See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/322201609

Automatic LPI Radar Waveform Recognition using CNN

Article  in  IEEE Access · January 2018

DOI: 10.1109/ACCESS.2017.2788942

CITATIONS READS

26 606

3 authors, including:

Seung-Hyun Kong Linh Hoang

Korea Advanced Institute of Science and Technology University of Technology Sydney
50 PUBLICATIONS   410 CITATIONS    3 PUBLICATIONS   29 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Seung-Hyun Kong on 09 February 2018.

The user has requested enhancement of the downloaded file.

 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Automatic LPI Radar Waveform

Recognition using CNN
SEUNG-HYUN KONG1 , (Senior Member, IEEE), MINJUN KIM1 , AND LINH MANH HOANG1
1
The CCS Graduate School of Green Transportation, Korea Advanced Institute of Science and Technology, Daejeon, Korea, 305-701 (e-mail: {skong, minjunk,
linhhm}@kaist.ac.kr)
Corresponding author: Seung-Hyun Kong (e-mail: [email protected]), Seung-Hyun Kong and Minjun Kim are co-first authors.
The authors gratefully acknowledge the support from Electronic Warfare Research Center at Gwangju Institute of Science and Technology
(GIST), originally funded by Defense Acquisition Program Administration (DAPA) and Agency for Defense Development (ADD).

ABSTRACT Detecting and classifying the modulation scheme of the intercepted noisy LPI (low probabil-
ity of intercept) radar signals in real-time is a necessary survival technique required in the electronic warfare
systems. Therefore, LPI radar waveform recognition technique (LWRT) has gained an increasing attention
recently. In this paper, we propose a convolutional neural network (CNN) based LWRT, where the input and
hyper-parameters of the CNN, such as the input size, number of filters, filter size, and number of neurons
are designed based on various signal conditions to guarantee the maximum classification performance. In
addition, we propose a sample averaging technique (SAT) to efficiently reduce the large computational cost
required when the intercept receiver needs to process a large amount of signal samples to improve the
detection sensitivity. We demonstrate the performance of the proposed LWRT with numerous Monte Carlo
simulations based on the simulation conditions used in the recent LWRTs introduced in the literature. In
addition, it is testified that the proposed LWRT offers significant improvement, such as robustness to noise
and recognition accuracy, over the recent LWRTs.

INDEX TERMS convolutional neural network, low probability of intercept, radar waveform recognition.

I. INTRODUCTION time-frequency analysis (TFA) techniques using an individ-

AST and efficient low probability of intercept (LPI) ual transform, such as short-time Fourier transform (STFT)
F radar waveform recognition is one of the core functions
required in the electronic warfare (EW) systems, such as elec-
[3], Wigner Ville Distribution (WVD) [2], [4], and Choi-
William Distribution (CWD) [2], [5], [6], and there are TFA
tronic support and electronic intelligence (ES/ELINT), elec- techniques utilizing a combination of two transforms, such
tronic protection (EP), and electronic attack (EA) systems. In as WVD with CWD [7], [8] and WVD with ambiguity
practice, automatic radar waveform recognition technique is function (AF) [9]. On the other hand, for the classification
a core survival technique for an intercept receiver performing techniques, there are hierarchical decision tree (HDT) [3],
threat recognition and radar emitter identification [1]. On [9] and artificial neural networks (ANN) based on multilayer
the contrary, radar signals should be designed to have LPI perceptron (MLP) [2], [5], [7], recurrent neural networks
properties [2] so that a radar signal is not easily detected (RNN) [8], and convolutional neural networks (CNN) [4],
and recognized by intercept receivers. Therefore, intercept [6].
receivers should be necessarily equipped with an automatic In [2], [5], the time-frequency image (TFI) produced by a
LPI radar waveform recognition function that has highly re- TFA technique is cropped and used as an input to an ANN
liable detection, classification, and identification capabilities based on MLP. The LWRTs in [7], [8] extract various signal
to recognize the presence of LPI radar signals in advance. features from the complex envelope (CE) of the intercepted
In the literature, there have been a few LPI waveform signal, such as the second order statistics and instantaneous
(i.e., LPI radar waveform) recognition techniques (LWRT) frequency and phase, and also extract TFI features of the
that utilize feature extraction techniques and classification intercepted signal. The signal features and the TFI features
techniques to extract features from the intercepted signal are then used as an input to an ANN based on MLP [7]
and to classify the intercepted signal based on the extracted or based on RNN [8]. The LWRTs in [3], [9] introduce
features, respectively. For the feature extraction, there are classification techniques based on the HDT, where the signal

VOLUME 4, 2016 1

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

features obtained using TFA techniques are compared to various simulations.

the thresholds to identify the correct branch in the decision In fact, most of the studies [2], [6], [7], [8] in the field
tree. These classification techniques based on the HDT can exploit less than two thousand signal samples and show a se-
achieve a fast classification performance, however, the binary vere performance degradation as the SNR of the intercepted
(i.e., yes or no) decisions made by comparing a feature value signal decreases below −4dB. To overcome this limit, the
against a threshold may have performance degradation when authors in [9] utilize a much larger number of signal samples,
a decision needs to consider multiple features at the same (e.g., 104 samples) to increase the processing gain (PG) and
time. Note that the LWRTs [3], [7], [8], [9] exclude all of utilize fractional Fourier transform (FRT) to reduce the com-
the four polytime codes introduced in [2] as representative putational cost required for the employed TFA techniques
LPI waveform modulations, and signal features to distinguish such as Radon ambiguity transform (RAT) and Radon-WVD
polytime codes are not studied. In addition, the performance (RWT). As a result, the LWRT in [9] can classify LPI radar
of LWRTs [7], [8] degrades significantly in the presence of waveforms at the SNR of −10dB.
carrier frequency estimation error (CFEE), since the calcu- In this paper, we propose an LWRT based on CNN, where
lated signal features from the CE can be corrupted by the the considered LPI waveforms are all of the twelve pulse
remaining frequency component. In fact, the LWRT in [8] compression waveforms defined in [2] which includes all
assumes that there is no remaining frequency component of the five polyphase and four polytime codes. In addi-
(i.e., zero center frequency) in the CE, however, in [7], the tion, all of the waveforms are collected only for a single
authors show that there is about 20% (or larger) performance period, and the input to the CNN is a grayscale TFI (i.e.,
degradation for LPI signals with SNR of −3dB (or less) in with amplitude information). The hyper-parameters of the
the presence of CFEE. proposed CNN are designed with various simulations and
The recent LWRT in [6] utilizes CWD to generate TFI and verified to guarantee the maximum performance. In addition,
uses the binary version of the generated TFI for an input to we exploit the sample averaging technique (SAT), used for
the CNN based classification technique. This technique has fast and high sensitivity GPS acquisition, in order to process
robustness against the CFEE, since the effect of CFEE on the a large number of signal samples with reduced computational
TFI is merely an offset in the object image location and the cost effectively. The performance of the proposed LWRT is
CNN is generally robust to the changes in the image topology compared to the recent LWRTs in [6], [8], [9] for the signal
[10]. The performance of [6] shows a strong improvement parameters and simulation conditions used in the studies [6],
compared to the conventional techniques in distinguishing a [8], [9]. The proposed technique without the SAT shows
set of LPI radar waveforms including the four polytime codes significant performance improvement over the recent LWRTs
[2] at SNR of −4dB or higher. However, the technique [6] in [2], [6], [7], [8], and the proposed technique with the SAT
does not consider four polyphase codes (e.g., P1, P2, P3, has much better performance than [9] in terms of both the
and P4) other than the Frank code. In addition, the LWRT sensitivity and classification accuracy for the same number
in [6] assumes that the Barker code based BPSK (binary of signal samples. The superior performance of the proposed
phase shift keying) waveform is repeating multiple times technique is demonstrated with numerous Monte Carlo sim-
during the pulse interval so that the collected signal samples ulations for a wide range of SNR of all waveforms.
of the BPSK waveform are from consecutive multiple code This paper is organized as follows. Section II presents
periods, while other waveforms (i.e., except the BPSK) are the proposed LWRT and describes the LPI radar waveforms
collected for a single period. This assumption makes the considered in this paper. Section III introduces the proposed
Barker modulated signal object in the TFI be very different pre-processing technique (i.e., the SAT), the TFA technique
from that of the T1 code (a polytime code), and otherwise, (i.e., CWD) used to generate TFI input to the CNN, and
the TFI of the two codes (i.e., BPSK and T1) may have a the waveform patterns of the twelve LPI waveforms on the
similar shape especially in the presence of large noise. In resulting TFIs. The TFI resizing process is also described in
addition, the input format (i.e., binary images) and hyper- Section III. The CNN input size and the hyper-parameters are
parameters (i.e., the parameters that define the CNN in detail analyzed and determined in Section IV, and the performance
such as input size, filter size, and number of filters) of the of the proposed technique is compared to the recent LWRTs
CNN in [6] are identically the same to those of [11] that in Section V. Finally, the conclusion of this paper is drawn in
is originally developed for hand-writing digit recognition, Section VI.
and the modification is only made to the fully connected
(FC) layer by adjusting the number of neurons (i.e., nodes) II. SYSTEM AND WAVEFORMS
according to the empirical formula [12]. In practice, larger In this section, we present the system structure of the pro-
TFI input size and amplitude (i.e., intensity) information posed LWRT and the definition of the twelve LPI radar
of the objects in the TFI can be better to the CNN, when waveforms considered in this paper.
the objects have subtle differences that are better observed
with finer image resolution and amplitude information. In A. PROPOSED LWRT
practice, the hyper-parameters of CNN should be designed Fig.1 illustrates the block diagram of the proposed LWRT.
and verified to maximize the recognition performance with The intercepted signal is down-converted to an intermediate
2 VOLUME 4, 2016

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

TABLE 1: Twelve LPI radar waveforms

RF
Signal Modulation
f [k][Hz] φi,j [k][rad] for subcode
Training type
data Training
Antenna
f0 +
LFM constant

Preprocessing-II
B
Preprocessing-I

IF Original Test τpw (kTs )

Front end Signal Time-
TFI data Modulation
Intercept Frequency Classifier Costas fj constant
Type
receiver y[k] ya [n] Analysis
BPSK constant 0 or π

2π
FIGURE 1: Proposed LPI radar waveform recognition system. Frank constant M (i − 1)(j − 1)

π
P1 constant −M [(M − (2j − 1))][(j − 1)M + (i − 1)]

π
frequency (IF) fI and then sampled at fs = (1/Ts ) to yield P2 constant − 2M [2i − 1 − M ][2j − 1 − M ]

y[k], where k is the sample index. A set of consecutive N π

− 1)2
P3 constant ρ (i
signal samples collected over τpw seconds (i.e. N = τpw · fs )
π 2
is fed into the preprocessing-I block for the proposed SAT. P4 constant ρ (i − 1) − π(i − 1)
 j k 
Similar to the previous studies [6], [7], [8], [9], this paper T1 constant mod N2π
ps
(N si (kT s ) − jτpw )
jNps
τpw
, 2π
focuses on classifying LPI radar waveforms and assumes that 
2π
j  
2j−Ns +1 Nps
k 

the signal samples are collected for a signal pulse interval T2 constant mod Nps (Nsi (kTs ) − jτpw ) τpw 2 , 2π
 j N B(kT )2 k 
(τpw ) and the coarse estimate of the carrier frequency is T3 constant mod N2π ps
2τ
s
, 2π
ps pw
obtained. Note that the preprocessing-I block is an optional  j N B(kT )2 N B(kT )
k 
ps s
T4 constant mod 2π
− ps 2 s , 2π
function used only when fs /(fI +Bs /2) is by multiple times Nps 2τpw

larger than Nsc that is the required number of samples per

mod{a, b} : remainder after division between a and b
cycle of a carrier wave of the highest frequency allowed
by the receiver bandwidth Bs . Note that Nsc ≥ 4 does bac: biggest integer smaller than or equal to a

not cause an SNR loss larger than 1.0dB. The details of

the proposed preprocessing-I are discussed in Section III.
As shown, depending on the use of the SAT in the pre- enough, each LPI radar waveform shows a distinguishable
processing-I, the input to the TFA block can be either y[k] pattern in the TFI produced by CWD (i.e., CWD-TFI). There-
(when the SAT is not used) or ya [n] (when the SAT is used). fore, it is useful to distinguish the waveform patterns using
In the TFA block, the signals are zero-padded to make input the CWD-TFI, which is discussed in detail in Section III. We
size of 2048 samples to the CWD for FFT operation. The also introduce the mathematical definition of the twelve LPI
TFI is then pre-processed in the preprocessing-II block to radar waveforms in this subsection.
resize the TFI to the required input size for the proposed In general, the discrete-time complex samples y[k] can be
classification technique (i.e., CNN). For the training of the expressed as
proposed LWRT, we generate training data set (the output of
the preprocessing-II block) to train the CNN in off-line. The y[k] = x[k] + w[k] = a[k]ejθ[k] + w[k], (1)
training phase of the system is depicted by the dashed arrows where x[k] represents the down-converted (to fI ) discrete-
in Fig. 1. When the CNN is trained enough, LPI radar signals time complex LPI radar signal samples, w[k] is the complex
are classified in the test phase. additive Gaussian white noise (AGWN) process with two-
sided power spectral density N0 /2, k is the sample index
B. LPI RADAR SIGNALS increasing every Ts for a sampling frequency fs , a[k] is the
In this subsection, we define the twelve LPI radar waveforms non-zero constant signal envelope (i.e., amplitude) within the
[2], [13] considered in this paper. There are LFM, Costas pulse interval τpw such that a[k] = A for 0 ≤ kTs < τpw ,
code, BPSK, five polyphase codes (such as Frank, P1, P2, P3, and θ[k] is the instantaneous phase of the LPI radar signal.
and P4 codes), and four polytime codes (such as T1, T2, T3, Similar to the previous studies [6], [8], [9], we assume that
and T4 codes) introduced in the literature [2]. Note that the the pulse period τpp = 1/fpp is larger than τpw , where
recent LWRTs introduced in the literature seldom considers fpp is the pulse repetition frequency, so that the duty cycle
all of the twelve waveform modulation schemes after [2] in DC = (τpw /τpp ) is assumed to be much smaller than 1.
2004. For example, the studies in [7], [8] consider LFM, The instantaneous phase θ[k] can be defined by instantaneous
Costas code, BPSK, and five polyphase codes, the study frequency f [k] and the phase offset φ[k] as
in [9] considers no modulation case, LFM, BFSK, and five
θ[k] = 2πf [k](kTs ) + φ[k]. (2)
polyphase codes, and the study in [6] considers LFM, Costas
code, BPSK, and Frank code, and four polytime codes, which In practice, there are two most common pulse compression
limits the application of the recent LWRTs. techniques that categorize the twelve LPI radar waveforms
When the SNR of the intercepted LPI signal is high considered in this paper into FM (Frequency Modulation)
VOLUME 4, 2016 3

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

1.5
SNR loss [dB]

Average
1 Standard deviation

0.5

0
2 3 4 5 6
Nsc

FIGURE 2: SNR loss with respect to Nsc .

(a) (b)

and PM (Phase Modulation) [14]. The instantaneous fre-

quency f [k] varies while the phase offset φ[k] is kept con-
stant in the FM, and f [k] is kept constant while φ[k] varies
accordingly in the PM [2], [13], as defined in Table 1, where a
subcode represents a partial pulse interval, within which φ[k]
is fixed to a constant.

III. PROPOSED PREPROCESSING TECHNIQUES (c)

A. PREPROCESSING-I: SAMPLE AVERAGING FIGURE 3: CWD-TFIs of a BPSK signal with and without the SAT. (a) CWD-TFI
for y[k] at SNR = −8dB, (b) CWD-TFI for ya [n] at SNR = −8dB, Na =
TECHNIQUE (SAT) Nsc = 4, (c) CWD-TFI for y[k] at SNR=−2dB.
When the intercept receiver needs higher detection sensitiv-
ity, the receiver may need to increase the processing gain
(PG) by increasing the sampling frequency fs by multiple SNR loss due to the SAT. As shown, Nsc of 4 or larger
times than the sampling frequency required by the maximum is required to have SNR loss less than 1.0dB in practice.
observable frequency fmax of the receiver, where Fig.3(a), 3(b), and 3(c) highlight the signal power spectrum
of a weak BPSK modulated LPI radar signal in the CWD-
fmax = fI + Bs /2, (3) TFIs from y[k] for an SNR of −8dB, from ya [n] for an
where Bs is the bandwidth of the receiver bandpass filter and SNR of −8dB, Na = 4, and Nsc = 4, and from y[k] for
is larger than the bandwidth of the signal B. However, as an SNR of −2dB, respectively, for an example. As shown,
the size of signal samples increases, the computational cost the signal power spectrum in Fig. 3(a) is much improved
for the TFA technique (i.e., CWD) increases exponentially, in Fig. 3(b), and the signal quality shown in Fig. 3(b) is
and the real-time detection of LPI radar signal may become similar to that in Fig. 3(c). In fact, the SNR values measured
difficult. The purpose of using the SAT is to reduce the at the maximum signal power to the mean noise floor in
computational cost required for the TFA block computing Fig. 3(a), 3(b) and Fig. 3(c) are 15.22dB, 20.73dB and
CWD by reducing the number of signal samples with at most 21.37dB, respectively, which shows that the SNR degradation
a slight loss of SNR. The goal is achieve by producing aver- is negligible with the proposed SAT. Therefore, as shown in
aged signal samples ya [n] which is the average (or coherent Fig. 2 and Fig. 3, the use of the SAT does not distort the signal
sum) of Na consecutive signal samples y[k] as component in the CWD-TFI, while, the reduction of both the
computational cost for the CWD and the SNR improvement
Na −1
1 X are significant. In general, the CWD processing requires
ya [n] = y[k + nNa ], (4) O(N 3 log2 N ) complex multiplications, which means we can
Na
k=0
reduce the size of the computational cost for the CWD by
where Na = N1 /Nsc and N1 = fs /fmax is the number of O(Na3 log2 N/(log2 N − log2 Na ) times using the proposed
samples per carrier (wave) cycle for the maximum frequency SAT.
fmax allowed by the receiver bandpass filter. Note that ya [n] Note that the concept of the SAT is already used in the
becomes the input to the TFA block in Fig. 1 when the SAT acquisition of GPS signals with unknown Doppler frequency.
is used, but, otherwise, y[k] becomes the input to the TFA In [15], [16], sample averaging is used when the obtained
block. In practice, the SAT is efficient, when Nsc and Na number of samples per a GPS chip is multiple times larger
are sufficiently large so that the phase difference between than 2. Since the consecutive samples being averaged can
neighboring samples of y[k] is negligible and the neighboring be from two neighboring GPS chips that may have opposite
samples can be added (or averaged) coherently with a slight polarities, the averaging process can be destructive when, for
loss of SNR (e.g., much less than 1.0dB). example, the resulting averaged sample may be zero when
Fig. 2 shows the relationship between Nsc and the mean a half of the original samples is from a +1 chip and the
4 VOLUME 4, 2016

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

other half is from a −1 chip. Therefore, in [15], the sample 1000 1000
60
averaging is performed for each possible sample offset and

frequency index
800

frequency index
800 50
includes another process to find the sample offset that leads 40
600 600
to the highest post-correlation SNR. However, exploiting the 30
SAT to reduce the LPI radar signal sample size is different 400 400
20
from GPS. For example, while a code chip has a very short 200 200
10
time interval (almost a micro-second) and the unknown fre- 0 0
quency is small in GPS, a subcode that may cause destructive 0 500 1000 0 500 1000
samples index samples index
averaging has a relatively very long interval but the unknown (a) LFM (b) Costas
frequency can be larger than GPS in the LPI radar waveform 1000 1000
with pulse compression. This is the reason for limiting Na to

frequency index

frequency index
800 800
maintain at least Nsc averaged samples per a carrier cycle
600 600
at the highest allowed frequency (i.e., fI + Bs /2) by the
400 400
intercept receiver.
200 200

0 0
0 500 1000 0 500 1000
B. TFA TECHNIQUE FOR CWD-TFI samples index samples index
(c) BPSK (d) Frank
1000 1000
This subsection provides the essential description of the
CWD used to produce the TFI of the intercepted signal in the
frequency index

frequency index
800 800

TFA block and discusses the unique patterns of the twelve 600 600
LPI waveforms based on their CWD-TFIs. 400 400

The CWD is based on the time-frequency distribution 200 200

of cohen0 s class [13] with an exponential distribution (ED) 0 0

0 500 1000 0 500 1000
kernel [17] as samples index samples index
∞
" ∞
(µ−`)2 (e) P1 (f) P2
X
−α2ωτ
X 1 − 1000 1000
EDy (`, ω) = 2 e p e 4τ 2 /σ
4πτ 2 /σ
frequency index

frequency index
τ =−∞ µ=−∞ 800 800
#
600 600
× y(µ + τ )y(µ − τ ) , (5)
400 400

√ 200 200
where α = −1 is the imaginary unit, ` and ω are the time
0 0
and angular frequency index variables, respectively, τ and µ 0 500 1000 0 500 1000
are discrete variables, and ξ and ω are continuous variables. samples index samples index
2 2
Note that the exponential kernel function ϕ(ξ, τ ) = e−ξ τ /σ 1000
(g) P3
1000
(h) P4
has a scaling factor σ that is an effective parameter for low
pass filtering in the ambiguity function (AF) of the cohen0 s
frequency index

frequency index

800 800

class. Since the CWD is generated from the AF using the ker- 600 600

nel function and the 2-dimensional (2D) Fourier transform, 400 400
the CWD of the intercepted signal y[k] comprised of multiple 200 200
frequency components has high power spectrum intensity
0 0
near the center frequency and some power spectrum compo- 0 500 1000 0 500 1000
nents away from the center frequency due to the auto-terms samples index samples index
and (non-zero frequency) cross-terms in the AF, respectively (i) T1 (j) T2
1000 1000
[18]. The scaling factor σ in (5) can be used to lessen the
frequency index

frequency index

800 800
cross-terms in the CWD, while the frequency resolution in
the CWD is degraded. In the proposed technique, we use 600 600

σ = 1 as a balance between the cross-terms suppression 400 400

and the frequency resolution of the auto-terms in the resulting 200 200
CWD-TFI.
0 0
0 500 1000 0 500 1000
In addition, we can limit the summations over τ and µ samples index samples index
in (5) by applying rectangular windows (RW) WP (τ ) and (k) T3 (l) T4
WQ (µ), respectively, to the ED kernel, and we exploit ω = FIGURE 4: CWD-TFIs of the 12 LPI radar waveforms considred in this paper for
SNR=10dB.
2k/N to modify (5) into a form of Fast Fourier Transform
VOLUME 4, 2016 5

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

(FFT) as Fig. 4(d)-4(h), respectively. Among the polyphase, Frank,

∞
X P1, and P2 codes are derived from step approximations
RWEDy (`, k) = 2 WP (τ )e−α2kτ /N to the LFM signal and P3 and P4 codes are from linear
τ =−∞ approximations to the LFM signal [2]. Due to the difference
" ∞
1 2 in the approximations to the LFM signal, the frequency
− µ
X
× WQ (µ) p e 4τ 2 /σ variation of the Frank, P1, and P2 codes has staircase patterns
µ=−∞ 4πτ 2 /σ in the CWD-TFIs, whereas P3 and P4 codes have linear
#
frequency variation, as shown in Fig. 4(d)-4(h). However, the
×y(` + µ + τ )y(` + µ − τ ) , (6) big jump in the middle of the Fig. 4(d) and 4(g), separating
the power spectrum into two distinctive components with
where WP (τ ) and WQ (µ) are the rectangular windows of weaker amplitude (intensity) than the intensities at both ends
unit amplitude for −P/2 ≤ τ ≤ P/2 and −Q/2 ≤ µ ≤ Q/2, (i.e., the beginning and the end), is because the Frank and P3
respectively. In the following, we discuss the unique patterns codes have the largest phase increments between subcodes
of the twelve LPI waveforms based on their CWD-TFI shown in the middle of the code. On the contrary, P1, P2, and
in Fig. 4. P4 codes have the largest phase increments at both ends of
the codes, which results in its power spectrum separation
1) Linear frequency modulation (LFM) and weaker intensities at both ends of the code and higher
There are two kinds of LFM [13] schemes; monotonous fre- intensity in the middle of the code. Note that P1 and P2
quency modulation, where f [k] increases or decreases from codes have multiple similarities as shown in Fig. 4(e) and
an initial frequency f0 to f0 ± B linearly within the pulse 4(f), respectively, in that there are staircase patterns and the
interval τpw , and triangular frequency modulation, where power spectrum is concentrated in the middle of the code,
f [k] increases from f0 to f0 + B and decreases back to f0 and the only difference is the opposite slope. Note also that
at the end of τpw . Since the LFM radar signal has its power while both of the LFM and P4 codes have linearly increasing
spectrum spread over the signal bandwidth B as shown in frequency patterns, the LFM has a uniform power spectrum
Fig. 4(a), the signal is good for an LPI radar waveform. distribution over the signal bandwidth, but the P4 code has
its power spectrum concentrated (i.e., higher intensity) in the
2) Frequency Shift Keying (FSK) middle of the code. The power spectrum separation at both
In FSK, f [k] = fj follows the frequency hopping (FH) ends of the P4 code may not distinctive when the signal is
sequence [f1 , f2 , . . . , fNF ], where NF is the number of FH corrupted by large noise.
sequence. In general, FSK modulated radar signal has LPI
property because of the pseudo-random frequency hopping 5) Polytime
pattern over a wide signal bandwidth B. In this paper, we use Polytime modulation is another PM, where the phase offset
the Costas code for a FH sequence [13], and its CWD-TFI is φ[k] is quantized into a user-selected number of states and
shown in Fig. 4(b). subcode intervals are not fixed to a constant, as defined in
3) Binary phase shift keying (BPSK)
BPSK modulation is a PM, where the phase offset φ[k] has 2000
2000 120
120
only 2 states (i.e., 0 and π) as shown in Table 1. In BPSK, the
frequency index

index

100
100
frequency index

1500
sidelobe level in the autocorrelation result is not larger than
frequency index

1500
8080
frequency

1/ρ of the main lobe, where the length of the sequence is 1000
1000 6060
ρ. This characteristic provides an advantage to cope with the 4040
500
500
target masking problem [13], but BPSK modulated signals 2020
can be easily detected due to the simple modulation, so it 1000 1500
500 500 1000 1500 2000
2000 20 20 4040 6060 8080 100 120
100 120
is not used for an LPI radar waveform modulation [2] in samplesindex
sample index samples
sample index
index
practice. However, its CWD-TFI in Fig. 4(c) has a noticeable Time- Cut off the
Resize
similarity to the CWD-TFI of the T1 code as shown in Frequency zero-padding
& Normalize
Classifier
Analysis part
Fig. 4(i). Therefore, we include the BPSK modulation for a
performance comparison as it is included in the recent studies 2000
2000

[6], [8], [9].

index

1500
1500
frequency index
frequency

4) Polyphase 1000
1000

Polyphase modulation is a PM, where the phase offset φ[k] 500

500

has more than 2 states. This characteristics is useful for

an LPI radar signal and provides a low sidelobe level in 200200400400600600 800
samples index
800 1000
1000
sample index
the autocorrelation result. The CWD-TFIs for the polyphase
codes (i.e., Frank, P1, P2, P3, and P4 codes) are shown in FIGURE 5: Preprocessing-II for cropping and resizing.

6 VOLUME 4, 2016

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

Conv-1 Pooling-1
Size: 7×7 Size: 2×2 Output
FC-1 FC-2 Softmax
Stride: 1 Stride: 2 Conv-2 Pooling-2 Layer
Zero-pad: 3 Zero-pad: 0 Size: 3×3 Size: 2×2
Flatten
Stride: 1 Stride: 2
Zero-pad: 1 Zero-pad: 0

Modulation
Type

30@64×64 60@64×64 60@32×32

Feature Feature Feature
1@128×128 30@128×128
Maps Maps Maps 61,440 200 12
Input Feature
Image Maps Neurons Neurons Neurons

Features extraction Classification

FIGURE 6: The CNN structure for the classification technique.

Table 1 [2]. In this paper, we consider two states (i.e., 0 and we crop the CWD-TFI to remove the pixels in the CWD-
π) for the phase offset φ[k], which is the most popular and TFI generated by the zero-padding. The cropped CWD-TFI
applicable case, because of the simplicity in generating the is then resized to further reduce the input size appropriately
waveforms [2], [19]. Among the polytime modulations, T1 to the CNN, which is to lessen the computational cost for
and T3 codes have φ[k] = 0 at the beginning and φ[k] = 0 or CNN while guaranteeing good classification performance.
π in the middle and at the end of the code. Due to this fact, T1 However, we should not reduce the cropped CWD-TFI to a
and T3 codes have constant frequency at the beginning and too small size, because the details of the object in the CWD-
symmetrical power spectrum distribution across the center TFI can be lost, which may result in a significant perfor-
frequency in the middle and at the end of the code. As a mance degradation. Therefore, we need a balance between
result, the CWD-TFIs of T1 and T3 codes have ‘<’ shapes, the computational cost and the classification performance in
as shown in Fig. 4(i) and 4(k). On the other hand, T2 and T4 regard to the input size of the CNN. In the proposed LWRT,
codes have φ[k] = 0 at the center of the code and φ[k] = 0 we use P = Q = 2048 for the rectangular window WP (τ )
or π at both ends of the code, which result in ‘X’ shapes and WQ (µ) so that the produced CWD-TFI has 2048×2048
in the CWD-TFIs as shown in Fig.4(j) and 4(l). Note that pixels, and we resize the TFI into 128×128 pixels using the
the CWD-TFIs for BPSK and T1 shown in Fig. 4(c) and nearest-neighbor interpolation (NNI) technique [20]. Note
Fig. 4(i), respectively, are similar and can be difficult to that the cubic interpolation (CI) technique [21] may be
distinguish when the power spectrum in the tail (i.e., samples used in the image resizing, however, we do not use the
with indices larger than 500) are buried under a strong noise CI technique, since each resulting pixel may be corrupted
(i.e., low SNR cases). Similarly, T2 and T4 codes shown in by multiple noisy pixels in the original image. The image
Fig. 4(j) and 4(l), respectively, can be difficult to distinguish cropping and resizing processes are illustrated in Fig. 5.
when the power spectrum at the both sides (i.e., samples with
indices less than 300 and those larger than 700) are buried IV. DESIGN OF THE PROPOSED CNN
under a strong noise (i.e., low SNR cases). In general, CNN shows a superb performance in image
classification problems such as handwriting recognition and
C. PREPROCESSING-II: IMAGE RESIZEING various object recognition including human faces, license
Studies in [7], [8] apply noise filtering to the input image to plates, hand gesture, logos, and texts [22]. This is the reason
lessen the effect of the noise and, thus, to improve SNR, but we employ the CNN in the proposed LWRT, and this section
it may result in a loss of signal information contained in the introduces the design of the proposed CNN based classifica-
details of the image. Therefore, preserving the details in the tion technique.
CWD-TFI (e.g., the staircase patterns in the Frank and P1 As shown in Fig. 6, CNN consists of a (image) feature ex-
codes, weaker intensities at both ends of the P4, at the tail of traction block and a classification block. Because the feature
BPSK and T1, and the sides of T2 and T4) is critical to the extraction block is integrated inside, CNN does not require
classification performance. Therefore, we use intact CWD- any prior feature extraction function [11]. In addition, the
TFI for any level of the noisy signal in the proposed LWRT. convolution and pooling processes in the CNN make the
In the proposed LWRT, we collect signal samples of less CNN robust to the geometrical distortions, such as scaling,
than 2048 and zero-pad to make an input size of 2048 shift, and rotation, and to the noise in the input image [11].
samples in order to use the FFT algorithm in the CWD. Then, Considering the number of classes to classify, the fact that
VOLUME 4, 2016 7

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

TABLE 2: Determination of the hyper-parameters of the CNN

First Conv Second Conv First Conv Second Conv Number of Average pcc for
Input size
Filter number Filter number Filter size Filter size neurons SNR=−4, −6, −8dB
64×64 10 20 5×5 5×5 400 83.08
64×64 20 40 5×5 5×5 400 84.11
64×64 30 60 5×5 5×5 400 89.56
64×64 40 80 5×5 5×5 400 86.83
128×128 10 20 5×5 5×5 400 89.64
128×128 20 40 5×5 5×5 400 87.86
128×128 30 60 5×5 5×5 400 90.08
128×128 40 80 5×5 5×5 400 89.19
256×256 10 20 5×5 5×5 400 88.22
256×256 20 40 5×5 5×5 400 88.72
256×256 30 60 5×5 5×5 400 87.72
256×256 40 80 5×5 5×5 400 87.47
3×3 3×3 400 88.78
5×5 3×3 400 90.89
5×5 5×5 400 90.08
7×7 3×3 400 92.78
128×128 30 60 7×7 5×5 400 88.86
7×7 7×7 400 91.75
9×9 3×3 400 91.89
9×9 5×5 400 87.72
9×9 7×7 400 91.94
9×9 9×9 400 89.50
100 91.94
200 93.67
300 92.53
128×128 30 60 7×7 3×3 400 92.78
500 89.19
600 90.72
700 90.25

the input image is in a gray scale, and the subtle shapes [25], but also because of the high complexity of the image
of the image objects of the twelve LPI radar waveforms features of the twelve LPI radar waveforms shown in Fig. 4.
in the CWD-TFIs, we can observe that there are multiple However, due to a large number of neurons in the FC layer,
similarities between the hand-writing recognition problem we may have an overfitting problem, for which we employ
[11] and the problem considered in this paper. Therefore, a Dropout layer between the two FC layers denoted as FC-1
we start with the basic structure of the CNN studied in [11], and FC-2 in Fig. 6.
[23], [24], [25] to develop an appropriate CNN structure for In the first step of the design, we determine the input size of
the proposed LWRT. The basic structure of the CNN can be the CNN and the number of convolution filters with various
described in a sequence of functions as Input−Conv−ReLU simulations, where the input size is related to the resolution of
− Pooling − Conv − ReLU − Pooling − FC − Dropout − FC, the objects in the CWD-TFI, and the number of convolution
where Conv represents the convolution layer, ReLU is the filters is to find elementary visual features such as oriented
rectified linear unit, Pooling is the pooling layer, FC denotes edges, end-points, and corners. The visual features are then
the full-connected layer, and Dropout is the dropout layer. combined by the subsequent layers to detect higher-order
Based on the basic structure, we design the hyper-parameters, features [11]. We develop multiple LWRTs shown in Fig.
such as the input size, convolution filter size, the number of 1 for various input sizes to the CNN, such as [64×64],
Conv feature maps, and the number of neurons in the FC, to [128×128], and [256×256], and for various numbers of con-
find the optimal structure for the LPI waveform classification volution filters used in the first/second convolutional layers,
problem based on numerous Monte Carlo simulations for such as 10/20, 20/40, 30/60, and 40/80, and test the LWRTs
various conditions. for twelve LPI radar waveforms with low SNR, such as
Due to the large number of independent parameters to −4dB, −6dB, and −8dB. Note that the choice of the test
determine, we exploit the Conv filter size [5x5] made in SNR range is effective since the performance of the LWRTs
previous studies [11], [23], [24] for an initial choice. How- starts to degrade for SNRs lower than −4dB and drops to
ever, we use 400 as an initial value (large enough) for the a very low percentage of the correct classification (pcc) for
number of neurons in the FC layer, which is different from SNRs lower than −8dB.
the number of neurons (i.e., 100 ∼ 200) used in the previous Table 2 shows some of the simulation results to testify the
studies [11], [23], [24], [25]. This initial choice of neurons is design of the proposed CNN. As shown, the input size of
not only because the input CWD-TFI size is much larger than [128×128] and the numbers of convolution filters for the first
the image size used in the previous studies [11], [23], [24], and second layers equal to 30 and 60, respectively, produce
8 VOLUME 4, 2016

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

TABLE 3: Confusion matrix for a comparison to [9]

100
No
BFSK LFM Frank P1 P2 P3 P4
Mod

pcc [%]
80 Proposed technique
No 100 0 0 0 0 0 0 0
Mod (95) (0) (5) (0) (0) (0) (0) (0)
0 100 0 0 0 0 0 0 Technique in [8]
BFSK 60
(5) (95) (0) (0) (0) (0) (0) (0) -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB]
0 0 100 0 0 0 0 0 (a) Overall
LFM
(5) (0) (95) (0) (0) (0) (0) (0)
0 0 0 100 0 0 0 0 100 100
Frank
(0) (0) (0) (85) (5) (0) (10) (0)

pcc [%]
0 0 0 0 100 0 0 0

pcc [%]
80
P1 80
(0) (0) (0) (5) (85) (0) (0) (10)
60
0 0 0 0 0 100 0 0
P2
(0) (0) (0) (0) (0) (95) (5) (0) 60
40
-10 -8 -6 -4 -2 0 2 4 6 8 10
0 0 0 0 0 0 100 0 -10 -8 -6 -4 -2 0 2 4 6 8 10
P3 SNR [dB] SNR [dB]
(0) (0) (0) (10) (0) (0) (85) (5) (b) LFM (c) BPSK
0 0 0 0 0 0 0 100
P4
(0) (0) (0) (0) (10) (5) (0) (85) 100 100

pcc [%]

pcc [%]
80

the best result. In the next 10 rows (i.e., from the 13th row to 60

the 22nd row) of the Table 2, we present the test results for 90 40
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
various filter sizes of the first and the second convolutional SNR [dB] SNR [dB]
layers, where it is assumed that the filter size of the first layer (d) Costas (e) Frank
is larger or equal to that of the second layer [11], [25]. It 100 100

turns out that the filter sizes of [7×7] and [3×3] for the two
pcc [%]

pcc [%]
80 80
layers produce the best result. As for the stride size in the
convolutional layers, since the CNN needs to extract features 60 60

from the subtle shapes of the image objects, we select a unit 40

-10 -8 -6 -4 -2 0 2 4 6 8 10
40
-10 -8 -6 -4 -2 0 2 4 6 8 10
stride size. For the determination of the hyper-parameters in SNR [dB] SNR [dB]
(f) P1 (g) P2
the pooling layers, we use the max pooling [26] with 2×2
filter size, stride size 2, and no zero-padding, which are the 100 100

common choice in the literature [23], [24]. Note also that

pcc [%]

pcc [%]
80 80
in the nonlinear layer (omitted in Fig. 6 for illustrational
60 60
simplicity), we use ReLU as a common choice. The last 7
rows of Table II shows that the performance of the LWRT 40
-10 -8 -6 -4 -2 0 2 4 6 8 10
40
-10 -8 -6 -4 -2 0 2 4 6 8 10

is the best when there are 200 neurons in the FC-2. Note SNR [dB] SNR [dB]
(h) P3 (i) P4
that there is a Dropout layer (of 50% rate) omitted in Fig.
FIGURE 7: Comparison with the LWRT in [8].
6 for illustrational simplicity, between the FC-1 and FC-2 to
avoid a possible overfitting problem [27]. The performance
with and without the Dropout layer is 93.67% and 91.51%, samples are collected for each trial, and the tested SNR
respectively. The details of the final design of the CNN are levels are from −10dB to 10dB. To apply the SAT, we use
described in Fig. 6. Na =10 which results in Nsc ≥ 4 and the CWD-TFIs of the
averaged signals are used in the testing phase. Table 3 shows
V. PERFORMANCE DEMONSTRATION AND the simulation result for SNR = −10dB, which is the same
COMPARISON TO THE CONVENTIONAL TECHNIQUES condition introduced in [9]. We observe that the proposed
In this section, we compare the performance of the proposed LWRT achieves perfect performance in classifying all of the
LWRT to the recent LWRTs [6], [8], [9] introduced in the eight waveform modulations, whereas the performance of the
literature with numerous Monte Carlo simulations. LWRT in [9], shown within parentheses, is 90% in average.

A. PERFORMANCE COMPARISON TO THE LWRT IN [9] B. COMPARISON WITH THE LWRT IN [8]
The first performance comparison is between the proposed The second performance comparison is between the pro-
LWRT with the SAT and the LWRT in [9] that utilizes FRT to posed LWRT without the SAT and the recent LWRT in [8].
reduce the computational cost for processing a large amount The waveform modulation schemes considered in [8] are
of signal samples. The waveform modulations considered LFM, BPSK, Costas, Frank, P1, P2, P3, and P4 codes, and the
are no modulation, BFSK, LFM, Frank, P1, P2, P3, and same simulation conditions in [8] are used for the simulations
P4 codes, and the same simulation conditions including the of the proposed LWRT. Fig. 7 shows that the classification
number of signal samples in [9] are used in the simulations performance of the proposed LWRT is superior to the LWRT
of the proposed LWRT with the SAT. Therefore, N=10, 000 in [8] by about 3dB in overall.
VOLUME 4, 2016 9

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

100

pcc [%]
Proposed technique

Technique in [6]
80
-10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB]
(a) Overall

100 100

pcc [%]

pcc [%]
90
80
FIGURE 8: CWD-TFI of BPSK signals in [6]. 80

70 60
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB] SNR [dB]
(b) LFM (c) BPSK

100 100

pcc [%]

pcc [%]
90

80

90 70
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB] SNR [dB]
(d) Costas (e) Frank

100 100

(a) (b)
pcc [%]

pcc [%]
90 80

FIGURE 9: Incorrect CWD-TFIs of T4 code. (a) Copy of the CWD-TFI of T4 code

in [6], (b) CWD-TFI of T4 code generated by the pseudo-code given in [2]. 80 60

70 40
-10 -8 -6 -4 -2 0 2 4 6 8 10 -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB] SNR [dB]
C. COMPARISON WITH THE LWRT IN [6] (f) T1 (g) T2

The third performance comparison is between the proposed 100 100

LWRT without the SAT and the CNN-based LWRT in [6].

pcc [%]

pcc [%]
90

The waveform modulation schemes considered in [6] are

80
LFM, BPSK, Costas, Frank, T1, T2, T3, and T4 codes,
and we use the same simulation conditions in [6] for the 80
-10 -8 -6 -4 -2 0 2 4 6 8 10
70
-10 -8 -6 -4 -2 0 2 4 6 8 10

simulations of the proposed LWRT. However, there are two SNR [dB] SNR [dB]
(h) T3 (i) T4
simulation conditions we do not follow [6]. Firstly, we do not
FIGURE 10: Comparison with the LWRT in [6].
assume that the BPSK waveform samples are obtained for
multiple consecutive periods as assumed in [6]. Fig. 8 shows
an example of a CWD-TFI of the BPSK waveform repeating
5 times. Therefore, the generated CWD-TFI of the BPSK should be as shown in Fig. 4(l). As a result, the correct CWD-
waveform to test the proposed LWRT should be similar to TFI of T4 code can be confusing with that of T2 code shown
the noise version of Fig. 4(c). Note that this condition for in Fig. 4(j) in noisy conditions, but the shape of the incorrect
the proposed LWRT makes the classification more difficult, CWD-TFI of the T4 code in Fig. 9(a) may be not easily
since the CWD-TFI of the repeating BPSK has a unique confused with other LPI radar waveforms shown in Fig.4
and distinctive shape when compared to other waveforms.
Secondly, we do not agree with the CWD-TFI of T4 code Fig. 10 shows the result of the proposed LWRT (without
shown in [6], which is copied in Fig. 9(a). In fact, the CWD- the SAT) compared to the result in [6], where the proposed
TFI of T4 code shown in Fig. 9(b), very similar to Fig. 9(a), LWRT has about 5dB improvement in overall. Notice that
is generated by using the pseudo-code given in the appendix there is no result from [6] for SNR below −4dB. This signifi-
of [2], where the phase offset follows cant improvement is because the proposed LWRT utilizes the
grayscale input that has amplitude information preserved in
( ) the CWD-TFI, and the input size and the hyper-parameters
2π j Nps B(kTs )2 Nps fc (kTs ) k of the CNN are designed to maximize the classification
φ[k] = mod − ,2π .
Nps 2τpw 2 performance. These improvements also allow the proposed
LWRT to classify a larger number of modulation schemes
However, the correct formula for the phase offset of the T4 (i.e., 12 in total) as presented in the next subsection.
code is stated in Table 1, which is the exact mathematical
definition in [2] and [14], and the CWD-TFI of T4 code
10 VOLUME 4, 2016

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

TABLE 4: Signal parameters and simulation conditions

Radar
Parameters Value of Range
waveforms

100
fc U (fs /6, fs /5)
80
LFM B U (fs /20, fs /16)

pcc [%]
N U [512, 1920] 60 With SAT
40

FH sequence {3, 4, 5, 6} 20
Without SAT
Costas fmin U (fs /30, fs /24) 0
-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
N U [512, 1920] SNR [dB]
(a) Overall
Lc {7, 11, 13}
100 100
BPSK fc U (fs /6, fs /5)
80 80
Ncc U [20, 24]

pcc [%]

pcc [%]
60 60

fc U (fs /6, fs /5) 40 40

Frank , P1 Ncc {3, 4, 5} 20 20

M {6, 7, 8} 0
-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
0
-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB] SNR [dB]
fc U (fs /6, fs /5) (b) LFM (c) BPSK
P2 Ncc {3, 4, 5}
100 100
M {6, 8}
80 80

pcc [%]

pcc [%]
60 60
fc U (fs /6, fs /5)
P3, P4 Ncc {3, 4, 5} 40 40

ρ {36, 49, 64} 20 20

0 0
-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
fc U (fs /6, fs /5) SNR [dB] SNR [dB]
T1, T2 Ng {4, 5, 6} (d) Costas (e) Frank
N U [512, 1920]
100 100

80 80
fc U (fs /6, fs /5)
pcc [%]

pcc [%]
B U (fs /20, fs /10) 60 60
T3, T4
Ng {4, 5, 6} 40 40

N U [512, 1920] 20 20

0 0
-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB] SNR [dB]
(f) P1 (g) P2
D. CLASSIFICATION PERFORMANCE OF THE 100 100

PROPOSED LWRT FOR ALL OF THE TWELVE LPI 80 80

pcc [%]

pcc [%]

60 60
RADAR WAVEFORMS.
40 40
In addition to the performance comparison of the proposed 20 20

LWRT to the recent LWRTs for a subset of LPI radar wave- 0

-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
0
-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
forms, we show the performance of the proposed LWRT with SNR [dB] SNR [dB]
(h) P3 (i) P4
and without the SAT for all of the twelve waveforms [2] in
this subsection. 100 100

80 80
Table 4 defines the signal parameters and simulation con-
pcc [%]

pcc [%]

60 60
ditions used for the performance evaluation of the proposed 40 40

LWRT without the SAT, where fc , B, N , fmin , Lc , Ncc , 20 20

M , ρ and Ng represent the center frequency, bandwidth of 0

-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
0
-20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB] SNR [dB]
the intercepted signal, number of collected signal samples, (j) T1 (k) T2
fundamental frequency of the Costas code, code length,
100 100
number of samples per cycle of the carrier frequency, number 80 80
of frequency steps (or number of subcodes in one frequency
pcc [%]

pcc [%]

60 60

step), number of the subcode in a code, and number of 40 40

segments in a code, respectively. The parameter range in 20 20

0 0
Table 4 is selected to satisfy the conditions such as Nsc = 4 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10
SNR [dB] SNR [dB]
and Ns (= 2048) > N and to allow much wider variations (l) T3 (m) T4
of the signal than those in subsections V-A, V-B, and V-C. FIGURE 11: Performance comparison of the proposed LWRT with and without
Notice that the bandwidth Bs of the receiver is assumed to the SAT for all of the twelve LPI radar waveforms in [2].

be larger or equal to that of the LPI radar signal B and that

the parameters related to the frequency and bandwidth of
the signal are expressed in terms of fs . In addition, to avoid
VOLUME 4, 2016 11

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

TABLE 5: Confusion matrix for SNR=−6dB

situations where the intercepted signals are easily classified
with the difference between the center frequencies of the BPSK Costas LFM Frank P1 P2 P3 P4 T1 T2 T3 T4
intercepted signals in the CWD-TFI, the center frequencies BPSK 99 0 0 0 0 0 0 0 1 0 0 0
of the twelve waveforms are set to be the same. Costas 0 99 0 0 0 0 0 0 1 0 0 0
For the training phase of the CNN in the proposed LWRT,
LFM 0 0 94 0 4 0 0 2 0 0 0 0
we generate 2700 signals for each of the twelve waveforms
with SNR ranging from −6dB to 10dB at a step of 2dB. Frank 0 0 0 89 0 0 11 0 0 0 0 0
Therefore, there are a total 32,400 signals generated, and P1 0 0 2 1 86 0 0 10 0 0 0 1
we divide the training signal set into two groups: 70% (i.e., P2 0 0 0 0 0 100 0 0 0 0 0 0
22,680) for the training set and 30% (i.e., 9,720) for the
P3 0 0 1 7 0 0 91 1 0 0 0 0
validation set. The validation set, that is used to estimate
the generalization error during the training phase, allows P4 0 0 4 0 9 0 0 85 1 0 0 1
updating the hyper-parameters accordingly [28]. Then, 1600 T1 1 0 0 0 0 0 0 0 97 0 2 0
test signals for each of the twelve LPI radar waveforms, T2 1 0 0 0 1 0 0 2 0 94 0 2
with SNR ranging from −20dB to 10dB at a step of 2dB, T3 0 0 0 0 0 0 3 0 1 0 95 1
are created to test the performance of the proposed LWRT.
T4 0 0 0 0 1 0 0 2 0 3 0 94
Therefore, there is a total 19,200 test data set generated
to test the proposed LWRT. Note that we do not train the
proposed LWRT by signals with SNR lower than −6dB since
the CWD-TFIs becomes too noisy and they may confuse the SAT to the intercepted LPI signal samples can effectively
system. Note also that we only use the signal set generated reduce the number of signal samples to be processed by
without using the SAT for the training phase. the intercept receiver and that the technique achieves a
In the tests of the proposed LWRT with the SAT, we significant performance improvement without increasing the
generate signal samples at Na (= 10) times higher sampling computation cost. Therefore, the proposed LWRT with and
frequency than fs , and use the signal samples to test the without the SAT can be useful for LPI radar signal detection
proposed LWRT with the SAT. For the proposed LWRT and classification in practice.
without the SAT, we down-sample the generated signal by Na
times to perform the test. We use Nsc = 4, and the simulation
results shown in Fig. 11 testify that there is more than 9dB
performance improvement by using the SAT with Na = 10 REFERENCES
and Nsc = 4. Note that slightly less than 10dB performance [1] R. G. Wiley, ELINT: The interception and analysis of radar signals.
improvement is expected from the slight SNR loss due to Artech House, 2006.
the SAT. Notice that the computation complexities of the [2] P. E. Pace, Detecting and classifying low probability of intercept radar.
Artech House, 2009.
proposed LWRTs (with and without the SAT) are similar, and [3] G. López-Risueño, J. Grajal, and A. Sanz-Osorio, “Digital channelized
that the performance of the proposed LWRT without the SAT receiver based on time-frequency analysis for signal interception,” IEEE
is slightly worse than the performance results shown in the Transactions on Aerospace and Electronic Systems, vol. 41, no. 3, pp. 879–
898, 2005.
previous subsections V-A, V-B, and V-C. This is because the
[4] C. Wang, J. Wang, and X. Zhang, “Automatic radar waveform recognition
proposed LWRT in this subsection is classifying all of the based on time-frequency analysis and convolutional neural network,” in
twelve waveforms, whereas, in the previous subsections, the Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE Interna-
proposed LWRT is required to classify only eight waveforms. tional Conference on. IEEE, 2017, pp. 2437–2441.
[5] E. R. Zilberman and P. E. Pace, “Autonomous time-frequency morpholog-
Table 5 shows the confusion matrix of the proposed LWRT ical feature extraction algorithm for lpi radar modulation classification,” in
without the SAT for SNR = −6dB. As shown, polyphase Image Processing, 2006 IEEE International Conference on. IEEE, 2006,
codes have low pcc, because the CWD-TFIs of the P1 and pp. 2321–2324.
[6] M. Zhang, M. Diao, and L. Guo, “Convolutional neural networks for
P4 codes can be confused with that of the LFM, and those of automatic cognitive radio waveform recognition,” IEEE Access, vol. 5, pp.
Frank and P3 codes can be confused, when the SNR is low. 11 074–11 082, 2017.
[7] J. Lundén and V. Koivunen, “Automatic radar waveform recognition,”
VI. CONCLUSION IEEE Journal of Selected Topics in Signal Processing, vol. 1, no. 1, pp.
124–136, 2007.
A CNN-based LPI radar signal classification technique has [8] M. Zhang, L. Liu, and M. Diao, “Lpi radar waveform recognition based
been proposed and tested for twelve intra-pulse modulated on time-frequency distribution,” Sensors, vol. 16, no. 10, p. 1682, 2016.
LPI radar waveforms introduced in the literature. The hyper- [9] T. R. Kishore and K. D. Rao, “Automatic intrapulse modulation classifica-
tion of advanced lpi radar waveforms,” IEEE Transactions on Aerospace
parameters of the proposed technique has been determined and Electronic Systems, vol. 53, no. 2, pp. 901–914, 2017.
to guarantee the maximum performance based on various [10] Y. Bengio, A. Courville, and P. Vincent, “Representation learning: A
simulations, and the proposed technique has demonstrated review and new perspectives,” IEEE transactions on pattern analysis and
machine intelligence, vol. 35, no. 8, pp. 1798–1828, 2013.
with numerous Monte Carlo simulations that it achieves a
[11] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
superior performance than the recent LWRTs introduced in applied to document recognition,” Proceedings of the IEEE, vol. 86, no. 11,
the literature. In addition, it has been found that applying the pp. 2278–2324, 1998.

12 VOLUME 4, 2016

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2017.2788942, IEEE Access

S.-H. Kong and M. Kim et al.: Automatic LPI Radar Waveform Recognition using CNN

[12] K. G. Sheela and S. N. Deepa, “Review on methods to fix number of hid- Seung-Hyun Kong (M’06-SM’16) received the
den neurons in neural networks,” Mathematical Problems in Engineering, B.S. degree in Electronics Engineering from So-
vol. 2013, 2013. gang University, Korea, in 1992, the M.S. degree
[13] N. Levanon and E. Mozeson, Radar signals. John Wiley & Sons, 2004. in Electrical Engineering from Polytechnic Uni-
[14] M. I. Skolnik, “Introduction to radar,” Radar Handbook, vol. 2, 1962. versity, New York, in 1994, and the Ph.D. degree
[15] J. Starzyk and Z. Zhu, “Averaging correlation for c/a code acquisition and in Aeronautics and Astronautics from Stanford
tracking in frequency domain,” in Circuits and Systems, 2001. MWSCAS University, CA, Jan. 2006. From 1997 to 2004, he
2001. Proceedings of the 44th IEEE 2001 Midwest Symposium on, vol. 2.
was with Samsung Electronics Inc.and Nexpilot
IEEE, 2001, pp. 905–908.
Inc., both in Korea, where he worked on devel-
[16] S.-H. Kong, “Sdht for fast detection of weak gnss signals,” IEEE Journal
on Selected Areas in Communications, vol. 33, no. 11, pp. 2366–2378, oping wireless communication system standards
2015. and UMTS mobile positioning technologies. In 2006 and from 2007 to
[17] H.-I. Choi and W. J. Williams, “Improved time-frequency representation 2009, he was a staff engineer at Polaris Wireless Inc., Santa Clara, and at
of multicomponent signals using exponential kernels,” IEEE Transactions Qualcomm Inc. (Corp. R&D), San Diego, respectively, where his research
on Acoustics, Speech, and Signal Processing, vol. 37, no. 6, pp. 862–871, was on Assisted-GNSS and wireless positioning technologies. Since 2010,
1989. he is with Korea Advanced Institute of Science and Technology (KAIST),
[18] P. Flandrin, “Some features of time-frequency representations of multi- where he is currently an associate professor at the CCS Graduate School
component signals,” in Acoustics, Speech, and Signal Processing, IEEE of Green Transportation. He serves as an Editor of IET Radar, Sonar and
International Conference on ICASSP’84., vol. 9. IEEE, 1984, pp. 266– Navigation, and an Associate Editor of IEEE Transactions on Intelligent
269. Transportation Systems and IEEE Access. His research interests include
[19] J. E. Fielding, “Polytime coding as a means of pulse compression,” IEEE signal processing for GNSS, neutral networks for sensing, and vehicular
Transactions on Aerospace and Electronic Systems, vol. 35, no. 2, pp. 716– communication systems.
721, 1999.
[20] J. A. Parker, R. V. Kenyon, and D. E. Troxel, “Comparison of interpolating
methods for image resampling,” IEEE Transactions on medical imaging,
vol. 2, no. 1, pp. 31–39, 1983.
[21] T. M. Lehmann, C. Gonner, and K. Spitzer, “Survey: Interpolation methods
in medical image processing,” IEEE transactions on medical imaging,
vol. 18, no. 11, pp. 1049–1075, 1999.
[22] Y. LeCun, K. Kavukcuoglu, and C. Farabet, “Convolutional networks and
applications in vision,” in Circuits and Systems (ISCAS), Proceedings of
2010 IEEE International Symposium on. IEEE, 2010, pp. 253–256.
[23] D. Ciregan, U. Meier, and J. Schmidhuber, “Multi-column deep neural
networks for image classification,” in Computer Vision and Pattern Recog-
Minjun Kim received the B.S. degree in Elec-
nition (CVPR), 2012 IEEE Conference on. IEEE, 2012, pp. 3642–3649.
tornics Engineering from Chung-Ang University,
[24] P. Y. Simard, D. Steinkraus, J. C. Platt et al., “Best practices for convo-
lutional neural networks applied to visual document analysis.” in ICDAR, Korea, in 2017. He is currently pursuing the M.S.
vol. 3, 2003, pp. 958–962. degree at the CCS Graduate School of Green
[25] C. Poultney, S. Chopra, Y. L. Cun et al., “Efficient learning of sparse repre- Transportation in the Korea Advanced Institute of
sentations with an energy-based model,” in Advances in neural information Science and Technology (KAIST), Korea. His re-
processing systems, 2007, pp. 1137–1144. search interests include signal processing, Radar,
[26] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification Deep learning and V2X for autonomous vehicle.
with deep convolutional neural networks,” in Advances in neural informa-
tion processing systems, 2012, pp. 1097–1105.
[27] N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut-
dinov, “Dropout: a simple way to prevent neural networks from overfit-
ting.” Journal of machine learning research, vol. 15, no. 1, pp. 1929–1958,
2014.
[28] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
no. 7553, pp. 436–444, 2015.

Linh Manh Hoang received the B.S. degree in

Mechatronics Engineering from Hanoi University
of Science and Technology, Vietnam, in 2012. He
is currently pursuing the M.S. degree at the CCS
Graduate School of Green Transportation in the
Korea Advanced Institute of Science and Tech-
nology (KAIST), Korea. His research interests
include signal processing, Radar, Deep learning
and V2X for autonomous vehicle.

VOLUME 4, 2016 13

2169-3536 (c) 2017 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
View publication stats http://www.ieee.org/publications_standards/publications/rights/index.html for more information.