REPORT DOCUMENTATION PAGE Form Approved OMB No.

0704-0188

Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing the burden, to Department of Defense, Washington Headquarters Services, Directorate for Information Operations and Reports (0704-0188), 1215 Jefferson
Davis Highway, Suite 1204, Arlington, VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to any penalty for failing to comply
with a collection of information if it does not display a currently valid OMB control number.
PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS.
1. REPORT DATE (DD-MM-YYYY) 2. REPORT TYPE 3. DATES COVERED (From – To)
21-12-2004 Final Report 4 March 2003 - 04-Jan -05
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
FA8655-03-1-3030
Detection and Characterization of Phase-Coded Radar Signals

5b. GRANT NUMBER

5c. PROGRAM ELEMENT NUMBER

6. AUTHOR(S) 5d. PROJECT NUMBER

Dr. Ernest R. Adams

5d. TASK NUMBER

5e. WORK UNIT NUMBER

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION

Cranfield University (RMCS) REPORT NUMBER
Shrivenham SN6 8LA
United Kingdom N/A

9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)

EOARD
PSC 802 BOX 14
FPO 09499-0014 11. SPONSOR/MONITOR’S REPORT NUMBER(S)
SPC 03-3030

12. DISTRIBUTION/AVAILABILITY STATEMENT

Approved for public release; distribution is unlimited.

13. SUPPLEMENTARY NOTES

14. ABSTRACT

This report results from a contract tasking Cranfield University (RMCS) as follows: The Grantee will investigate various mathematical
concepts of detection and characterization of simulated radar signals. As detailed in the technical proposal, at least four approaches will be
investigated: 1. A varied matched filter and spectral analysis. 2. A neural network (NN) approach, based on the use of NNs to detect pulse
radar. 3. Methods derived from CDMA (code division multiple access) multiuser detection. 4. Independent component analysis (ICA). He will
employ MATLAB and SIMULINK to generate signals and MATLAB for their analysis. Efficiency of algorithms will be an important
consideration in assessing the performance of the mathematical approaches developed.

15. SUBJECT TERMS

EOARD, Signal Processing, radar

16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18, NUMBER 19a. NAME OF RESPONSIBLE PERSON
ABSTRACT OF PAGES PAUL LOSIEWICZ, Ph. D.
a. REPORT b. ABSTRACT c. THIS PAGE UL
UNCLAS UNCLAS UNCLAS 21 19b. TELEPHONE NUMBER (Include area code)
+44 20 7514 4474
Standard Form 298 (Rev. 8/98)
Prescribed by ANSI Std. Z39-18
CONTRACT TITLE: Detection and Characterization of Phase-coded
Radar Signals

SPONSOR: AFRL/SNRP, WPAFB through EOARD, London

SPONSOR’S REFERENCE: 033030

FINAL REPORT: Research carried out up to December 2004 by

Dr E R Adams, Cranfield University (RMCS)
1. Summary

Relatively simple correlation/matched filter techniques were developed to detect and

identify phase-coded radar signals based on true Barker codes and longer, combined
Barker codes. Detection and identification rates were good, though not so good as
those possible with more sophisticated techniques like neural networks. However, the
simplicity and computational efficiency of the new methods make them worthwhile
for some applications.

2. Description of the problem

The problem is to detect and characterize pulsed phase-coded radar signals in noise.
Gaussian noise is assumed, although further work would consider non-Gaussian noise
and interference from DS/SS-based CDMA transmissions. Partial knowledge of the
radar signal is assumed, in particular the carrier frequency and the type of pseudo-
random sequence used to determine phase. Barker codes are investigated first but, as
the longest known true Barker code is only 13 bits and short codes require high power
levels for a given energy, longer codes created from true Barker codes are also
considered. The extended problem of detecting multi-phase Barker codes, other
pseudo-random codes like m-sequences and forms of wideband interrupted CW was
not attempted.

1
3. Properties of phase-coded radar signals

A medium/long-range surveillance radar typically requires a range resolution ∆R of

15 m. This ∆R implies a waveform RF bandwidth B of 10 MHz. From Gabor’s
theorem, bit duration is 1/B = 10−7s. Assuming the 13-bit Barker code, the pulse in
Figure 1 results.

1
= 10 −7 s
B

τ = 1.3 × 10 −6 s

Figure 3.1: Radar pulse for a 13-bit Barker code

Time delay td is related to range-to-target R by the equation

2R
ct d = 2 R ⇒ t d = , (3.1)
c

so time-delay resolution ∆td is related to range resolution ∆R by the equation:

2∆R c∆t d
∆t d = ⇒ ∆R = (3.2)
c 2
But ∆td is known to be 1/B, giving range resolution

c∆t d c
∆R = = (3.3)
2 2B

Thus B = 10 MHz gives the desired range resolution of 15 m. Pulse duration τ results
from the choice of the 13-bit Barker code.

The pulse-repetition interval T, illustrated in Figure 3.2, is a more fundamental

variable than τ. It is determined by the maximum unambiguous range, Ru:

2 Ru
T= (3.4)
c

τ should then be chosen to satisfy 3τ ≤T in order to avoid eclipsing. In practice, for

short pulses like the 13-bit Barker code, T >> 3τ. Greater efficiency results (power
reduced) if τ approaches T/3 (τ = T produces CW), which provides the motivation for
examining the longer codes in section 6.

2
 τ

T T

Figure 3.2: Train of regularly repeating pulses

The length of the pulse train is t0 = nT, i.e. the train has n pulses. The variable n is
determined by the requirement that 2E/N0 at maximum range must be 20 dB, where
E = Sτn, S being signal power. This may alternatively be stated in the relation
between output and input SNRs:

⎡S⎤
= 2n(Bτ )⎢ ⎥ ,
2E
(3.5)
N0 ⎣N ⎦

i.e. 2nBτ is the integration improvement factor.

Finally, Ts is the interval between pulse trains, satisfying:

Ω
Ts = t0 (3.6)
Ωb

where Ω is the solid angle of search and Ωb is the solid angle of the radar beam. In
practice, the choice of Ts is more arbitrary than for other parameters. In simulations,
Ts was initially assumed to be constant. Later, it was assumed to be uniformly
distributed between Ts − t0/2 and Ts + t0/2.

The following values were used in initial simulations:

1
B = 10 MHz, (bit length) = 10−7 s
B
τ = 1.3 × 10−6 s (for 13-bit Barker code)

τ = m × 10−7 s for pseudo-random m-bit code

2 Ru
T = 7 × 10−4 s, based on Ru = 100 km, T =
c
n = 20 ⇒ t0 = nT = 1.4 × 10−2 s

Ts = 3 s (constant)

Ts = 3 + (rand − 0.5)t0, where rand is the Matlab function returning a

uniformly distributed value between 0 and 1.

3
4. Barker and other pseudo-random codes for phase coding

4.1 True Barker codes

The nine known Barker codes are represented in Table 4.1.

Length N cn
2 ++
2 −+
3 ++−
4 ++−+
4 +++−
5 +++−+
7 +++−−+−
11 + + + − − − +− − + −
13 +++++−−++−+−+

Table 4.1: Barker codes

Barker codes have autocorrelation functions (ACFs) with equal time sidelobes: they
are the only codes for which ACF sidelobes at zero Doppler have levels ≤N (code
length). As explained in section 2, they are too short for some applications – the
greatest sidelobe reduction, for N = 13, is −22.3 dB. However it is possible to
combine Barker codes to create longer codes with good properties.

4.2 Combined Barker codes

This method involves phase coding with one Barker code within each segment of
another Barker code. For example, the Barker code of length 3, B3, may be combined
with that of length 5, B5, to create a code of length 15 (Figure 4.1):

B53 = (+ + + − +, + + + − +, − − − + −) (4.1)

Note that Bnm is the Kronecker product Bn⊗Bm , regarding sequences as row vectors.

Figure 4.1: Combined Barker Code B53

The order in which the two codes are combined is significant, i.e. Bn,m ≠ Bm,n. [1,2]
give examples of the different code properties of the alternative orderings. E.g. B4,13,
for which each bit of the 13-bit code B13 is coded into four bits, has a zero Doppler
ACF with four side peaks of amplitude 13 at shifts ±1 and ±3, and 12 peaks of
amplitude 4. B13,4 has the same number of side peaks of amplitude >1, but the side
4
peaks of amplitude 13 occur at shifts of ±13 and ±39. The main peak of the ACF in
both cases has amplitude 52. B4,13 is useful if expected interference is well separated
in range from the target.

Combining pairs of Barker codes simply in this way can give codes up to
11 × 13 = 143 bits long. The properties of 3-way combinations, e.g. 3 × 5 × 7, are not
known. Note that Bnn is not used because of high partial correlations.

4.3 n-Phase Barker codes

n-phase Barker sequences, e.g. using phase alphabets such as multiples of π/6, are
defined and their properties explored in [3,4]. They will be investigated if time
allows.

5
5. Methods for detection and characterization of Barker sequences

5.1 Correlation/pattern recognition techniques

These newly developed techniques begin with calculating the ACF of a length Tp of
signal, where Tp > n′T′, the highest reasonable value of the unknown nT, and Tp << Ts,
i.e. Tp is selected so there is a high probability it will include a single pulse train of n
pulses of length T. FFTs are used to estimate the ACF to improve efficiency.
Figure 5.1 illustrates the ACF for a noiseless pulse train of B3 pulses.

…
−T 0 T (n − 1)T nT

Figure 5.1: ACF of 3-length code

Pattern recognition algorithms are under development to recognize the repeating

waveform, linearly decreasing in amplitude at multiples of |T|, with the aim of
estimating T. Alternatively, pattern recognition may be carried out in the frequency
domain. Assuming signal x(t) is sampled at intervals of ∆t, i.e. xi = x(i∆t), and
T = M∆t, then the pulse train of n Barker codes bi, 0 ≤ i ≤ N − 1, may be written:

n −1
yi = bi * ∑ δ (i − kM ) (5.1)
k =0

The DFT of the pulse train yi is thus the product of the DFT of the Barker code with
the DFT of the finite train of impulse functions. If n → ∞, a line spectrum of the
Barker code results, with spectral line spacing 1/T Hz. For finite n, this spectrum is
convolved with sinc (nTf), i.e. a very narrow sinc function with zeros at integer
multiples of 1/nT. Assuming T = M∆t and Tp = P∆t, the frequency resolution of the
FFT for window Tp is 1/P∆t, so the spectral lines at 1/M∆t will be sampled only if M
divides P. This is illustrated in Figure 5.2. Algorithms are being developed to
estimate T from the DFT or power spectrum of xi.

6
 Envelope of FT[code]

Sampling at
intervals of
1
P∆t

0
1
f
1 M∆t
P∆t ⎛1⎞
⎜ ⎟
⎝T ⎠
1
nM∆t

Figure 5.2: Fourier transform of signal with Tp-length window

The estimate Tˆ , either from the ACF or FFT method, is used to identify the unknown
waveform by dividing the interval Ta into K segments of length Tˆ and averaging (K is
the largest integer for which KTˆ ≤ Ta ). Assuming Tˆ = M ′∆t ,

1 K −1
ui = ∑ xi + kM ′ 0 ≤ i ≤ M −1 (5.2)
K k =0

The averaging reduces noise and allows easier detection of the single Barker
waveform in ui with a matched filter or neural network.

5.2 Simulation results for correlation techniques

The performance of a combination of the techniques in 5.1 was assessed in initial

simulations. To reduce computational effort, T was taken to be 7 × 10−5 s. Otherwise,
the values in section 3 were used. A two-stage method detected the presence of a
repeating pulse-waveform and estimated T, the pulse-repetition interval. Because of
the high cost of ACF estimation, detection and initial estimation of T were based on
the signal’s Fourier spectrum. This allowed a more precise estimate of T, following
detection and initial T estimation, from a narrow search of ACF values. The refined T
estimate was then used to produce an averaged pulse-waveform, i.e. one with greatly
reduced noise. The final stage was to identify the waveform by correlation with
Barker codes – a matched filter approach. In more detail, the stages were:

7
(1) As described in 5.1, the DFT of a signal containing a repeating waveform
consists of narrow sinc functions spaced at 1/T Hz intervals, the sinc amplitudes
depending on the waveform’s Fourier transform. The FFT Xk of the P-length signal
(same symbols used as in 5.2) was used to test for repeating sinc functions by finding
the value i = r which maximizes

X i + X 2i + X 3i + L + X kii
Yi = (5.3)
ki
So

i ≠ r ⇒ Yr ≥ Yi (5.4)

This Yr was then tested against a threshold to detect the presence of a repeating
waveform. If a repeating waveform was detected the following approximate estimate
of T was calculated:

P∆t
Tˆ1 = (5.5)
r

The estimate is not exact as, in general, 1/T is not a sampled frequency of the FFT, i.e.
the sinc functions’ centres are not sampled.

Detection rates for B13 pulse-trains in alternative noise environments are given in
table 5.2. False alarm rates, i.e. detection rates when no pulse-train is present, are also
given.

(2) Given Tˆ1 from the first stage (5.5), a more accurate estimate of T (and one that
may be exact) may be derived by examining ACF values for shifts close to Tˆ1 . For
such a corresponding m, a shift index for which m∆t ≅ Tˆ , the ACF is
1

C m = E[xi xi + m ] , (5.6)

estimated by averaging xixi+m over all possible i (1 ≤ i ≤ P − m). For m values

satisfying

m∆t − Tˆ1 ≤ ∆T , (5.7)

∆T being determined by trial and error, the following statistic was maximized:

C m + C 2 m + C 3m + L + C k m m
Dm = (5.8)
km

For M ′ corresponding to this maximum, i.e.

m ≠ M ′ ⇒ D M ′ ≥ Dm (5.9)

8
the final estimate for T becomes:

Tˆ = M ′∆t (5.10)

(3) Tˆ was then used to produce a waveform with reduced noise by averaging
signal segments of length Tˆ (5.2). The waveform was identified by correlating the
averaged segment with alternative Barker codes. A match was found if the maximum
correlation exceeded a pre-determined threshold. Although only B13 signals were
simulated, the averaged segment was also correlated with B11 and false detection rates
(deciding B11 was present instead of the actual B13) were estimated. Also estimated
were failure rates to detect any pulse-train when B13 was present. The identification
results may be seen in table 5.3. The confusion between B13 and B11 may be
explained by their high cross-correlation:

CC j (13,11) = ∑ B13 (i ) B11 (i + j )

i
= 0L 0 − 1 0 − 1 − 2 − 1 0 − 3 − 4 1 0 − 3 7 3 0 − 1 − 4 3 0 1 − 2 1 0 1 0L 0

(5.11)

More sophisticated pattern recognition algorithms are being developed to distinguish

Barker codes. The cross-correlation properties of Barker codes are also under
investigation.

Noise % Detection % False alarms

6 dB 96 2
3 dB 86 8
0 dB 62 42

Table 5.2: Detection/false alarm rates for B13

% Failure to
Noise % Identification (B13) % False identification (B11)
identify
6 dB 100 4 0
3 dB 98 12 0
0 dB 90 26 6

Table 5.3: Identification rates for B13

Each estimated % in tables 5.2 and 5.3 is based on 50 simulations. Note that a 0 dB
noise level corresponds to noise with the same mean amplitude (i.e. 1) as the Barker
code. However, the noise persists throughout the signal at this level. 3 dB
corresponds to a mean noise amplitude of 0.707 (power 0.5) and 6 dB to 0.5 (power
0.25)

9
Note also that a radar receiver needs an output SNR (i.e. 2E/N0) of 100 (20 dB), and

= n(Bτ ) .
E S
(5.12)
N0 N

So the receiver needs

S 100
= . (5.13)
N n(Bτ )

In the above case, n = 20 and Bτ = 13, so the required S/N = 0.38, i.e. a mean-
amplitude ratio of 0.62.

10
6. Methods for detection and identification of combined Barker sequences

6.1 Correlation/matched-filter techniques

Combined Barker sequences may be expressed as the Kronecker product of a pair of

Barker sequences. This is worthwhile as the Matlab routine ‘kron’ may be used to
produce combined sequences. The Barker sequences of lengths r and s may be
represented by the vectors:

B r = (α i ) = (α 1 α 2 Kα r )
( )
B s = β j = (β 1 β 2 K β s )
(6.1)

A combined Barker sequence Brs is produced by the Kronecker product Br ⊗ Bs , i.e.

taking Bs = (βj) and replacing each entry βj by the r-vector βj(αi):

B rs = B r ⊗ B s = (γ k )
= (β 1 (α i ) β 2 (α i ) K β s (α i ) ) (6.2)
= (β 1α 1 β 1α 2 K β 1α r β 2α 1 β 2α 2 K β 2α r . . . β sα 1 β sα 2 K β sα r )

Correlating the combined Barker sequence Brs with the inner sequence Br produces
the matched-filter output:

1 r −1
ck = ∑ α j +1γ k + j (6.3)
r j =0

Note that ck produces βj values every r–th value of k:

c1 =
1
r
(
β 1 α 12 + α 22 + K + α r2 )
c1+ r
1
(
= β 2 α 12 + α 22 + K + α r2
r
) (6.4)
M

c1+( s −1) r
1
(
= β s α 12 + α 22 + K + α r2
r
)
As, ∀i, α i = ±1 ⇒ α i2 = 1, all the (K) terms in (6.4) equal r , giving :

c1+ jr = β j +1 , 0 ≤ j ≤ s −1 (6.5)

For all other ck terms:

1
k ≠ 1 + jr ⇒ c k ≤ (6.6)
r

11
Thus, correlating every r-th value of (ck) with the outer sequence Bs produces the
maximum matched-filter output:

1 s −1
cc1 = ∑ β j +1c1+ jr
s j =0
1 s
= ∑ β 2j , from 6.5 (6.7)
s j=1
= 1 , as β 2j = 1

The above suggests a 2-dimensional search for the presence of Brs in a sample.
Successively searching for Br and Bs is more efficient than directly searching for Brs,
i.e. for the presence of a repeating (γk) in the sample. Correlating with an r-length
sequence and an s-length is clearly more efficient than correlating with one of length
rs. Furthermore, it is possible to estimate r before going on to estimate s. Because of
poor cross-correlation properties of some Barker sequences, initial correlation of the
sample with an incorrect Br may lead to a weak repeating pattern, increasing the
likelihood of false identification.

6.1.1. Estimation of Bj components

For a particular Br and sample Y=(yi), the following matched filter results:

1 r −1
ck = ∑ α j +1 y k + j , k ≥1 (6.8)
r j =0

The first stage in estimating the outer sequence is to correlate alternative Bs with ck
values at k-intervals of r , producing the ccks sequence (matched-filter output):

1 s −1
ccks = ∑ β j +1c k + jr , k ≥1 (6.9)
s j =0

If the correct Bs is used, the maximum value of 1 is produced when that Bs is aligned
with identical r-spaced c values, as in (6.7), which happens once for each Brs
waveform (every M-th value of k in (6.9), T = M∆t). Thus, for noiseless (yi):

max [cck ] = 1
s
(6.10)
k

and, for other k ′ :

1 1 s −1 1
| c k ′+ jr |≤ and β j = ±1 ⇒ ccks′ = ∑ β j +1c k ′+ jr ≤ (6.11)
r s j =0 r

12
Tests for Bs exploit the expected pulse train of 1’s in ccks for the correct s. The train
of n pulses at intervals of T (=M∆t), beginning at mT, may be represented:

m + n −1 ⎧t − t0 ⎫ ∞
∑ δ (t − iT ) = rect ⎨ ⎬ ∑ δ (t − iT )
i =m ⎩ nT ⎭i = −∞
(6.12)
⎛ n − 1⎞
t0 = ⎜ m + ⎟T
⎝ 2 ⎠

Its Fourier transform is:

⎧m + n −1 ⎫ ∞ ⎛ i⎞
FT ⎨ ∑ δ (t − iT )⎬ = ne − j 2πft0 sinc(nTf ) * ∑ δ ⎜ f − ⎟ (6.13)
⎩ i =1 ⎭ i = −∞ ⎝ T⎠

Its amplitude spectrum is thus:

∞ ⎛ i⎞
FT{ } = n sinc(nTf ) * ∑ δ ⎜ f − ⎟ , (6.14)
i = −∞ ⎝ T⎠

1 2
i.e. a series of absolute sinc functions centred on frequencies 0 Hz, ± Hz, ± Hz
T T
1 2
etc. The sinc centred on 0 Hz has zero values at frequencies ± , ± Hz etc,
nT nT
1 1
where << . Tests for Bs may thus be based on the amplitude spectrum of ccks
nT T
- a phase independent test, i.e. m-independent.

An alternative and simpler test estimates M by taking average of ccks values at

p-intervals of k. Different values for p are tried within the likely range for M, which
may be derived by examining the amplitude spectrum of ccks as above, or may be
known a priori. The p-interval averages are:

1 L p −1 s N
ccsip = ∑ cci + jp , L p ≤ (6.15)
L p j =0 p

where N is the number of available y samples. The largest of the p possible values
(i = 1,2,…, p) is taken for each possible p and the overall maximum over p is
compared with a threshold to detect s:

max [ccsi ] ≥ threshold ?

p
(6.16)
p ,1≤i ≤ p

13
A final alternative is to test the series resulting from taking every r-th value of the ck
in (6.8) using the techniques of section 5, i.e. test for a true Barker code.

Tests for a single true Barker code or a combined Barker code in the sample may be
combined. Using an initial Br matched filter may result in a pulse train of 1’s or a
recurring pattern associated with a further Bs. Thus it is possible to decide there is
just a repeating true Barker code Br present or go on to detect the additional Bs of a
combined code.

6.1.2. Effects of additive Gaussian noise

Applying the matched filter of (6.8) to a sample consisting of a signal (xi) in Gaussian
noise (ni):

y i = x i + ni , (6.17)

where ni are independent samples from a zero-mean Gaussian distribution of variance

σ2, produces

ck =
1 r −1
(
∑ α j +1 x k + j + nk + j
r j =0
)
(6.18)
1 r −1 1 r −1
= ∑ α j +1 x k + j + ∑ α j +1nk + j
r j =0 r j =0

As αj+1 = ±1, the noise in ck has zero-mean and variance σ2/r. This effect is repeated
for ccks in (6.9), which has a variance of σ2/rs. This clear aids detection, particularly
for longer sequences, as does the smaller value for cc when βj is not perfectly aligned
(6.11).

6.2 Simulation results for matched filter/correlation techniques

As an illustration, identification rates are given for B5,3 and B13,7. ck and ccks were
calculated as in (6.8, 6.9) and tests for Bs used (6.15). M was assumed to be in the
range 600-800, when the actual value was 700.

% Failure to
Noise % Identification % False identification
identify
6 dB 84 13 11
3 dB 81 14 16
0 dB 70 23 27

Table 6.1: Identification rates for B5,3

14
 % Failure to
Noise % Identification % False identification
identify
6 dB 91 17 7
3 dB 88 21 10
0 dB 79 34 21

Table 6.2: Identification rates for B13,7

Each estimated % in tables 6.1 and 6.2 is based on 100 simulations, with the same
noise assumptions as described in 5.2.

Clearly, identification is better for the longer B13,7, as expected, but false
identification is also higher due to higher cross-correlations for the constituent Barker
codes.

15
7. Alternative methods for Barker sequence detection

7.1 Neural network detection

Neural networks have previously been trained to recognize particular codes [5]. The
current investigation has focussed on a 3-layer perceptron network with 13 input
elements trained to detect B13. The output layer has one neuron, which may be shown
to give a sufficient statistic. All neurons in the hidden and output layers use a sigmoid
function, and neuron thresholds are zero.

Sequence Training input

1 0 0 0 0 0 0 0 0 0 0 0 0 0
2 1 0 0 0 0 0 0 0 0 0 0 0 0
3 −1 1 0 0 0 0 0 0 0 0 0 0 0
4 1 −1 1 0 0 0 0 0 0 0 0 0 0
5 −1 1 −1 1 0 0 0 0 0 0 0 0 0
.
.
.
13 1 1 1 1 −1 −1 1 1 −1 1 −1 1 0
14 1 1 1 1 1 −1 −1 1 1 −1 1 −1 1 B13
15 0 1 1 1 1 1 −1 −1 1 1 −1 1 −1
.
.
.
24 0 0 0 0 0 0 0 0 0 0 1 1 1
25 0 0 0 0 0 0 0 0 0 0 0 1 1
26 0 0 0 0 0 0 0 0 0 0 0 0 1

Table 7.1: Training set, without noise, for B13

Training was carried out with the 26 possible sequences in Table 7.1 with additive
noise, to attempt to get output 1 for sequence 14 and 0 for all other sequences.

[5] has shown that after training with noiseless codes, the signal-to-sidelobe ratio of
this kind of network can achieve 42.7 dB for B13 and 49.7 dB for the 63-bit
m-sequence, much higher than the respective 17 dB and 22 dB for matched filters.
The neural network also performed much better in the presence of non-Gaussian
noise. Current results so far for B13 with additive Gaussian noise have achieved
around 35 dB, still a much superior detector to the conventional matched filter.

Future work will investigate the effects of the number of neurons in the hidden layer
and alternative learning algorithms. Clearly, different networks would be needed for
different codes, and subsequent analysis required to determine other parameters,
e.g. T(or n). An obvious disadvantage of neural networks is their computational
complexity.

16
7.2 Methods derived from CDMA multiuser detection

The pulse-compression technique in phase-coded radar involves the transmission of a

long-duration wide-bandwidth signal code, properties shared by CDMS transmissions.
Thus more severe problems arise when interfering CDMA signals are present. In
these cases, it may be beneficial to use a modified multiuser detector (MUD),
designed to reject multiple access interference (MAI) in CDMA, possibly blindly and
adaptively. MUDs use information about all users’ codes to cancel MAI, i.e. all users’
signals except the wanted user’s: in general, the more information used the better, and
slower, the detection. In this application, all users constitute the MAI and the radar
signal the desired user. MUDs using subspace methods [6] will be investigated: the
basis of this approach is that, by performing an eigendecomposition of the
autocorrelation matrix of signal samples, the signal is separated into radar signal and
CDMA user subspace and the noise subspace.

MUDs exploiting the cyclostationarity and resulting high degree of spectral

correlation of signals will also be investigated. Silent periods between pulse trains
may be exploited to derive the (cyclostationary) CDMA interference model, removing
the need for training sequences [7].

7.3 Independent component analysis (ICA)

This is a statistical technique which is suitable for blind estimation and detection
when signals from multiple sensors are available. ICA finds a linear transformation of
the signal vector to separate it into its component, statistically independent, parts.
Phase-coded radar signals, any CDMA users’ signals present, and noise will be
mutually independent. The following model is assumed for the signal vector y:

y = Mx + n (7.1)

where x is a vector containing the radar signal and CDMA users’ signals, M is a
mixing matrix, and n an additive noise vector. ICA computes a linear transform W
such that the elements of

s = Wy (7.2)

are as statistically independent as possible: these s-elements are estimates of the

independent source signals, i.e. the radar and CDMA users’ signals.

ICA has been used as a MUD for CDMA with little knowledge of even the wanted
signal [8], and also for separating monopulse radar signals [9].

17
8. Conclusions and future work

Only the simple, low-complexity correlation and matched-filter methods of sections 5

and 6 have been investigated in detail. They produce good detection and
identification rates for true Barker codes, less good for combined Barker codes.
Further investigation into the cross-correlation properties of true Barker codes is
needed to derive theoretical performance limits. It seems unlikely that the slightly
improved performance of more sophisticated methods like neural networks would
justify the extra complexity. Neural networks are not real-time methods for this
application.

The techniques of section 6 may in principle be extended to analyse signals

employing Barker-like codes constructed from three true Barker codes, Bu⊗Bv⊗Bw.
E.g B13⊗(B11⊗B7) and (B13⊗B11)⊗B7 are codes of length 1001. However, the
properties of such codes are unknown and need investigation.

Finally, alternative noise environments need to be explored, in particular modified

MAI techniques to cancel interfering signals.

18
9. References

1. E Hollis: Comparison of combined Barker codes for coded radar use, IEEE
Trans Aerospace Electron Sys, AES-3/1, pp 141-143, 1967.

2. F E Nathanson: Radar Design Principles, McGraw Hill, 1969.

3. N Chang and S W Golomb: On n-phase Barker sequences, IEEE Trans on

Information Theory, Vol 40/4, pp 1251-1253, 1994.

4. M Friese: Polyphase Barker sequences up to length 36, IEEE Trans on

Information Theory, 42/4, pp 1248-1250, 1996.

5. H K Kwan and C K Lee: A Neural Network Approach to Pulse Radar

Detection, IEEE Trans on Aerospace and Electron Sys, 29/1, pp 9-21, 1993.

6. X Wang and H V Poor: Blind Multiuser Detection: a Subspace Approach,

IEEE Trans on Information Theory, Vol 44, pp 677-691, 1998

7. S S H Wijayasuriya et al: A Novel Interference Rejection Scheme for DS-

CDMA Using Adaptive Noise Cancellation, ICCS’94, pp 364-358, 1994.

8. T Ristaniemi and J Joutsensalo: Independent Component Analysis With Code

Information Utilization in DS-CDMA Signal Separation, IEEE Globecom’99,
Part A, pp 320-324, 1999.

9. E Chaumette, P Comon and D Muller: ICA-based Technique for Radiating

Sources Estimation: Application to Airport Surveillance, IEE Proc-F, Vol
140/6, 1993.

19
10. Glossary

10.1 Abbreviations

ACF autocorrelation function

CDMA code division multiple access
CW continuous wave
DFT discrete Fourier transform
FFT fast Fourier transform
ICA independent component analysis
MAI multiple-access interference
m-sequence maximal-length sequence
MUD multi-user detector
RF radio frequency

10.2 Symbols

B bandwidth
Bn Barker code of length n
Bn,m combined Barker code from Bn and Bm
c velocity of light
Cm ACF value for shift m
CCj Cross-correlation function value for shift j
E input energy, i.e. S t 0 ( S is the mean input signal power)
m metres
n number of pulses in train
N noise input to receiver, i.e. N0B
N0 input-noise power spectral density (one-sided)
R range to target
∆R range resolution
Ru maximum unambiguous range
s seconds
S input signal power
t0 length of pulse train
td time delay
∆td time-delay resolution
T pulse-repetition interval
Ts interval between pulse trains
τ pulse duration
Ω solid angle of search
Ωb solid angle of radar beam
≅ approximately equal to
* convolution
⊗ Kronecker product, e.g. (a1 a2) ⊗ (b1 b2) = (b1a1 b1a2 b2a1 b2a2)

20