IEEE - Aerospace and Electronic Systems - August 2023 - 26

Feature Article:
DOI. No. 10.1109/MAES.2023.3268020
Incremental Deinterleaving of Radar Emitters
Stefan Scholl and Stefan Br€uggenwirth , Fraunhofer Institute for High
Frequency Physics and Radar Techniques (FHR), 53343 Wachtberg, Germany
INTRODUCTION
Systems for electronic support (ES) aim to detect and
identify radar emitters present in the electromagnetic
spectrum. For that purpose, three main processing steps
can be identified: detection of radar pulses, pulse deinterleaving,
and emitter classification (including specific emitter
identification). This article considers the task of
deinterleaving, which groups detected radar pulses, that
belong to the same emitter. Successful deinterleaving is a
prerequisite for accurate emitter classification.
Typically, deinterleaving is based on pulse descriptor
words (PDW) [16]. A PDW contains a set of several properties
for each detected radar pulse, such as time-of-arrival
(TOA), radio frequency (RF), angle-of-arrival (AOA), or
pulse width (PW).
ES systems usually collect a batch of pulses or PDW
data for some period of observation time, for example,
some hundred milliseconds. After this time, the deinterleaver
processes the batch ofPDWs and forwards them to the
emitter classifier. After that, the PDWs of the next period
of observation time are collected and processed. We refer
to this processing scheme as nonincremental (see
Figure 1). Nonincremental processing introduces a delay
in the reaction time of system, that is, the time from the
first reception of an emitter to the emitter recognition:
Radar emitters can only be detected after the observation
time period has finished. The deinterleaver reaction time
is in average approximately half the observation time plus
the deinterleaver processing time.
To reduce the reaction time, a deinterleaver is required
to output its result more frequently-not only at the end of
the observation time. Such an deinterleaver is called
Authors' current address: Stefan Scholl and Stefan
Br€uggenwirth are with the Fraunhofer Institute for
High Frequency Physics and Radar Techniques (FHR),
Wachtberg 53343, Germany (e-mail: stefan.scholl@fhr.
fraunhofer.de; stefan.brueggenwirth@fhr.fraunhofer.de).
Manuscript received 19 May 2022, revised 21 December
2022; accepted 6 April 2023, and ready for publication
26 April 2023.
Review handled by Peter Willett.
0885-8985/23/$26.00 ß 2023 IEEE
26
incremental and must be able to handle additional pulses
after an intermediate deinterleaving result has been output.
In this article, we consider the ultimate case, where
deinterleaving is performed after each pulse. We call this
pulse-to-pulse incremental deinterleaving. Pulse-to-pulse
deinterleaving enables maximal fast reaction on the occurrence
of radar emitters. It allows to detect an emitter
already during its first illumination of the receiver rather
than only at the end of the observation time. Figure 1
depicts the difference between nonincremental and pulseto-pulse
incremental deinterleaving concepts.
RELATED WORK AND CONTRIBUTION
Early works on deinterleaving put a strong focus on utilizing
TOA data to detect typical patterns of pulse repetition
intervals (PRI) [5], [10]. These approaches based on PRI
relies on time difference histograms. These histogram
methods require the knowledge of many pulses before
deinterleaving can take place. Therefore, these methods
are not well suited for incremental processing.
Other approaches include clustering algorithms to group
pulses with similar PDW properties, for example, PW and
RF. However, popular clustering algorithms, such as kmeans
or agglomerative clustering, are not suited for deinterleaving.
These algorithms require in their basic variants the
number of clusters (i.e., the number of radar emitters) to be
known in advance, which is typically not available.
More advanced clustering algorithms, such as
DBSCAN, do not require the number of clusters in
advance and have been successfully applied to deinterleaving
[14], [15]. However, the algorithm in Mottier
et al. [15] apparently uses very long observation times in
the order of seconds, which is contrary to incremental
processing. In Mahmod, [14] a partially incremental
approach is presented, but it requires a training phase with
a standard nonincremental DBSCAN and is therefore also
not able to perform pulse-to-pulse deinterleaving.
Algorithms that can naturally perform incremental
clustering are neural nets based on adaptive resonance theory
(ART). These networks have been applied to deinterleaving
in [9] and [1], but are trained with several
epochs. It means, that all pulses from an observation
IEEE A&E SYSTEMS MAGAZINE
AUGUST 2023
https://orcid.org/0000-0002-3152-4072 https://orcid.org/0000-0003-1118-0433 http://www.fraunhofer.de
IEEE - Aerospace and Electronic Systems - August 2023

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