Noise Jamming Detection Algorithm

Anatolii Kononov, Kim Jin Ha and Shin Young Cheol

STX Engine Co., Ltd., Yongin-si, Gyeonggi-Do, 446-915 Korea
Email: kaa@ieee.org

Abstract—An algorithm for detection of continuous-wave noise caused, e.g. by the presence of asynchronous interference, the
jamming in two-dimensional surveillance radar is proposed. The algorithm employs non-coherent integration with censoring or
algorithm is highly immune to strong interfering spikes due to censored video integration (CVI) [1, 2]. The results of CVI for
using censored video integration of range noise samples collected each antenna dwell represent azimuth samples associated with
during dwell time for each azimuth resolution cell. To detect noise the corresponding azimuth bins. To detect the presence of
jamming the algorithm applies order statistics CFAR to a se- jamming the algorithm applies order statistics (OS) CFAR
quence of azimuthal samples resulted from censored video inte- technique to a sequence of azimuthal samples. The reason for
gration in range. The detection performance of the algorithm in
using OS CFAR is that it prevents the misdetection of jam-
the absence of interference is evaluated numerically by means of
ming if some of azimuth reference samples are affected by
an integral equation and verified using statistical simulations. In
the presence of infinitely strong spikes the detection and false
strong spikes, which pass through the stage of CVI when cen-
alarm performances are estimated using statistical simulations. soring is not able to cope with excessive number of spikes.
The paper is organized as follows. In Section II, the detec-
I. INTRODUCTION tion algorithm is described in detail. An integral equation for
the detection performance is considered in Section III. In Sec-
One of the ECM techniques commonly used against sur- tion IV some results of performance analysis are reported and
veillance radars is continuous-wave noise jamming, which is discussed. Concluding remarks are presented in Section V.
intended to prevent detection of targets by raising the receiver
noise level. A so-called spot noise jammer has a relatively nar- II. ALGORITHM
row spectral band and may be effective if its band matches the Consider search radar, which looks for surface targets in az-
radar signal bandwidth. If the radar signal bandwidth is not imuth and range using fan antenna beam (Fig.1). The radar
known to the jammer or if the radar changes operating fre- antenna rotates to scan full azimuthal area around the radar
quency, the spot jammer loses its effectiveness. The jammer (360° coverage in azimuth) and receives returns from some
must then employ a wide frequency band that covers the ex- range interval for each azimuth bin, which angular extent is
pected radar frequency extent. Such a jammer, called a barrage equal to the antenna beamwidth in azimuth.
noise jammer, needs more power than a spot jammer because Radar returns received during the dwell time correspond-
only some portion of its power lies in the instantaneous band- ing to each azimuth bin are converted into digital samples,
width of the radar receiver. Thus, an effective ECCM tech-
processed in a matched filter followed by a square-law detec-
nique against noise jamming is varying radar frequency since
it forces the jammer to spread its energy over the wider band- tor and then collected into a data matrix. Hence, the radar
width that results in reducing the power density of jamming, if returns received over a one full scan in azimuth can be repre-
the mean jammer power remains constant, at the input of radar sented in the form of an N-by-M data matrix, where N is the
receiver. total number of range resolution cells (range bins) and M is
Frequency changing by the radar can be performed in a the total number of azimuth bins. Each sample in this matrix
manner of frequency agility when transmit frequency changes is associated with a particular range-azimuth bin. We assume
on a pulse-to-pulse basis for non-coherent signal processing that the jamming detection algorithm uses an Nr-by-M matrix
modes and on a dwell-to-dwell basis when radar echoes re-
ceived over dwell intervals are coherently processed. Another
technique for coherent radars is frequency hopping when radar
360° coverage
changes transmit frequency every time after detecting the pres-
ence of jamming.
To implement frequency hopping the radar must be able to Antenna
Azimuth
Beamwidth
reliably detect the presence of jamming. In this paper, we in-
troduce an algorithm for detection of continuous-wave noise
jamming in two-dimensional (range-azimuth) search radars.
The algorithm non-coherently integrates a set of range noise
Figure 1. Scan pattern for surveillance radar
samples collected during a dwell time interval (coherent proc-
essing interval) corresponding to each azimuth resolution cell
(azimuth bin). To ensure immunity to strong interfering spikes

2013 IEEE Radar Conference (RadarCon13)

978-1-4673-5794-4/13/$31.00 ©2013 IEEE
 1 2 … j j+1 … Naz-1 Naz CVI is reasonable because this CFAR technique is known to
1
have immunity to the presence of interfering targets in the
reference window. Hence, it will also be able to prevent the
2
misdetection of jamming even if some number of the refer-
Input data for
… … … … … … … … … …
jamming detection ence samples yj are affected by strong spikes that may pass
Nr-1
algorithm through the CVI when censoring is not able to cope with ex-
cessive number of spikes.
Nr CVI works as follows. Let z1, z2, …, zNr be the range sam-
ples for an arbitrary column of the input data matrix. First,
CVI sorts these samples according to magnitude in order to
y1 y2 … yj Y yj+1 … yNaz-1 yNaz Results of integration
obtain the ordered sequence of samples z(1) ≤ z(2) ≤ …≤ z(Nr).
CUT After sorting, it integrates the lowest Kc samples. However,
this is not a simple addition. Whereas each of the lowest Kc–1
Figure 2. Structure of input data for jamming detection algorithm and inte- ordered samples is given unit weight, the Kc-th ordered sam-
gration of range samples; CUT is the cell under test
ple is given a weight Nr–Kc +1. With this weighting, the inte-
(Nr<N) of noise samples. For instance, such a matrix of noise grator output can be represented as
only associated samples can be extracted from the N-by-M Kc
data matrix corresponding to a “passive” pulse repetition in- y = ( N r − K c ) z( K c ) + ∑ z ( n ) (1)
terval, intentionally scheduled for each antenna dwell, when n =1
the radar does not transmit pulses.
The structure of the Nr-by-M input data matrix is de- where Kc is the highest order used in the censoring of range
scribed in Fig. 2. Each j-th column of this matrix represents samples, Kc ∈{1, 2, …, Nr}.
noise samples corresponding to Nr range bins associated with Clearly, CVI provides immunity against Nr–Kc strong
the j-th azimuth bin. It is assumed that Naz/2columns to left spikes. CVI also possesses an important property that is used
and Naz/2 columns to the right of the cell under test (CUT) in this paper later to derive the detection performance. As has
represent the Naz azimuth reference bins. Integration of Nr been proven in [3], when the z1, z2, …, zNr are independent
range samples means that each column is added up. The sums exponentially distributed RVs the statistical distribution of y
appear in the bottom row. in formula (1) is the same as that for the sum of only Kc inde-
The random variable (RV) representing the sum of the j-th pendent exponentially distributed RVs without sorting.
reference column is designated yj, while the RV describing The proposed jamming detection algorithm is described be-
the sum of the column under test (shaded column in Fig. 2) is low. The description assumes that when the algorithm decides
designated Y. For the jamming detection algorithm the col- on the presence of jamming for a current antenna dwell (azi-
umn under test corresponds to the current azimuth bin (cur- muth bin) all the reference samples yj corresponding to the
rent antenna dwell) and all the samples in this column col- previous dwells (azimuth bins) are already computed using
lected from the previous scan are replaced with the samples CVI. It is also assumed that for the OS CFAR, which is used at
collected over the current dwell. The reference columns cor- the final decision stage of the algorithm, the rank Kos of the
respond to the azimuth bins adjacent to the CUT. For each representative reference sample and a scalar factor α needed to
current azimuth bin the RV Y is an updated sum of the range compute the adaptive detection threshold are known. How to
samples from the current dwell, while the reference samples determine α for OS CFAR in case of the full noncoherent inte-
yj, j=1,2,…, Naz correspond to the previous dwells. gration and CVI is shown in Section III.
Let’s assume that a standard cell averaging (CA) CFAR ALGORITHM
algorithm is used for noise jamming detection. To detect the
presence of jamming in the current azimuth bin the CA 1. Sort the range samples z1, z2, …, zNr from the current dwell
CFAR algorithm compares Y with an adaptive detection 2. Having obtained the ordered sequence z(1) ≤ z(2) ≤ …≤ z(Nr)
threshold computed using the sum of all Naz reference sam- update the test statistic Y for the CUT using equation (1)
ples yj. It is clear that if Y and yj are computed by integrating 3. Sort the azimuthal reference samples yj, j=1, 2, …, Naz and
all of the Nr range samples, the CA CFAR algorithm is sensi-
then select the Kos-th ordered sample X = y(Kos)
tive to interfering spikes. Indeed, the presence of a single in-
finitely strong spike in the column under test will guarantee a 4. Decide on the presence of jamming using the rule
false alarm. On the other hand, if a single infinitely strong If Y ≥ α X jamming is present
spike is present in one of the reference columns then the CA If Y < α X jamming is absent
CFAR will miss the presence of jamming.
To cope with interfering spikes the proposed algorithm III. PERFORMANCES
combines the CVI and order statistics (OS) CFAR technique.
As has been shown in [1, 2] the CVI is immune to spikes, yet In this paper, we assume that the radar receiver noise and
suffers a small loss with respect to the full noncoherent inte- jamming are statistically independent zero-mean white noises
gration that sums all Nr range samples. Using OS CFAR after that obey Gaussian distribution and the input of a square-law
detector is normalized with respect to the receiver noise umns out of the Naz columns (Fig. 2) are affected by infinitely
RMS. Under these assumptions, the RVs zn, n =1, 2, …, Nr at strong spikes with J ≤ Nr – Kc spikes per each column. These
the output of the square-law detector are independent and spikes occupy the top J ranks and their effect is to reduce the
identically distributed variables that obey exponential distri- effective number of samples [2], i.e. the number of independ-
bution with probability density function (PDF) p(z) = exp(-z) ent and identically distributed (i.i.d.) samples. Hence, the ef-
when the input is only the receiver noise, and with the PDF fective number of samples in each affected reference column
p(z) = Dexp(-Dz) when the input is the receiver noise plus is Nr – J. However, the integration rule of equation (1) as-
jamming. The parameter D is given by D=1/(1+JNR), where sumes that the effective number of samples is strictly equal to
JNR is the jamming-to-noise power ratio at the input of the Nr. Thus, the PDF of the weighted sum y is not identical any
square-law detector. more to that of Kc unsorted samples. As a result, the reference
window becomes nonhomogeneous, since the PDF of the
A. Detection Performance Nspaz reference samples is not equal to the PDF of remaining
The derivation of the detection performance in the ab- Naz – Nspaz reference samples corresponding to the non-spiky
sence of interfering spikes is straightforward. First, we refer reference columns. Expression for the PDF of y given by (1)
to the detection performance of the OS CFAR with full non- in case of any number of spikes J is too elaborate [2] and the
coherent integration, which has been studied in [4]. As has derivation of closed form expression for the PDF of the Kosth
been shown, the detection probability PD in this case is given ordered sample in this case is also not a simple task.
by the following integral equation Discursively, the detection probability decreases when the
reference columns are affected by strong spikes (assuming no
⎛N ⎞ ∞⎡ N r −1
(α Dy ) j ⎤ spikes in the CUT). Because the J infinitely strong spikes
PD = K os ⎜ az ⎟ ∫ ⎢exp(−α Dy ) ∑ ⎥
⎝ K os ⎠ 0 ⎣ j =0 j! ⎦ occupy the top J ranks the Kc-th ranked sample out of Nr – J
K os −1 N az − K os i.i.d. samples is statistically higher than the Kcth ranking sam-
⎡ N r −1 y j exp(− y ) ⎤ ⎡ N r −1 y j exp(− y ) ⎤ ple out of i.i.d. Nr samples. Since the Kcth ranked sample has
× ⎢1 − ∑ ⎥ ⎢∑ ⎥
⎣ j =0 j! ⎦ ⎣ j =0 j! ⎦ the higher weight in the sum of equation (1) the reference
N r −1 samples yj corresponding to the spiky reference columns be-
y exp(− y ) come statistically higher than those for the reference columns
× dy
( N r − 1)! without spikes. Due to this the Nspaz reference samples in the
(2) CFAR reference window are statistically higher than those for
To compute PD, equation (2) has to be integrated numerically. the non-affected reference columns. Hence, the Kosth ranked
The false alarm probability PFA is obtained from (2) after set- sample in the CFAR reference window becomes statistically
ting D = 1 (JNR = 0).To determine the scalar factor α corre- higher than it would be if there were no spikes and the detec-
sponding to the desired PFA one should use (2) iteratively with tion probability decreases.
D = 1. It should also be noted that the OS CFAR with full nonco-
To derive the detection performance for the OS CFAR with herent integration totally collapses (the detection probability
CVI we recall that the statistical distribution (PDF) of y in for- is zero) when J ≤ Nr – Kc and Nspaz > Naz – Kos, while the OS
mula (1) is the same as that for the sum of only Kc independent CFAR with CVI is able to operate without a total collapse
exponentially distributed RVs without sorting. Hence, all the even if all the reference samples are spiky, i.e. Nspaz = Naz.
results obtained for OS CFAR with noncoherent integration of Because the analytical analysis of the effect of strong in-
Nr range samples apply to the OS CFAR with CVI after replac- terfering spikes on the detection performance of the proposed
ing Nr with Kc. Thus, equation (2) completely describes the method is not straightforward we have resorted to statistical
detection performance of the OS CFAR with CVI for any inte- simulations.
ger Kc, 1 ≤ Kc ≤ Nr. B. False Alarm Performance
As follows from the discussion above, if Kc = Nr, the de-
tection performance of the OS CFAR with CVI is equivalent In the analysis of the false alarm performance it is assumed
to that of the OS CFAR with full noncoherent integration. If that the probability of interference is small and therefore the
Kc < Nr, the detection performance of the former degrades detection threshold is set up at a nominal level so that the
with respect to that of the latter, since CVI suffers inherent desired PFA is reached in the absence of interfering spikes.
loss with respect to the noncoherent integration using all Intuitively the false alarm probability increases if strong
available Nr samples. However, this loss is relatively small spikes are present only in the CUT. Indeed, the J ≤ Nr – Kc
(on the order of 1 dB when Kc is close to Nr [1, 2]). Hence, infinitely strong spikes occupy the top J ranks. To get a false
one should expect relatively small loss for the OS CFAR with alarm, the CUT sample Y, which is the Kcth ranked sample
CVI when Kc < Nr and remains close enough to Nr. out of remaining Nr – J samples in the CUT, has to cross the
The analysis of the effect of strong interfering spikes on nominal threshold computed using the Kosth ordered sample
the detection performance of the OS CFAR with CVI is not out of the Naz reference samples yj, which are not affected by
straightforward. Let us consider the most interesting scenario spikes. The Kcth ranked sample out of the Nr – J i.i.d. samples
(in terms of detection performance) when the CUT column is statistically higher than the Kcth ranking sample out of the
does not contain spikes and the Nspaz azimuthal reference col- Nr i.i.d. samples. Because the Kcth ranked sample has the
higher weight in the sum of equation (1) the CUT sample Y is 1
statistically higher than each of the reference samples yj. N = 32
r
0.9
Hence, the false alarm probability increases. It is clear that if N = 32
az
J > Nr – Kc then the probability of a false alarm is unity. 0.8 K = 24

D
os

Detection Probability P
On the contrary, if the spikes are not present in the CUT
0.7 P = 10-7
and the Nspaz ≤ Naz reference columns are affected by infi- FA
nitely strong spikes, the false alarm probability decreases be- 0.6
cause all the spiky reference samples yj are statistically higher K = 32
0.5 c
the CUT sample Y. It is also clear that when J > Nr – Kc and 30
the number of spiky reference columns Nspaz > Naz – Kos the 0.4 28
false alarm probability is zero. 0.3 26
If the strong spikes appear in the CUT and in the reference 24
0.2
columns, the analysis of the false alarm performance is too
elaborate because the false alarm probability is an intricate 0.1
function of the J, Nr, Kc, Naz, Nspaz and the PDF of the Kosth 0
ordered sample for a nonhomogeneous reference window. -2 -1 0 1 2 3 4 5 6
Jamming-to-Noise Ratio, dB
To analyze the false alarm performance we have also re-
sorted to statistical simulations. Figure 4. Effect of Kc on performance of OS CFAR with CVI

IV. NUMERICAL RESULTS

seen, the corresponding estimated and theoretical curves are
A. Detection Performance very close. The computations show that for both of the cases,
To verify the theoretical result presented by equation (2) 1– full noncoherent integration with Nr=32 (this is equivalent
we have resorted to statistical simulation. Fig. 3 shows the to CVI with Kc=32) and 2– CVI with Kc=24, the absolute
graphs of PD (in the absence of interfering spikes) versus JNR relative error does not exceed 3% for 0 ≤ JNR ≤ 6 dB.
at a nominal PFA= 10–7 for the case of Nr = 32 samples, Kc = Fig. 4 demonstrates the effect of Kc on the theoretical de-
24, Naz= 32, and Kos= 24. With this selection, the CVI is im- tection performance of the OS-CFAR with CVI (in the absence
mune to as many as Nr–Kc = 8 spikes. It is also found using of interfering spikes), with other parameters equal to those
equation (2) iteratively with PFA= 10–7 and D = 1 that α = from the previous analysis. Inspecting the graphs of Fig. 4 at
2.060293 for the OS CFAR with noncoherent integration of point PD = 0.8 shows that decreasing Kc from 32 to 24 yields
Nr = 32 samples and α = 2.267472 for the OS CFAR with an additional loss that increases evenly from 0.2 dB (Kc = 30)
CVI when Kc=24. In Fig. 3, the dotted curves represent the to 0.9 dB (Kc = 24). This relatively small loss is the price paid
theoretical performances computed using numerical integra- for the CVI immunity to spikes: when Kc decreases from 30 to
tion in (2), while the solid curves are the performances esti- 24 the immunity to spikes improves from 2 to 8 spikes. It
mated by Monte-Carlo trials using 10000 runs. As can be should also be noted that even for Kc = 24, when the loss in

1 1 Nr = 32 Naz = 32 CUT: J=0 -5

PFA = 10
N = 32 Kc = 24 Kos = 24 Ref window: J = 4
r
0.9 K = 24 FNCI Equation (2) Nspaz = 0
c
0.8 CVI Equation (2) Nspaz = 0
D

0.8 N = 32
Detection Probability PD

az 1 2
Detection Probability P

FNCI Simulation Nspaz = 6

0.7 K = 24 CVI Simulation Nspaz = 6
os
-7 FNCI Simulation Nspaz = 8
0.6 P = 10 0.6
FA CVI Simulation Nspaz = 8

0.5 FNCI Simulation Nspaz = 9

CVI Simulation Nspaz = 9
0.4 0.4 CVI Simulation Nspaz = 32

0.3

0.2 0.2

0.1

0 0
-2 -1 0 1 2 3 4 5 6 -4 -3 -2 -1 0 1 2 3 4
Jamming-to-Noise Ratio, dB Jamming-to-Noise Ratio, dB
Figure 3. Theoretical and estimated detection performances Figure 5. Estimated detection performances in the presence of infinitely
1– OS CFAR with full non coherent integration of Nr = 32 samples (this is strong spikes for different number of spiky reference columns Nspaz
equivalent to OS CFAR with CVI when Kc = 32); 2– OS CFAR with CVI for No spikes in CUT (J = 0); number of spikes per each spiky reference column
Kc = 24 J = 4; Nspaz = 6, 8, 9 and 32; FNCI and CVI stand for OS CFAR with full
noncoherent integration and CVI, respectively.
TABLE I. ESTIMATED FALSE ALARM PROBABILITY IN THE PRESENCE TABLE II. ESTIMATED FALSE ALARM PROBABILITY IN THE PRESENCE
OF STRONG INTERFERING SPIKES IN CUT OF STRONG INTERFERING SPIKES IN CUT AND REFERENCE WINDOW
Nr = 32, Kc = 24, Naz = 32, Kos = 24, nominal PFA = 10-5 Nr = 32, Kc = 24, Naz = 32, Kos = 24, nominal PFA = 10-5
Nspaz = 0 J=4

J 2 4 6 8 9 Nspaz 2 4 6 8 10

CVI 2.21×10-4 3.66×10-3 5.53×10-3 0.509 1 CVI 2.78×10-3 1.99×10-3 1.37×10-3 9.78×10-4 6.1×10-4
PFA PFA
FNCI 1 1 1 1 1 FNCI 1 1 1 1 1

JNR is about 1 dB, the OS CFAR with CVI is quite effective: reference columns (Table II). For both of these cases, the false
to ensure the jamming detection probability PD = 0.8 at PFA= alarm probability for the OS CFAR with full noncoherent inte-
10–7 the required input JNR is about 3.2 dB. gration is always equal to unity.
The effect of strong spikes on the detection performances Table I illustrates the increase of PFA for the OS CFAR
of the OS CFAR with CVI as well as with full noncoherent with CVI when the number of strong spikes J in the CUT in-
integration for the scenario when spikes are present only in the creases from 2 to 9. Table II shows the decrease of PFA when J
reference window, number of spikes in each affected reference is fixed (J = 4 for both the CUT and reference samples) and
column J = 4 (J < Nr – Kc) and Nspaz = 6, 8, 9 and 32 is shown the number of spiky reference samples Nspaz increases from 2
in Fig. 5 for PFA = 10-5. The graphs in this figure are computed to 10. Analyzing the data in Table I and II yields that the false
using Monte-Carlo simulations (104 trials). Analysing the de- alarm performance of the proposed algorithm is highly im-
tection curves in Fig. 5 yields that OS CFAR with CVI is supe- mune to the spiky interference.
rior to OS CFAR with full noncoherent integration in terms of
immunity to strong interfering spikes. As one can see, when
Nspaz increases the detection performance of the former (solid V. CONCLUSION
lines) does not degrade significantly with respect to that in the The results presented in this paper have demonstrated that
absence of spikes (solid line computed from equation (2)): for the proposed algorithm based on the combination of the order
PD = 0.8 the loss in the JNR is 0.6 dB when Nspaz = 9 and 1.8 statistics CFAR with censored video integration is a reliable
dB even when all the reference samples are spiky, i.e. Nspaz = method for detection of noise jamming.
32. At the same time the performance of the latter (dashed In the absence of interfering spikes the algorithm ensures
lines) significantly degrades with respect to that in the absence sufficiently high detection probability at a quite low false
of spikes (dashed line computed from equation (2)) and with alarm probability even when the input jamming-to-noise ratio
respect to the performance of the former. As can be seen from is on the order of 3 dB. At the same time, the algorithm has
Fig. 5 the detection performance of the OS CFAR with full also been shown, from the point of view of detection and
noncoherent integration for Nspaz = 8 is even inferior to the false alarm performances, to be highly immune to strong in-
detection performance of the OS CFAR with CVI for Nspaz = 9 terfering spikes.
and totally collapses when Nspaz = 9.
Fig. 5 clearly demonstrates the significant advantage, in
terms of detection performance in the presence of strong inter- REFERENCES
fering spikes, of the OS CFAR with CVI over the OS CFAR [1] N. Levanon, “Censored video integration in radar detection,” in Proc.
with full noncoherent integration: even when Nspaz > Naz – Kos IEEE International Radar Conference, Arlington, Virginia, May 7-10, pp.
the former remains highly robust, while the latter totally col- 511-513, 1990.
lapses. [2] N. Levanon, “Analytic comparison of four robust algorithms for post-
detection integration,” IEE Proceedings-F, vol. 139, no. 1, pp. 67-72,
B. False Alarm Performance 1992.
[3] J. A. Ritcey, “Performance analysis of the censored mean-level detectors,”
To show the effect of strong spikes on the false alarm per- IEEE Transactions on Aerospace and Electronic Systems, vol. AES-22,
formance of the proposed algorithm we present the results of no. 4, pp.443-454, 1986.
Monte-Carlo simulations (106 trials are used) for the two cases [4] M. Shor and N. Levanon, “Performances of order statistics CFAR,” IEEE
when the infinitely strong spikes occur only in the CUT col- Transactions on Aerospace and Electronic Systems, vol. AES-27, no. 2,
umn (Table I) and when they are present in both the CUT and pp. 215-224, 1991.