IEEE - Aerospace and Electronic Systems - April 2023 - 14

Feature Article:
DOI. No. 10.1109/MAES.2023.3238704
LPI Waveform Recognition Using Adaptive Feature
Construction and Convolutional Neural Networks
Hui Huang , Yi Li, Jiaoyue Liu , Dan Shen , and Genshe Chen , Intelligent
Fusion Technology, Inc., Germantown, MD 20876 USA
Erik Blasch
, Air Force Research Laboratory, Arlington, VA 22203 USA
Khanh Pham , Air Force Research Laboratory, Kirtland AFB, NM 87117 USA
INTRODUCTION
Low probability of interception (LPI) radar signal is
designed to have a low probability of detection and recognition
by intercept receivers via various waveform construction
approaches. Currently, recognizing the waveform
of the LPI radar signals is one of the crucial functions for
the intercept receivers in modern electronic warfare applications,
such as radar emitter recognition, cognitive radar,
and threat detection [1]. Recent years, with the development
of artificial intelligence, numerous approaches have
been proposed for the LPI waveform recognition based on
various machine learning (ML) and signal analysis techniques
[2], [3], [4], [5], [6]. The signal analysis techniques
are typically utilized to extract the features of the LPI
waveforms and the ML techniques are further performed to
achieve the LPI waveform recognition by classifying them
with the extracted features [7], [8], [9], [10].
For the feature extraction, different kinds of signal
analysis techniques are used including time-frequency
analysis, autocorrelation functions, instantaneous frequency
analysis, and so on. In [8], the time-frequency technique
of Wigner-Ville distribution (WVD) coupled with
the Convolutional Neural Network (CNN) achieves LPI
Authors' addresses: Hui Huang, Yi Li, Jiaoyue Liu, Dan
Shen, and Genshe Chen are with Intelligent Fusion Technology,
Inc., Germantown, MD 20876 USA (e-mail: hui.
huang@intfusiontech.com, yi.li@intfusiontech.com, jay.
liu@intfusiontech.com, dshen@intfusiontech.com, and
gchen@intfusiontech.com). Erik Blasch is with the Air
Force Research Lab, Arlington, VA 22203 USA (e-mail:
erik.blasch.1@us.af.mil). Khanh Pham is with the Air
Force Research Laboratory, Kirtland AFB, NM 87117 USA
(e-mail: khanh.pham.1@us.af.mil).
Manuscript received 25 January 2022, revised 3 October
2022; accepted 16 January 2023, and ready for publication
23 January 2023.
Review handled by Stefan Br€uggenwirth.
0885-8985/23/$26.00 ß 2023 IEEE
14
waveform recognition. Other time-frequency techniques,
such as Choi-William distribution (CWD) and wavelet
transform, have also been extensively used for the LPI
waveform recognition [10], [11], [12]. In addition to the
time-frequency techniques, the autocorrelation functionsbased
approaches are developed in [13], [14], where the
instantaneous frequency analysis is utilized in [15] and the
high-order spectral analysis is applied in [16] to extract frequency
features, and the principle component analysis is
performed in [17] for the radar waveform classification.
For the classification techniques, various popularly used
ML methods are used to classify the extracted features
including support vector machine [17], hierarchical decision
tree [18], recurrent neural network[19], and CNN [10].
As the LPI signal mostly have spread bandwidth and
relatively small peak power for staying silent from the noncooperate
receivers, the low SNR condition of the intercepted
LPI signal is an inevitable challenge for the LPI
signal recognition. Although the existing studies described
previously make a great improvement for the LPI recognition
performance, most ofthem suffer in a mediocre performance
in low signal-to-noise ratio (SNR) environments.
According to the results of several studies [7], [8], [10],
[20], [21], the LPI waveform recognition performance
shows a severe degradation with the SNRwhen the LPI signal
decreases below -4 dB [7]. One of the reasons may be
that most of the well-defined analytical expressions were
developed under rigorous mathematical rules, which
mainly focus on discovering efficient distinctions for classification.
These techniques may be inadequate for revealing
the physical mechanisms hidden in signals and may
also introduce irrelevant interference, such as the cross
term in WVD [1].
In order to tackle the challenge ofLPI signal recognition
in low SNR environments, we propose to use the adaptive
features, which are the main components extracted from
adaptive signal decomposition approaches including empirical
mode decomposition (EMD) [22] and variational mode
decomposition (VMD) [23]. Since the adaptive features are
IEEE A&E SYSTEMS MAGAZINE
APRIL 2023
https://orcid.org/0000-0002-2656-9380 https://orcid.org/0000-0003-4474-7037 https://orcid.org/0000-0001-5777-1541 https://orcid.org/0000-0003-1834-5456 https://orcid.org/0000-0001-6894-6108 https://orcid.org/0000-0002-4768-9365
IEEE - Aerospace and Electronic Systems - April 2023

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - April 2023
