be assured in within the
a 90% detection probability for the shortest pulse. block). The MSKs feature symbol durations of 64
Below this value of SNR, the errors in the estimation samples (approximately 99% of the signal energy
of the PA tend to be very high, and it is not possible within the channel [25]) and 512 samples (separation
to achieve low probability of errors. For greater SNR in frequency equal to 1=1024, which would be the
values, the errors in the PA estimate decrease, so channel bandwidth if the whole processing were
that the probability of error will decrease as well and coherent, i.e., by a 1024-point DFT). The signals
become very small. are present during the whole sample block and their

896 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005
carrier frequencies are centered in a channel. The SNR [7] Fields, T. W., Sharpin, D. L., and Tsui, J. B. Y.
values are the same as in the hypothesis Ha . Digital channelized IFM receiver.
In IEEE Microwave Theory and Techniques Symposium
The constant ca is associated with the threshold Digest, Vol. 3, San Diego, CA, 1994, 16671670.
that helps to distinguish LFM-modulated from [8] Zahirniak, D. R., Sharpin, D. L., and Fields, T. W.
nonmodulated signals (Section IVB). This threshold is Hardware-efficient, multirate, digital channelized receiver
defined as: tha = ca a (see Appendix B for a definition architecture.
of a ), and the corresponding hypothesis test is IEEE Transactions on Aerospace and Electronic Systems,
34 (Jan. 1998), 137151.
Hb : LFM [9] Tsui, J. B.
a ? tha : (45) Microwave Receivers with Electronic Warfare Applications.
Ha : No Modulation New York: Wiley, 1986.
[10] Lo pez-Risuen o, G., Grajal, J., Yeste-Ojeda, O. A.,
The representatives of the hypothesis Ha are the Sanz-Osorio, A., and Moreno, J. A.
nonmodulated pulses used in the computation of Two digital receivers based on time-frequency analysis
cLFM , and the representatives of the hypothesis Hb for signal interception.
In International Conference on Radar (Radar 2003),
are the chirped pulses also used to compute cLFM
Adelaide, Australia, Sept. 2003, 394399.
plus two chirped pulses both with bandwidth 1=1024 [11] Baraniuk, R. G., and Jones, D. L.
(the channel bandwidth if the whole processing were A signal-dependent time-frequency representation:
coherent, i.e., if a 1024-point DFT were used) and Optimal kernel design.
duration 1024 and 500, respectively. These new IEEE Transactions on Signal Processing, 41 (Apr. 1993),
signals represent the smallest LFM modulation that 15891602.
[12] Bultan, A.
should be detected. The SNR values are the same as
A four-parameter atomic decomposition of chirplets.
for the constant cLFM . IEEE Transactions on Signal Processing, 47 (Mar. 1999),
Threshold thf is defined as a constant for the 731745.
sake of simplicity, and is utilized to distinguish PSK- [13] Mallat, S.
from FSK-modulated signals (Section IVB). The A Wavelet Tour of Signal Processing.
corresponding hypothesis test is New York: Academic Press, 1998.
[14] Oppenheim, A. V., and Schafer, R. W.
Hb : PSK Discrete-Time Signal Processing.
max jf(n)j ? thf (46) Englewood Cliffs, NJ: Prentice-Hall, 1989.
Ha : FSK [15] Mallat, S. G., and Zhang, Z.
Matching pursuits with time-frequency dictionaries.
where f(n) is the first-order difference of the DIFM IEEE Transactions on Signal Processing, 41 (Dec. 1993),
sequence. The representatives of the hypothesis Hb 33973415.
are the same BPSK signals used to compute cLFM , and [16] Trunk, G. V.
the representatives of the hypothesis Ha are the MSK Comparison of the collapsing losses in linear and
square-law detectors.
signals also used to compute cLFM . The SNR values Proceedings of the IEEE (June 1972), 743744.
are the same as well. [17] Mills, R. E., and Prescott, G. E.
A comparison of various radiometer detection models.
IEEE Transactions on Aerospace and Electronic Systems,
REFERENCES 32 (Jan. 1996), 467473.
[18] Maranda, B. H.
[1] Fuller, K. L. On the false alarm probability of an overlapped FFT
To see and not be seen. processor.
IEE Proceedings Pt. F, 137 (Feb. 1990), 19. IEEE Transactions on Aerospace and Electronic Systems,
[2] Stephens, J. 32 (Oct. 1996), 14521456.
Advances in signal processing technology for electronic [19] Priestley, M. B.
warfare. Spectral Analysis and Time Series.
IEEE Aerospace and Electronic Systems Magazine, (Nov. New York: Academic Press, 1981.
1996), 3138.
[20] Kay, S.
[3] Tsui, J. A fast and accurate single frequency estimator.
Digital Techniques for Wideband Receivers. IEEE Transactions on Acoustics, Speech and Signal
Dedham, MA: Artech House, 1995. Processing, 37 (Dec. 1989), 19871990.
[4] Tsui, J., and Stephens, P. J. [21] Mengali, U., and DAndrea, A. N.
Digital microwave receiver technology. Synchronization Techniques for Digital Receivers.
IEEE Transactions on Microwave Theory and Techniques, New York: Plenum Press, 1997.
50 (Mar. 2002), 699705. [22] Azzouz, E. E., and Nandi, A. K.
[5] Hlawatsch, F., and Boudreaux-Bartels, G. Automatic Modulation Recognition of Communication
Linear and quadratic time-frequency signal Signals.
representations. Boston: Kluwer Academic Publishers, 1996.
IEEE Signal Processing Magazine, (April 1992), 2167. [23] Ho, K. C., Prokopiw, W., and Chan, Y. T.
[6] Qian, S., and Chen, D. Modulation identification by the wavelet transform.
Joint Time-Frequency Analysis. Methods and Applications. In IEEE Military Communications Conference (MILCON
Englewood Cliffs, NJ: Prentice-Hall, 1996. 95), Vol. 2, San Diego, CA, Nov. 1995, 886890.

LOPEZ-RISUE ET AL.: DIGITAL CHANNELIZED RECEIVER BASED ON TIME-FREQUENCY ANALYSIS
NO 897
[24] Skolnik, M. [27] Sa nchez-Marcos, M. A., Garrido-Ga lvez, M., Lo pez-Vallejo,
Radar Handbook (2nd ed.). M. L., and Grajal, J.
New York: McGraw-Hill, 1990. Implementing the FFT algorithm on FPGA platforms: A
[25] Proakis, J. G. comparative study of parallel architectures.
Digital Communications (3rd ed.). In XIX International Conference on Design of Circuits and
New York: McGraw-Hill, 1995. Integrated Systems (DCSI 2004), Bourdeax, France, Nov.
[26] Liedtke, F. 2004.
Computer simulation of an automatic classification [28] Dy4 Systems
procedure for digitally modulated communications signals Digital Signal Processors Catalog (July 2004).
with unknown parameters. http://www.dy4.com/products/dsp/index.htm.
Signal Processing, 6 (Aug. 1984), 311323.

Gustavo Lo pez-Risueno (SM99M03) was born in Barcelona, Spain, in

1974. He received the Ingeniero de Telecomunicacio n and the Ph.D. degrees
from Universidad Polite cnica de Madrid, Madrid, Spain, in 1998 and 2002,
respectively.
From 2000 to 2003, he was an assistant professor at the Department of
Signals, Systems and Radiocommunications, Universidad Polite cnica de Madrid,
working in the areas of statistical signal processing for active and passive radars,
and passive radar design and technology. He spent the fall of 2000 and the
summer of 2002 as a visiting researcher at the Adaptive System Lab, McMaster
University, Hamilton, Canada. Since 2004, he has been with the European Space
Research and Technology Centre (ESTEC) of the European Space Agency
(ESA), Noordwijk, Netherlands, working on advanced receivers for global
navigation satellite systems. His current research interests include statistical signal
processing, spread spectrum communications, and global navigation satellite
systems.

Grajal was born in Toral de los Guzmanes (Leo n), Spain, in 1967. He
Jesus
received the Ingeniero de Telecomunicacio n degree and the Ph.D. degree from the
Technical University of Madrid, Madrid, Spain in 1992 and 1998, respectively.
Since 2001 he has been associate professor at the Signals, Systems, and
Radiocommunications Department of the Technical School of Telecommunication
Engineering of the same university. His research activities are in the area of
hardware design for radar systems, radar signal processing, and broadband digital
receivers for radar and spectrum surveillance applications.

Alvaro Sanz-Osorio was born in Madrid, Spain, in 1979. He received his
telecommunication engineering degree from the Technical University of Madrid
(UPM) in 2003, where he carried out his master thesis on the design of a digital
radar receiver at the Microwaves and Radar Group (GMR). His research activities
are focused on the area of signal processing for radar receivers.

898 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 3 JULY 2005