The problem of developing the architecture of modern cognitive radar system
The problem of developing the architecture of modern cognitive radar systems using artificial intelligence technologies is considered. The main difference from traditional systems is the use of a trained neural network. The heterogeneous multiprocessor system is rebuilt in the process of solving the...
Gespeichert in:
| Datum: | 2022 |
|---|---|
| Hauptverfasser: | , |
| Format: | Artikel |
| Sprache: | English |
| Veröffentlicht: |
PROBLEMS IN PROGRAMMING
2022
|
| Schlagworte: | |
| Online Zugang: | https://pp.isofts.kiev.ua/index.php/ojs1/article/view/490 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Назва журналу: | Problems in programming |
| Завантажити файл: |  |
Institution
Problems in programming| id |
pp_isofts_kiev_ua-article-490 |
|---|---|
| record_format |
ojs |
| resource_txt_mv |
ppisoftskievua/7c/e5764afece68c277510346658c338e7c.pdf |
| spelling |
pp_isofts_kiev_ua-article-4902022-07-12T19:40:17Z The problem of developing the architecture of modern cognitive radar system Проблеми розробки архітектури сучасних когнітивних радіолокаційних систем Коsovets, M. Tovstenko, L. Perception-Action Cycle;Artificial Intelligence; Signal to Noise Ratio; Active Electronically Scanned Array;Environmental Dynamic Database; Signal to Noise Ratio; Radar Resource Management; multiprocessor UDC 517.9:621.325.5:621.382.049.77 штучний інтелект; нейронна когнітивна мережа; сенсорні мережеві додатки УДК 517.9:621.325.5:621.382.049.77 The problem of developing the architecture of modern cognitive radar systems using artificial intelligence technologies is considered. The main difference from traditional systems is the use of a trained neural network. The heterogeneous multiprocessor system is rebuilt in the process of solving the problem, providing reliability and solving various types of problems of one class and deep learning of the neural network in real time. This architecture promotes the introduction of cognitive technologies that take into account the requirements for the purpose, the influence of external and internal factors.Problems in programming 2022; 4: 75-86 Розглянуто проблему розробки архітектури сучасних когнітивних радіолокаційних систем із використанням технологій штучного інтелекту. Основною відмінністю від традиційних систем є використання навченої нейронної мережі. Гетерогенна багатопроцесорна система перебудовується в процесі розв’язування задачі, забезпечуючи надійність і вирішення різних типів задач одного класу і глибоке навчання нейронної мережі в режимі реального часу. Така архітектура сприяє впровадженню когнітивних технологій, які враховують вимоги по призначенню, вплив зовнішніх і внутрішніх факторів. Глибоке навчання нейронної когнітивної мережі радарних датчиків є функцією штучного інтелекту, який моделює роботу людського мозку таким чином, що обробляє дані та створює шаблони, які використовуються у прийнятті рішень. Система виявлення вчиться виявляти зміни не тільки в рівнях сигналу, а й у формі та параметрах сигналу. Експерименти показали, що система виявлення змін радіолокаційної інформації на основі нейронної мережі з глибоким навчанням є оптимальною для розробки сенсорних мережевих додатків і може бути успішно реалізована на доступних технологічних платформах.Problems in programming 2022; 4: 75-86 PROBLEMS IN PROGRAMMING ПРОБЛЕМЫ ПРОГРАММИРОВАНИЯ ПРОБЛЕМИ ПРОГРАМУВАННЯ 2022-05-30 Article Article application/pdf https://pp.isofts.kiev.ua/index.php/ojs1/article/view/490 10.15407/pp2022.01.075 PROBLEMS IN PROGRAMMING; No 1 (2022); 075-086 ПРОБЛЕМЫ ПРОГРАММИРОВАНИЯ; No 1 (2022); 075-086 ПРОБЛЕМИ ПРОГРАМУВАННЯ; No 1 (2022); 075-086 1727-4907 10.15407/pp2022.01 en https://pp.isofts.kiev.ua/index.php/ojs1/article/view/490/489 Copyright (c) 2022 PROBLEMS IN PROGRAMMING |
| institution |
Problems in programming |
| baseUrl_str |
https://pp.isofts.kiev.ua/index.php/ojs1/oai |
| datestamp_date |
2022-07-12T19:40:17Z |
| collection |
OJS |
| language |
English |
| topic |
Perception-Action Cycle;Artificial Intelligence Signal to Noise Ratio Active Electronically Scanned Array;Environmental Dynamic Database Signal to Noise Ratio Radar Resource Management multiprocessor UDC 517.9:621.325.5:621.382.049.77 |
| spellingShingle |
Perception-Action Cycle;Artificial Intelligence Signal to Noise Ratio Active Electronically Scanned Array;Environmental Dynamic Database Signal to Noise Ratio Radar Resource Management multiprocessor UDC 517.9:621.325.5:621.382.049.77 Коsovets, M. Tovstenko, L. The problem of developing the architecture of modern cognitive radar system |
| topic_facet |
Perception-Action Cycle;Artificial Intelligence Signal to Noise Ratio Active Electronically Scanned Array;Environmental Dynamic Database Signal to Noise Ratio Radar Resource Management multiprocessor UDC 517.9:621.325.5:621.382.049.77 штучний інтелект нейронна когнітивна мережа сенсорні мережеві додатки УДК 517.9:621.325.5:621.382.049.77 |
| format |
Article |
| author |
Коsovets, M. Tovstenko, L. |
| author_facet |
Коsovets, M. Tovstenko, L. |
| author_sort |
Коsovets, M. |
| title |
The problem of developing the architecture of modern cognitive radar system |
| title_short |
The problem of developing the architecture of modern cognitive radar system |
| title_full |
The problem of developing the architecture of modern cognitive radar system |
| title_fullStr |
The problem of developing the architecture of modern cognitive radar system |
| title_full_unstemmed |
The problem of developing the architecture of modern cognitive radar system |
| title_sort |
problem of developing the architecture of modern cognitive radar system |
| title_alt |
Проблеми розробки архітектури сучасних когнітивних радіолокаційних систем |
| description |
The problem of developing the architecture of modern cognitive radar systems using artificial intelligence technologies is considered. The main difference from traditional systems is the use of a trained neural network. The heterogeneous multiprocessor system is rebuilt in the process of solving the problem, providing reliability and solving various types of problems of one class and deep learning of the neural network in real time. This architecture promotes the introduction of cognitive technologies that take into account the requirements for the purpose, the influence of external and internal factors.Problems in programming 2022; 4: 75-86 |
| publisher |
PROBLEMS IN PROGRAMMING |
| publishDate |
2022 |
| url |
https://pp.isofts.kiev.ua/index.php/ojs1/article/view/490 |
| work_keys_str_mv |
AT kosovetsm theproblemofdevelopingthearchitectureofmoderncognitiveradarsystem AT tovstenkol theproblemofdevelopingthearchitectureofmoderncognitiveradarsystem AT kosovetsm problemirozrobkiarhítekturisučasnihkognítivnihradíolokacíjnihsistem AT tovstenkol problemirozrobkiarhítekturisučasnihkognítivnihradíolokacíjnihsistem AT kosovetsm problemofdevelopingthearchitectureofmoderncognitiveradarsystem AT tovstenkol problemofdevelopingthearchitectureofmoderncognitiveradarsystem |
| first_indexed |
2025-12-02T15:42:57Z |
| last_indexed |
2025-12-02T15:42:57Z |
| _version_ |
1850411767287513088 |
| fulltext |
75
Моделі та методи машинного навчання
Introduction
Modern radar systems operate in a
wide variety of dynamically changing sce-
narios: the detection, tracking and classifi-
cation of very small and slow targets. Such
objects as drones, missiles, boats, along with
a complex spectrum are important system
requirements. Cognitive radars, combining
many well-known and new methods, offer
a promising solution to these problems. We
consider the functional architecture of cog-
nitive radar from the perspective of the user
and the manufacturer.
Cognitive radar is an updated tech-
nology, the origins of which go back to sci-
ence - «cybernetics», human-machine inter-
action, signal processing. The evolution of
cognitive radar is aimed at achieving cogni-
tion, as in its natural counterparts, such as
the radar capabilities of bats and dolphins,
or human intellectual decision-making. This
article provides an overview and trends in
the development of cognitive radar systems.
Organization of cognitive
computing
The term «cognitive radar» was first
introduced by Dr. Simon Haikin [1], follow-
ing the ideas of cognitive neurology, which
is based on works of cybernetics, artificial
neural networks, self-organized learning
and solutions of Bayesian theory. Engineer-
ing analogues for the implementation of
the main cognitive features identified by
Faster: memory, attention and intelligence
(PAC Perception-Action-Cycle: Cycle-per-
ception-action) have been proposed [2]. In
studies of cybernetics, Rasmussen [3], [4]
described human behavior in terms of three
levels: based on skills, rules and knowledge.
He described behavior-based behavior as a
subconscious that reflects basic signal pro-
cessing and generation blocks in a radar
system [5]. Rule-based behavior is used
in familiar situations. The basis of parallel
work is modeling and analysis of previous
experience.
Build cognitive radar developers
have inspired research in the field of bio-
mimetics. Artificial intelligence is modeled
on the basis of observations of living intel-
ligence. Thus, masters of echolocation -
bats and dolphins can detect and track very
small prey, using complex waveforms that
are changed dynamically [6]. Moreover,
knowledge of the intelligence of living be-
ings allows us to better understand living
nature. It helps to create artificial intelli-
gence, which is superior to «living» and is
used in technical systems.
Information in the radar system is
perceived by «smart» sensors, i.e. sensors
with primary processing and control of the
measurement process [7], as well as through
network sensors that demonstrate «distrib-
УДК517.9:621.325.5:621.382.049.77 http://doi.org/10.15407/pp2022.01.75
M. Коsovets, L. Tovstenko
THE PROBLEM OF DEVELOPING
THE ARCHITECTURE
OF MODERN COGNITIVE RADAR SYSTEM
The problem of developing the architecture of modern cognitive radar systems using artifi cial intelligence
technologies is considered. The main diff erence from traditional systems is the use of a trained neural
network. The heterogeneous multiprocessor system is rebuilt in the process of solving the problem, providing
reliability and solving various types of problems of one class and deep learning of the neural network in
real time. This architecture promotes the introduction of cognitive technologies that take into account the
requirements for the purpose, the infl uence of external and internal factors.
Keywords: Perception-Action Cycle, Artifi cial Intelligence, Signal to Noise Ratio, Active Electronically
Scanned Array, Environmental Dynamic Database, Signal to Noise Ratio, Radar Resource Management,
multiprocessor.
© M. Коsovets, L. Tovstenko, 2022
ISSN 1727-4907. Проблеми програмування. 2022. № 1
76
Моделі та методи машинного навчання
uted intelligence» with self-monitoring ca-
pabilities, automatic solution of changes in
their environment [8] , [9]. Cognitive radar
has the ability to adapt to transmission in
the probing process, imitating human per-
ception as an interactive process where the
cognitive entity responds to or changes its
behavior as a result of external stimuli.
In traditional radar systems, the flow
of information is one-way: the radar inter-
rogates the environment by transmitting a
fixed, predetermined pulse signal, regard-
less of any changes in the environment.
Adaptive processing is performed on recep-
tion, but the results of such processing do
not control any radar function for transmis-
sion. An overview of the cognitive direc-
tions of radar construction research over
the last decade gives an idea of the methods
being developed for a wide range of radar
applications. Technical problems in the de-
velopment of cognitive radars are the moti-
vation for further work in this area. Central
to these works is the idea of closed-loop
data collection, where the dynamic state
is interpreted as an adaptive measurement
determined by Kalman filtering. This ap-
proach allows the antenna array to be adap-
tively directed in the direction and width of
the beam, as well as to place zeros so as to
reject any unwanted signals or noise outside
the main particle.
A database of problems has been
developed that allows comparing methods
that use beam control of phased array an-
tennas to optimize tracking, minimizing
false alarms [10], [11]. Sometimes several
hypotheses and filtering interactions of sev-
eral tracking models are tested [12-13] to
optimize performance: such as signal-to-
noise ratio, interference effects, track, and
detection threshold.
Common features of increasing
adaptability: prediction of adaptive sched-
uling review time, adaptive choice of de-
tection thresholds (eg, constant false alarm
rate detectors) [14], and adaptive interfer-
ence suppression using adaptive-spatio-
temporal processing mode [15] to improve
target detection. Adaptive tracking meth-
ods vary the measurement time, as well as
the signals used to update the trajectory,
based on the measurements obtained by the
tracker. This feedback is used to control
the radar so that frequent measurements
are made during an unpredictable or rapid
dynamic maneuver, while infrequent mea-
surements are made during predicted peri-
ods or steady dynamics.
The simultaneous change of intra-
pulse signal modulation is studied on the
basis of the provided measurements on the
tracker. This leads to the choice of optimal
signal methods [16-17] and their adaptive
extensions [18-19]. Optimization of radar
signal in dynamics, to maximize perfor-
mance according to specific scenarios and
tasks, includes the use of some components
of radar, such as antenna, radiation pattern
(both transmission and reception), time,
frequency, coding and polarization. The
signals are selected from several classes
of signals, such as linear or nonlinear fre-
quency modulation, phase or encoding fre-
quency, and ultra broadband signals. This
also includes adapting parameters within
the signal class, such as changing the pulse
repetition interval, bandwidth, or center
frequency [20]. The optimal signal, which
maximizes the signal / noise, arises as a so-
lution of the generalized eigenvalue on the
waveform [21], developing in the frame-
work of «joint optimization of transmission
and reception by the choice of waveform.
The approach was used in Bayesian theory
of decision making and development, de-
signed to optimize the system by selecting
the signal at the transmitter and minimiz-
ing interference at the receiver. There are
also difficulties in choosing the criteria of
optimality and accurate distribution of in-
terference.
According to the IEEE, the defini-
tion of «cognitive radar» is a radar that has
the ability to learn: «Radar system, which
automatically generates a constant percep-
tion of the target scene and takes appropri-
ate action. It can use short-term and long-
term memory to increase the performance
of a given function. Compared to adaptive
radar, cognitive radar is trained to adapt
operating parameters as well as processing
parameters, and can do so over longer pe-
riods of time.”
77
Моделі та методи машинного навчання
Fig.1. FMCW Radar Imaging Cognitive
Ability Modeling Complex
with Deep Learning Package.
Cognitive radar diff ers from traditional
active radar due to the following features: de-
velopment of rules of conduct for self-orga-
nization through a process called experiential
learning, which is the result of long-term in-
teraction with the environment. According to
Charlish, cognitive radar is a radar system that
acquires knowledge and understanding of the
work environment through online assessment
and training from databases that contain con-
textual information. Cognitive radar uses this
knowledge to improve information: search,
data processing and management of radar re-
sources. With the development of cognitive
radars began a new era in the creation of mod-
ern radar systems. The number of publications
on the development of cognitive radar archi-
tecture with artifi cial intelligence is increas-
ing avalanche. Artifi cial intelligence is used in
the construction of neural networks, methods
of deep learning, signal processing, pattern
recognition, classifi cation. It should be noted
that elements of cognition in the construction
of radars have always been presented (power,
pulse width, repetition rate, modulation, etc.).
There is also the ideology of the neural net-
work, and accordingly their in-depth training
(multiprocessor with restructuring). The ex-
plosive growth of the latest developments in
radar systems is related to public demand (de-
fense, security, medicine, subsurface sound-
ing, mine search, unmanned aerial vehicles)
and the ability to meet them: the use of artifi -
cial intelligence technology and the develop-
ment of a new component base.
Cognitive radar methods use mimic
elements of human cognition, such as the
cycle of perception-action, deep learning, in-
telligence and the use of existing knowledge
[22]. Cognitive radar vision methods use ra-
dar spectrum [23], [24], radiation optimiza-
tion [25], tracking [26-28], beam control [29],
interference reduction [30], network [31], re-
source management [32] .
To develop a cognitive radar system
that adapts in real time, the multiprocessor
must be an integral part of the simulation tool.
It helps to analyze the behavior of the radar at
the simulation stage. The post-process stage
consists of two steps. In the fi rst step, a radar
sensor and a proven circuit are developed us-
ing non-adaptive settings. In the second step,
the multiprocessor is tested and confi gured,
replacing the Front-end sensor. Pre-recorded
raw datasets that include all radar parameters
are optimized. Such datasets use the same
measurement for all parameters of environ-
ment, target, and trajectory. At this stage of
development, the feedback cycle is closed by
obtaining an interval of coherent processing
of raw data relating to the selected optimal pa-
rameters in real time.
The modeling complex consists of
an adaptive radar sensor that perceives the
environment with optimized radar param-
eters and a multiprocessor that tracks the
target and selects the optimal radar param-
eters for each new measurement. The sensor
consists of an adaptive signal generator, a
radar interface, an ADC and a real-time sig-
nal converter and a display. The controller
consists of a Kalman filter tracker and an
optimizer that selects both optimal signals
and real-time processing parameters based
on the latest measurements. Multiprocessor
processing is simulated in Matlab and runs
on an Ubuntu Linux PC.
The radar data processing system
consists of three modules: FPGA, worksta-
tion and graphics processing unit. FPGA
provides primary signal processing and
hardware management. Its most important
role in signal processing is to perform digi-
tal down-conversion of the received signal
so that the true baseband can be transmitted
to the workstation for further processing.
Subsequent implementation of the appro-
78
Моделі та методи машинного навчання
priate fi lter on the FPGA can be useful for
freeing resources on the graphics processor
for its other tasks: fi ltering, discrete Fourier
transform, and processing the constant rate
of false alarms (CFAR). The detection task
is implemented on a graphics processor. The
CPU is an Intel multi-core processor.
LabView National Instruments sup-
port control functions. In addition, LabView
can be used to write FPGA software. Simula-
tion on FPGA Xilinx, RF-path on on-a-chip
is used. The shell is written in Python for the
C ++ library. Algorithms are tested on fl ex-
ible radar equipment. We use digital trans-
ceiver systems - such as Universal Software
Radio Peripheral (USRP).
In recent years, research on cognitive
radar design has been conducted covering a
wide range of programs, using many diff er-
ent methods based on previous advances in
Bayesian theory, information theory, theoret-
ical solutions, approaches, including fuzzy
logic, rule-based systems, metaheuristic al-
gorithms and Markov solutions. processes,
dynamic programming, optimization and
game theory.
Future systems are learning the abil-
ity to predict the behavior of radars in the
operational environment and to adapt its
transmission in the available spectrum. Ra-
dar cognition in this case is based on two
main concepts: spectrum probing and spec-
trum distribution. The sounding spectrum is
aimed at recognizing the frequency used by
other systems and occupying the same spec-
trum in real time.
System performance is measured in
terms of standard metrics such as target de-
tection probability and false alarms, root
mean square error in tracking systems, and
classifi cation accuracy in automated target
recognition systems - cognitive systems re-
quire additional metrics that quantify perfor-
mance gains and achievement use of system
resources. Two key issues for cognitive radar
research are the development of assessment
and assessment tools, as well as experimental
testing of the methodology.
A related but unique problem with
the cognitive design of radar is experimental
testing, as the shape of the transmitted signal
and the settings are adapted during opera-
tion. With more sophisticated modeling, new
development and qualifi cation processes can
be developed, including software testing that
will help test cognitive radars.
Cognitive radars are evaluated
through simulations, or using pre-recorded
data. The infrastructure of testing, calibra-
tion and debugging tools is being developed
in parallel. For example, SPC Quantor has
developed real-time tests for cognitive ra-
dars. Reliability of modeling and compu-
tational errors is an important issue that
should be investigated [33].
Radars differ in their qualitative and
quantitative parameters. A typical approach
is to determine the number of worst cases
and make them work in the worst cases. This
is true for non-cognitive radar systems that
do not change the configuration depending
on the current environment, because a sin-
gle radar configuration is used. Cognitive
radars change their configuration, rebuild-
ing the neural network and learning or self-
learning to solve various problems within
certain limits. Moreover, with the develop-
ment of neural network design tools, cogni-
tive radars will have better characteristics
and lower design costs and the ability to
self-improvement.
The power of the transmitter should
not exceed the limits imposed by regulatory
requirements, because there is unwanted ra-
diation due to the nonlinearity of the transmit-
ter and a sharp increase and decrease in radar
pulses [34]. Especially in cognitive systems,
dynamic reconfi guration of the transmission
spectrum is not always easy to implement and
can lead to out-of-band transmissions, which
cause a slight spectral expansion outside the
designated radar band.
Cognitive radar architecture
The cognitive radar architecture is
built through the extension of the perception-
action cycle by introducing an evaluation
process that forms a perception-evaluation-
action cycle (PEAC). The purpose of this ad-
ditional step is to emphasize the assessment
of the currently perceived situation supported
by artifi cial intelligence. It is done regarding
the purpose of the sensor and depends on the
purpose certain perception results will lead
79
Моделі та методи машинного навчання
to very diff erent actions, such as observation,
where the overall picture is important. When
observed, all identifi ed targets will receive an
equal share of available resources.
Fig.2. Functional diagram
of the cognitive radar.
The radar sensor, which operates in
PEAC, consists of a programmable sensor that
provides data processing and data evaluation,
and management of resources generated de-
pending on recently received environmental
information. In future systems, much of the
intelligence will be located in a multiplatform
cloud. Cognitive radar responds intelligently
to real-time scenario variations.
A software-defi ned sensor is a system
that performs radar measurements, ie emits,
receives and processes electromagnetic sig-
nals in accordance with the requirements of
the sensor in order to obtain new informa-
tion in its environment. A software-defi ned
sensor requires hardware capabilities (eg, in-
stantaneous bandwidth, operating bandwidth,
waveform, polarization, etc.) to meet resource
management requirements.
Radio compatibility is achieved by
methods relating to the emission of radar sig-
nal or signal processing. Methods of reduc-
ing interference from other radio frequency
systems are achieved by dividing in time, fre-
quency, space or signal modulation. Consider
the main measures that allow coexistence:
waveform, illumination in radar images,
adaptive zeroing of interference, frequency
adaptation, dynamic adaptation of the search
circuit and the level of radiation power, de-
tection of interfering samples in the receiving
signal, suppression of interfered samples in
the radar signal [35].
A Doppler shift around a target ap-
pears if the target contains moving, vibrat-
ing, or rotating parts and can be observed
externally. For example, the wings of birds,
the wheels of cars, the arms and legs of peo-
ple walking, as well as the rotors of helicop-
ters and tank tracks have a unique Doppler
spectrum, which is observed using a radar
system with a fairly high Doppler resolu-
tion. The spectrum of Doppler shift strongly
depends on such parameters as the angle of
illumination, the absolute velocity of the
target and the composition of the underly-
ing surface.
To adjust the detection thresholds ac-
cording to the actual background, it is nec-
essary to determine and assess the level of
interference. This assessment can be per-
formed on a one-time basis or over a long
period of time by studying the characteristics
of the obstacles. The mapping features can
be characterized by its spatial composition,
amplitude statistics and Doppler spectrum.
This allows you to reliably adjust the detec-
tion thresholds according to the character-
istics of the interference, while maintaining
a low level of false positives and providing
the ability to detect targets against the back-
ground of interference.
Multi-beam radiation on the sea surface
is characteristic of marine radars. The imposi-
tion of a direct and refl ected path on the sea
surface leads to the appearance of zones with
attenuation and loss of target detection, as
well as to errors in altitude measurement. By
detecting fading situations, tracking can be
more resistant to detection errors by changing
the transmission frequency and thus avoiding
the fading situation.
Active grating and electronic scanning
antennas, which are the latest in modern radar
systems, provide great fl exibility in the direc-
tion of the beam and waveforms.
The various actions that require re-
sources from the system are called tasks. Each
task is an implementation of the radar func-
tion that the sensor is capable of. Examples
of such tasks: search tasks, tracking tasks,
classifi cation tasks, visualization tasks, envi-
ronmental detection tasks, externally assigned
search and tracking tasks ordered by a higher
system.
QoS resource management techniques
use quality measures to optimize overall sys-
80
Моделі та методи машинного навчання
tem performance for this metric and avail-
able resources. Therefore, this approach
requires a good understanding of the qual-
ity measures used. Especially when a large
number of diff erent tasks are used, which
will certainly be the case in future cognitive
radars, it is necessary to fi nd a strong bal-
ance between individual tasks and mission
needs. The advantage of the QoS approach is
that even in dense scenarios, the available re-
source is distributed among the tasks still to
maximize system performance, and no pre-
determined priority, which may or may not
be applied in the current situation, should be
used to resolve confl icts. However, adaptive
rules as well as the QoS approach require
more environmental information to adapt the
waveform to the evolving situation, taking
into account the various infl uences imple-
mented in the model used to assess the ex-
pected performance of the system. In addi-
tion, for QoS it is necessary to keep a list of
all active tasks of the radar sensor.
The hardware capabilities needed to
take advantage of modern resource manage-
ment capabilities are high if you use the full
potential of algorithms. In this case, you need
a fully fl exible interface that allows you to
confi gure all available parameters (such as
waveform, parameters, and viewing direc-
tions) within the physical boundary of the
external interface. However, to speed up the
process of optimizing resource management,
the available degrees of freedom can be lim-
ited in advance, for example, by limiting the
repetition rates of the selected pulses. If the
limit is chosen adequately, the decrease in
productivity is insignifi cant.
Deep learning of the cognitive
radar neural network
Probably, today only the lazy are not
engaged in machine learning. But when we
look at cognitive radar software, we are talk-
ing about deep learning. This is when the
feedback covers the entire radar.
The backpropagation algorithm is an
extension of the perception of multilayer
neural networks. Thus, the backpropagation
algorithm uses three or more levels of pro-
cessing units (neurons). In a typical 3-tier
network architecture for a backpropagation
algorithm, the leftmost layer of ones is the
input layer that receives the input data. Lat-
er, this is a hidden layer, where processing
blocks are linked to the layers before and af-
ter it. The rightmost layer is the output layer.
The levels are fully interconnected, which
means that each processing unit is connected
to each unit at the previous level and at the
next level. However, the units are not linked
to other units in the same layer. Backpropa-
gation networks are not fully interconnected,
which means that any number of hidden lay-
ers can be used. [31].
Traditional event detection in cogni-
tive imaging radar is based on batch or offl ine
algorithms: it is assumed that there is one
event in each radar information stream. The
stream is usually processed using a prepro-
cessing algorithm that requires a huge amount
of computation. Neural networks can eas-
ily cope with such tasks with the appropriate
deep learning. This is an analogue of infor-
mation processing tasks “on the fl y” as they
become available.
Neural networks are also an eff ective
method for diagnosing faults based on non-
linear mapping of input and output data, par-
allel processing and a high degree of self-or-
ganization and self-learning ability [36]. In
the structure of closed-loop neural networks
the only suitable connections are between
the outputs of each level and the input of the
next level [37]. A backpropagation neural
network is one known method for creating a
trained machine or system that can provide a
fi nal classifi cation decision through a series
of learning processes. It can be developed
using the tools provided in MATLAB, but
sometimes this leads to diff erent detection
and recognition accuracy of objects for each
experiment [30–31].
We achieve troubleshooting by rebuild-
ing the computational resource of the neural
network. Moreover, a quick response occurs
by changing the course of the computing pro-
cess and in case of failures, readjustment of
the network with a change in its resource. It
is possible to draw an analogy with a living
intellect, where homeostasis is provided at
the hormonal level and a quick response to
changes in the external environment by ner-
vous signaling.
81
Моделі та методи машинного навчання
We always willingly or unwillingly
use bionic models. Now it has resulted in a
separate science of imitation of nature - bio-
mimetics. Creating a model in biomimetics
is half the battle. To solve a specific prac-
tical problem, it is necessary not only to
check the presence of the model properties
of interest to practice, but also to develop
methods for calculating the predetermined
technical characteristics of the device, to
develop synthesis methods that ensure the
achievement of the indicators required in
the problem.
And therefore, many bionic models,
before they receive technical implementa-
tion, begin their life on a computer. A math-
ematical description of the model is con-
structed. Based on it, a computer program
is compiled - a bionic model. On such a
computer model, various parameters can be
processed in a short time and design flaws
can be eliminated.
Traditionally, deep learning algorithms
update the weight of the network, while the
architecture of the network is selected manu-
ally using the trial-and-error method. This
study proposes two new approaches that auto-
matically update the structure of the network,
as well as studying its weight. The novelty
of this approach is parameterization, where
depth or additional complexity is constantly
encapsulated in the space of parameters that
give additional complexity.
Deep learning includes several levels
of nonlinear information processing. This al-
lows us to study architectures that implement
functions through repetitive compositions of
simpler functions, thereby exploring levels of
abstraction with the best generalization and
representation.
Although in-depth training is useful,
keeping multiple layers can be problematic.
First, when more layers, weight, space, and
computational complexity are higher; second,
when there are more free parameters, there is
a higher risk of retraining; third, if the net-
work is deep, there is the problem of disap-
pearing gradients when the error spreads over
many layers.
There have been many approaches to
optimizing the network architecture - from
early incremental methods of bringing hid-
den modules one after the other (or start-
ing from a large network and reducing it) to
more sophisticated modern approaches such
as evolutionary algorithms or reinforced
learning and stimulus style techniques. The
purpose of the study is to study network ar-
chitecture based on data. The main differ-
ence is that instead of searching in discrete
space for all architectures that have param-
eterized models in such a way that the very
notion of complexity or depth is itself con-
tinuous, making the model differentiated
from beginning to end.
Two methods are proposed for con-
structing and studying the structure of a
deep neural network, where the complexity
of the network at the level of a hidden block
or layer is encoded by continuous parame-
ters. These parameters are adjusted together
with the network weights during the gradi-
ent descent, which implies a slight change
in the structure of the network together with
the network weights. The first method in
tunnel networks associated with each hid-
den block is a continuous parameter. If this
parameter is not active, the block simply
copies its input to its output to bypass non-
linearity, effectively increasing the depth of
the network. In the second method, the per-
ceptron has parameters associated with each
layer, indicating whether further nonlinear
processing is required. We start with one
layer first, and when training with a gradient
descent, when necessary, this parameter can
become active, which causes the creation of
another complete layer, increasing the depth
of the network.
Experiments on synthetic double-
helix data like tunnel networks and novice
perceptrons can be adapted to different sizes
for different complexity of problems using
the same set of hyperparameters, adapting
the number of units for the tunnel networks
and the number of layers for the initial per-
ceptrons. With regard to real problems of
recognition of numbers and images, we ob-
serve that tunnel networks achieve better
performance, providing a better regularized
model and using fewer parameters com-
pared to backbone networks. Also, novice
perceptrons showed comparable or better
performance. Compared to tunnel networks,
82
Моделі та методи машинного навчання
novice perceptrons appear to grow larger
and shrink less. By setting the learning rate
in descending order, it is observed that dif-
ferent layers grow at different rates and are
used in different ways. Combined with reg-
ularization, this allows tunnel networks to
keep some of the unused upper layers lin-
ear, thus effectively removing them from
the network at the end.
Deep learning is an AI function that
mimics the workings of the human brain in
such a way that it processes data and cre-
ates patterns for use in decision making
[35]. Cognition is a fundamental feature of
natural intelligence. Sensory cognitive net-
works provide new technological support to
dramatically increase the quantity and qual-
ity of information that can be collected and
transmitted in complex adaptive systems.
Their application can significantly increase
the level of intelligence in the design and
implementation of the system to the levels
at which the effects of cognition will begin
to manifest themselves. Cognitive abilities
can be thought of as a shared sensory net-
work. The detection system learns to detect
changes not only in signal levels, but also in
the shape and parameters of the sensor sig-
nal, which is a more difficult task. The ar-
chitecture can significantly reduce resource
consumption without sacrificing change
detection performance. Experiments prove
that a neural network-based change detec-
tion system is feasible for developing sensor
network applications and can be success-
fully implemented on available technology
platforms.
Designing cognitive radars has sev-
eral stages. At the first stage we develop
user requirements. The second is the for-
malization of user requirements. Next, we
develop a model of radar operation, check
the receipt of the declared quality charac-
teristics of the radar. In the fourth stage,
we conduct in-depth training of the gener-
ated neural network with an inverse loop,
for which the radar is calibrated. Based on
the calibration results, the development
of the developing cognitive radar system
is adjusted. Consider in more detail the
calibration of 3D-Imaging radar, devel-
oped and manufactured in SPC «Quantor»
on the example of obtaining a 3D image
of the internal structure of the multilayer
material.
The possibility calibration of is
studied in the exploring of material proper-
ties to the example of multilayer structure,
depending on the distance between the
sample and the antenna using an absorber.
The results of preliminary studies indicate
the possibility of measuring the thickness
of the material. On the calibration, a small
metal plate and several measurement cy-
cles for averaging the noise were used. It is
shown that the accuracy of measurements
is influenced by the width of the radiation
pattern, the number of measurement cycles
at one point, the accuracy of positioning
and moving the head during the measure-
ments, and the time interval between the
calibrations.
We have developed algorithms and
have obtained the required accuracy. We will
try to test the radar system, having previously
calibrated it.
Before carrying out the measurements,
we set:
1. The horn and the sample close the
absorber to reduce the refl ections;
2. We measure the signal without a
sample;
3. We place the sample (5-10) mm and
begin to measure;
4. Very carefully, a thin conductive
fi lm is pasted from above and measurements
are taken;
5. Very gently fl ip back with a conduct-
ing medium and measure.
Fig.3. Implementation of 3D scanning
of small objects by cognitive FMCW
3D-Imaging Radar terahertz range.
83
Моделі та методи машинного навчання
Fig.4. Functional diagram of 3D Imaging
FMCW Terahertz Radar.
If all done carefully, there should be a
shift of even a fraction of a millimeter. Scan
must be disabled. We will make a point of 50
measurements.
Fig.5. Propagation of radar radiation
in a multilayer material.
1. Diff erent materials have a diff erent
permittivity and diff erent velocity of phase
of electromagnetic wave. This gives that real
thickness between upper and lower plane of
samples is equivalent to virtual thickness be-
tween real upper and virtual lower plane in
air. We can estimate equivalent virtual thick-
ness between metal planes in air and then cor-
rect result for real material.
2. In real materials we have a multiple
refl ection. This gives several spectral lines for
one thickness of the sample.
3. Refl ections from virtual metal
planes do not fully correspond to refl ections
from real metal planes during calibration.
There are numerous errors of discrepancy
between virtual planes and the nearest cali-
bration levels. This gives multiple errors in
the spectral lines and creates some diffi cul-
ties in estimating the thickness.
4. Some calibration levels must be pre-
sented lower than baseline to estimate posi-
tions of virtual metal planes.
5. We try to fi x existing mathematic
problems and to get a mathematical tool for
universal measurement device.
6. We check the additional measure-
ment confi guration. For calibration, we use
a special sample with a higher accuracy –
Plexiglas.
As a result of the measurement cycle,
a frequency dependence of the attenuation in
the microwave channel D (f) =Uref (f)/Uinc (f)
is obtained.
Unknown parameters of the dielec-
tric structure are determined by procedure of
global minimization of discrepancy between
the measured attenuation in channel D(f) and
one calculated theoretically Dth(f, p)
f
th fDfDF 2,pp .
Here Dth(f, p) is defi ned according to
the formula
2
233
10 )1)(1( cc
c
th VVkVkVk
VVkkD
,
and 0 3( ),.., ( )k f k f are complex coeffi cients,
which are determined experimentally using
reference samples and describe properties of
the microwave channel; f is the frequency of
sounding waves; VC (f) is the complex refl ec-
tion coeffi cient (CRC) of the reference arm
3; V (f,p) is a theoretically calculated CRC
of the dielectric structure, which depends on
a vector of the structure parameters p (thick-
ness of layers and electrical parameters of
materials).
We consider that in free space extends
a plane electromagnetic wave and normally
incident on the infi nite (M-1)-layer medium
with fl at boundaries. The CRC V (f, p) is relat-
ed of the CRC of the structure in free space VS
(f, p) through the scattering matrix of the an-
tenna S, which is determined experimentally:
S
S
VS
VS
SV
22
21
11 1
.
84
Моделі та методи машинного навчання
The CRC of the structure in free space
depends on the thickness and electrophysical
parameters of structure layers:
VS=VS(f,h1,..,hM-1,ε1,..,εM,tgδ1,...,tgδM),
where hm, εm, tgδm is thickness, permittivity
and loss tangent of m-th layer. The CRC of
the plane wave from dielectric plane-lay-
ered medium VS(f,p) is determined by the
known formulas:
10
10
YW
YW
VS
,
,
)1(exp)1(exp
)1(exp)1(exp
1
1
mmmm
mmmm
mm qYqW
qWqYWY
,MM WY
0/mmW ;
)sin1(22 2
0mmm hfjq ;
)tg1(0 mmm j ,
where ε0, is permittivity and μ0 is the perme-
ability of free space.
During the setup process, we do not
need to change the distance between the sig-
nal and the base line, but we need to will
move the device and perform a calibration at
the center of each step. This calibration pro-
cess is simpler and can be performed in auto-
matic mode without an additional table with a
micrometer.
Сonclusions
This article provides a summary of
the development of modern radar systems.
It is shown that with the development of
Artificial Intelligence technologies, mod-
ern radars use deep learning neural net-
works, as a result, radars have become
cognitive. There is no alternative to satisfy
the consumer in terms of quality indicators
using old technologies. Scientific, techno-
logical, information base is ready for such
challenges. The need for modern radars is
also huge: medicine, security, defense, the
Internet of things, and others. Scenarios
are becoming more complex and require
creative solutions. Cognitive radar is one
potential solution that has long been dis-
cussed in the literature.
It has been shown that cognitive ra-
dars can coexist in a congested spectrum,
including with random and intentional in-
terference, and be invisible. Cognitive radar
systems can adapt to a changing environment
using internal and external sources of infor-
mation. It is possible to control the resources
of the radar, and therefore, the radars are in-
herently fault-tolerant.
References
1. S. Haykin, “Adaptive radar: Evolution to
cognitive radar,” in Proc. IEEE Int. Symp.
Phased Array Syst. Technol., Boston, MA,
USA, Oct. 2003, p. 613.
2. S. Haykin, Y. Xue, and P. Setoodeh, “Cogni-
tive radar: Step toward bridging the gap be-
tween neuroscience and engineering,” Proc.
IEEE, vol. 100, no. 11, pp. 3102–3130, Nov.
2012.
3. J. Rasmussen, “Skills, rules, and knowledge;
signals, signs, and symbols, and other dis-
tinctions in human performance models,”
IEEE Trans. Syst., Man, Cybern., vol. SMC-
13, no. 3, pp. 257–266, May 1983.
4. J. Rasmussen, Information Processing and
Human Machine Interaction: An Approach
to Cognitive Engineering. New York, NY,
USA: Elsevier Science, 1983.
5. S. Bruggenwirth, “Design and implementa-
tion of a three-layer cognitive radar architec-
ture,” in Proc. 50th Asilomar Conf. Signals,
Syst. Comput., Nov. 2016, pp. 929–933.
6. A. Balleri, H. Griffi ths, and C. Baker, Bio-
logically Inspired Radar and Sonar: Lessons
from Nature. IET, 2017.
7. S. Middelhoek, J. B. Angell, and D. J. W.
Noorlag, “Microprocessors get integrated
sensors: Sensing devices and signal process-
ing built into one silicon chip portend a new
class of smart sensors,” IEEE Spectr., vol.
17, no. 2, pp. 42–46, Feb. 1980.
8. L. Reznik, G. Von Pless, and T. Al Karim,
“Distributed neural networks for signal
change detection: On the way to cognition in
sensor networks,” IEEE Sensors J., vol. 11,
no. 3, pp. 791–798, Mar. 2011.
9. R. V. Kulkarni, A. Forster, and G. K. Venay-
agamoorthy, “Computational intelligence in
wireless sensor networks: A survey,” IEEE
Commun. Surv. Tut., vol. 13, no. 1, pp. 68–
96, Jan.–Mar. 2011.
85
Моделі та методи машинного навчання
10. W. D. Blair, G. A. Watson, and S. A. Hoff -
man, “Benchmark problem for beam point-
ing control of phased array radar against ma-
neuvering targets,” in Proc. Amer. Control
Conf., vol. 2, pp. 2071–2075, Jun. 1994.
11. W. D. Blair, G. A. Watson, T. Kirubarajan,
and Y. Bar-Shalom, “Benchmark for ra-
dar allocation and tracking in ECM,” IEEE
Trans. Aerosp. Electron. Syst., vol. 34, no. 4,
pp. 1097–1114, Oct. 1998.
12. G. van Keuk and S. S. Blackman, “On
phased-array radar tracking and parameter
control,” IEEE Trans. Aerosp. Electron.
Syst., vol. 29, no. 1, pp. 186–194, Jan. 1993.
13. S. S. Blackman, R. J. Dempster, M. T.
Busch, and R. F. Popoli, “IMM/MHT solu-
tion to radar benchmark tracking problem,”
IEEE Trans. Aerosp. Electron. Syst., vol. 35,
no. 2, pp. 730–738, Apr. 1999.
14. T. Kirubarajan, Y. Bar-Shalom, W. D. Blair,
and G. A. Watson, “IMMPDAF for radar
management and tracking benchmark with
ECM,” IEEE Trans. Aerosp. Electron. Syst.,
vol. 34, no. 4, pp. 1115–1134, Oct. 1998.
15. J. Ward, Space-Time Adaptive Processing
for Airborne Radar. Cambridge, MA: MIT
Lincoln Laboratory, 1994.
16. W. K. Staff ord, “Real time control of a mul-
tifunction electronically scanned adaptive
radar (mesar),” in Proc. IEE Colloq. Real-
Time Manage. Adaptive Radar Syst., Jun.
1990, pp. 7/1–7/5.
17. S. Haykin, “Radar vision,” in Proc. IEEE
Int. Conf. Radar, May 1990, pp. 585–588.
18. S. J. Anderson, “Adaptive remote sensing with
HF skywave radar,” IEE Proc. F—Radar Signal
Process., vol. 139, no. 2, pp. 182–192, Apr. 1992.
19. D. J. Kershaw and R. J. Evans, “Optimal
waveform selection for tracking systems,”
IEEE Trans. Inf. Theory, vol. 40, no. 5, pp.
1536–1550, Sep. 1994.
20. D. J. Kershaw and R. J. Evans, “Waveform
selective probabilistic data association,”
IEEE Trans. Aerosp. Elect. Syst., vol. 33,
no. 4, pp. 1180–1188, Oct. 1997.
21. B. F. La Scala, W. Moran, and R. J. Evans,
“Optimal adaptive waveform selection for tar-
get detection,” in Proc. IEEE. Int. Conf. Radar
(Cat. 03EX695), Sep. 2003, pp. 492–496.
22. J. Fuster, Cortex and Mind: Unifying Cog-
nition. New York, NY, USA: Oxford Univ.
Press, 2010
23. P. Stinco, M. S. Greco, and F. Gini, “Spec-
trum sensing and sharing for cognitive ra-
dars,” IET Radar, Sonar Navigation, vol. 10,
no. 3, pp. 595–602, 2016.
24. A. Aubry, V. Carotenuto, A. D. Maio,
and S. Iommelli, “Cognitive radar wave-
form design for spectral compatibility,”
in Proc. Sensor Signal Process. Defense,
2016, pp. 1–5.
25. N. A. Goodman, “Closed-loop radar with
adaptively matched waveforms,” in Proc.
Int. Conf. Electromagn. Adv. Appl., 2007,
pp. 468–471.
26. K. L. Bell, C. J. Baker, G. E. Smith, J. T.
Johnson, and M. Rangaswamy, “Fully adap-
tive radar for target tracking part I: Single
target tracking,” in Proc. IEEE Radar Conf.,
2014, pp. 0303–0308.
27. J. M. Christiansen, K. E. Olsen, and G. E.
Smith, “Fully adaptive radar for track update
control,” in Proc. IEEE Radar Conf., 2018,
pp. 0400–0404.
28. L. Ubeda-Medina and J. Grajal, “Multiple
target tracking in the fully adaptive radar
framework,” in Proc. IEEE Statist. Signal
Process. Workshop, 2016, pp. 1–5.
29. R. Romero and N. Goodman, “Cognitive
radar network: Cooperative adaptive beams
teering for integrated search and track ap-
plication,” IEEE Trans. Aerosp. Electron.
Syst., vol. 49, no. 2, pp. 915–931, Apr. 2013.
30. R. Oechslin, P. Wellig, S. Hinrichsen, S.
Wieland, U. Aulenbacher, and K. Rech,
“Cognitive radar parameter optimization in
a congested spectrum environment,” in Proc.
IEEE Radar Conf., 2018, pp. 0218–0223.
31. F. Smits, A. Huizing, W. van Rossum, and
P. Hiemstra, “A cognitive radar network: Ar-
chitecture and application to multiplatform
radar management,” in Proc. 5th Eur. Radar
Conf., 2008, pp. 312–315.
32. A. Charlish and F. Hoff mann, “Cognitive
radar management,” in Novel Radar Tech-
niques and Applications: Waveform Diversi-
ty and Cognitive Radar and Target Tracking
and Data Fusion 2, London U.K.: Institution
of Engineering and Technology, 2017.
33. L. Ubeda-Medina, A F. Garca-Fernndez, and
J. Grajal, “Robust sensor parameter selection
in fully adaptive radar using a sigma-point
gaussian approximation,” in Proc. IEEE Ra-
dar Conf., Apr. 2018, pp. 0263–0268.
86
Моделі та методи машинного навчання
34. R. Turyn, “On barker codes of even length,”
Proc. IEEE, vol. 51, no. 1, pp. 230–231, Sep.
1963. [34] D. DeLong and E. Hofstetter, “On
the design of optimum radar waveforms for
clutter rejection,” IEEE Trans. Inf. Theory,
vol. 13, no. 3, pp. 454–463, Jul. 1967.
35. K. L. Bell, C. J. Baker, G. E. Smith, J. T.
Johnson, and M. Rangaswamy, “Fully adap-
tive radar for target tracking part I: Single
target tracking,” in Proc. IEEE Radar Conf.,
2014, pp. 0303–0308.
36. J. M. Christiansen, K. E. Olsen, and G. E.
Smith, “Fully adaptive radar for track update
control,” in Proc. IEEE Radar Conf., 2018,
pp. 0400–0404.
37. L. Ubeda-Medina and J. Grajal, “Multiple
target tracking in the fully adaptive radar
framework,” in Proc. IEEE Statist. Signal
Process. Workshop, 2016, pp. 1–5.
38. R. Rovol. 49, no. 2, pp. 915–931, Apr. 2013.
Received: 05.02.2022
About the authors:
Mykola Коsovets
Leading Constructor,
Number of scientifi c publications
in Ukrainian publications -56
Number of scientifi c publications
in foreign publications -17
Index Girsh - 5
https://orcid.org/0000-0001- 8443-7805
Scopus Author ID: 5644007500
Lilia Tovstenko
Leading Software Engineer
Number of scientifi c publications
in Ukrainian publications -24
Number of scientifi c publications
in foreign lands -8
Index Girsh -7
https://orcid.org/0000-0002-3348-6065
Scopus Author ID: 56439972800
Place of work:
Mykola Коsovets
SPE “Quantor”, Chief
03057, c. Kyiv-57, str. Е. Potye, 8-А
Ph.: (380)66-2554143
E-mail: quantor.nik@gmail.com
Lilia Tovstenko
Institute of Cybernetics of Glouchkov
National Academy of the National Academy
Sciences Ukraine
03187, Kyiv-187,
Academician Glouchkov Avenue, 40.
Ph.: (380)67-7774010
E-mail: 115lili@incyb.kiev.ua
|