Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку
Problem of signal processing and network traffic optimization is solved at the hardware level and is interesting, modern and relevant from the point of view of the application level. It is necessary to propose an approach that combines machine learning methods with network bandwidth tasks and traffi...
Збережено в:
| Дата: | 2026 |
|---|---|
| Автори: | , , , , , |
| Формат: | Стаття |
| Мова: | Англійська |
| Опубліковано: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2026
|
| Теми: | |
| Онлайн доступ: | https://journal.iasa.kpi.ua/article/view/358086 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | System research and information technologies |
| Завантажити файл: |  |
Репозитарії
System research and information technologies| _version_ | 1862949225453584384 |
|---|---|
| author | Zolotukhin, Oleg Kudryavtseva, Maryna Bodyanskiy, Yevgeniy Filatov, Valentin Antilikatorov, Andriy Kalinin, Dеnys |
| author_facet | Zolotukhin, Oleg Kudryavtseva, Maryna Bodyanskiy, Yevgeniy Filatov, Valentin Antilikatorov, Andriy Kalinin, Dеnys |
| author_sort | Zolotukhin, Oleg |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2026-04-19T21:53:19Z |
| description | Problem of signal processing and network traffic optimization is solved at the hardware level and is interesting, modern and relevant from the point of view of the application level. It is necessary to propose an approach that combines machine learning methods with network bandwidth tasks and traffic over the network. To solve this problem, it is proposed to use concept of Informed Machine Learning (IML), that is the Taxonomy of IML, the principles of constructing deep machine learning systems based on information about the physical properties of the data transmission network under study. The platform for developing is deep machine learning models using PINN neural networks. The PINN represents the class of deep learning algorithms that can integrate data with or without physical processes description. As an algorithm it is proposed to use the popular algorithm in the field of deep learning – ADAM (Adaptive Moment Estimation) for optimizing network traffic. Using the PINN trained with the ADAM algorithm to transmit data the efficiency has increased. Thanks to this method, it is possible to obtain a low noise signal in practice, due to which network traffic is optimized. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2026.1.11 |
| first_indexed | 2026-04-20T01:00:29Z |
| format | Article |
| fulltext |
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin, 2026
Системні дослідження та інформаційні технології, 2026, № 1 155
UDC 004.8:519.816
DOI: 10.20535/SRIT.2308-8893.2026.1.11
PHYSICAL-INFORMED NEURAL NETWORK
IN SIGNAL PROCESSING
AND NETWORK TRAFFIC COMMUNICATIONS
O.V. ZOLOTUKHIN, M.S. KUDRYAVTSEVA, Y.V. BODYANSKIY,
V.O. FILATOV, A.V. ANTILIKATOROV, D.V. KALININ
Abstract. Problem of signal processing and network traffic optimization is solved at
the hardware level and is interesting, modern and relevant from the point of view of
the application level. It is necessary to propose an approach that combines machine
learning methods with network bandwidth tasks and traffic over the network. To solve
this problem, it is proposed to use concept of Informed Machine Learning (IML), that
is the Taxonomy of IML, the principles of constructing deep machine learning
systems based on information about the physical properties of the data transmission
network under study. The platform for developing is deep machine learning models
using PINN neural networks. The PINN represents the class of deep learning
algorithms that can integrate data with or without physical processes description. As
an algorithm it is proposed to use the popular algorithm in the field of deep learning
– ADAM (Adaptive Moment Estimation) for optimizing network traffic. Using the
PINN trained with the ADAM algorithm to transmit data the efficiency has increased.
Thanks to this method, it is possible to obtain a low noise signal in practice, due to
which network traffic is optimized.
Keywords: Adaptive Moment Estimation, artificial intelligence, artificial neural
network, Data Visualization, dynamic neuron, gradient decent, Informed Machine
Learning, learning algorithm, network bandwidth utilization, network traffic
optimization, neuro-fuzzy logic, Python.
INTRODUCTION
The main task of optimizing information networks is to maintain the required
level of performance characteristics and network capacity under conditions of load
changes.
To assess the quality of any network, characteristics such as data transfer
speed, reliability of transmitted information, bandwidth and reliability of the
communication channel are used.
The main characteristic of information transmission channels is their
bandwidth, that is the amount of data that is going threw a network with a given
speed depending on the channel capacity.
Currently, the load on the data transmission channel can vary by an order of
magnitude depending on the time of day, and the nature of the transmitted data is
determined by the user’s field of activity (for example, large videos, project files,
high quality images), therefore, optimization problems must be solved in real time.
First of all, it is necessary to formulate criteria for the effectiveness of the
information network. Most often, these criteria are performance and reliability,
which in turn require the selection of specific evaluation indicators. For example,
the performance of an information network is determined by the response time (the
time interval between the occurrence of a user request for any network service and
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin
ISSN 1681–6048 System Research & Information Technologies, 2026, № 1 156
the receipt of a response to this request), and the reliability of the network is
determined by the readiness ratio of the network equipment to perform it’s
functions at an arbitrary point in time.
In addition, it is necessary to determine many variable network parameters
that directly or indirectly affect the performance criteria. Settings can be devices,
protocols, or transmission technologies.
Thus, to optimize data transmission over a network, it is advisable to consider
three successive stages of network management.
1. Bringing the network into working condition, which usually includes:
searching for faulty network elements at the physical or data link levels;
checking the compatibility of equipment and software (at the network level);
selection of correct values for key parameters of programs and devices that
ensure data transfer between network nodes — addresses of networks and nodes,
protocols used, types of frames, packets (at the transport level).
In this case, optimization comes down to diagnosing faults and bringing the
network into an operational state.
2. Primary configuration — selection of parameters that significantly affect
the characteristics of the network. If the network is efficient, but communication is
very slow due to unacceptable latency or communication sessions are frequently
interrupted, then you need to look for the key factors that degrade the network.
Typically, the reasons for a noticeable decrease in network performance or unstable
network operation are found in an incorrectly operating element or an incorrectly
set parameter, but due to the large number of such parameters, solving this problem
may require long-term monitoring of network operation, collecting statistics, and
searching through options. At this stage, a certain threshold value of the efficiency
indicator is also set and it is required to find a network variant for which this value
would be no worse than the threshold.
3. The final configuration of network parameters is directly optimizing the
network operation. In the case of a normally operating network, further improving
it’s quality, as a rule, requires finding some optimal combination of values for a
large number of parameters.
During the final network setup, in which the parameters of the operating
network, for example, the frame size or the size of the window of unacknowledged
packets, can be varied in order to improve performance (for example, the average
response time) by at least a few percent.
As a rule, network optimization is understood as some kind of compromise
approach. It is necessary to select values of network parameters such that it’s
efficiency indicators are at least not lower than the maximum permissible values
specified when choosing the global Quality of Service level.
In real conditions, it is enough to find a solution close to the optimal one, i.e.
it is necessary to find some rational version of the structure and parameters of the
network.
MATERIALS AND METHODS
The business of geographically distributed companies with an extensive branch
network largely depends on the speed and stability of information exchange
between departments. The WANs (Wide Area Networks) used for this usually do
not always meet the requirements for the productive operation of such critical
Physical-informed neural network in signal processing and network traffic communications
Системні дослідження та інформаційні технології, 2026, № 1 157
applications as VoIP, video conferencing, Enterprise Resource Planning systems
and other software.
WAN optimization solutions help companies increase data transfer speeds
without expanding the channel, and network administrators can improve
application performance, avoid data transfer delays and packet losses, and ensure
guaranteed service quality.
At present, when companies are actively centralizing IT infrastructures to
improve security and reduce management costs, and end users are increasingly
working in mobile office mode, the most pressing task for network administrators
is increasing data transfer speeds.
The modern network architecture is presented on Fig. 1. The distributed
network architecture is presented in the form of a chain, which consists of: the
user’s local network, the user’s provider, the global Internet, the recipient’s
provider and the recipient’s local network.
Based on this complex hierarchical architecture, the problem of packet data
transmission becomes clear — the amount of data transmitted exceeds the amount
of data received due to network noise.
Fig. 1. The distributed network architecture
So, today problem of signal processing and network traffic optimization is
solved at the hardware (or protocol level) and is interesting, modern and relevant
from the point of view of the application level. Representation of the protocol and
user data transfer levels you can see on the Fig. 2.
It іs necessary to propose an approach that combines machine learning
methods, for example, neural networks with network bandwidth tasks and traffic
over the network.
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin
ISSN 1681–6048 System Research & Information Technologies, 2026, № 1 158
This will allow us to present not only a theoretical, but also a practical solution
to the problem.
Fig. 2. Representation of the protocol and user data transfer levels
One way to optimize of data traffic at the hardware level is solved by
configuring protocol parameters for packet transmission: packet size and number
of packets. The packet transmission rate is determined by the speed of light and
depends on the number of nodes from point A to point B of network.
For example, we established a connection with a remote computer, sent 16
packets, received confirmation that 12 packets were received, and in some packets
the checksum does not match. We transmit the following packets, to them we re-
add packets in which the correct checksum was not received. This is already an
optimization compared to the fact that it is possible to transmit data in one packet
and wait for confirmation each time, and in case of incorrect transmission of the
packet, resend it.
To solve this problem in [1, 2] works it is proposed to use concept of Informed
Machine Learning (IML). IML describes learning from a hybrid information source
that consists of data and prior knowledge. The prior knowledge comes from an
independent source, is given by formal representations, and is explicitly integrated
into the machine learning pipeline.
Physics-informed machine learning integrates seamlessly data and
mathematical physics models, even in partially understood, uncertain and high-
dimensional contexts [1].
Taxonomy of Informed Machine Learning is represented on Fig. 3. This
taxonomy serves as a classification framework for informed machine learning and
structures approaches according to the three above analysis questions about the
knowledge source, knowledge representation and knowledge integration. Authors
identified for each dimension a set of elements that represent a spectrum of different
approaches [2].
This paper proposes to use the Taxonomy of Informed Machine Learning, the
principles of constructing deep machine learning systems based on information
about the physical properties of the information data transmission network under
study. The platform for developing intelligent tools is deep machine learning
models using PINN neural networks.
Physical-informed neural network in signal processing and network traffic communications
Системні дослідження та інформаційні технології, 2026, № 1 159
Fig. 3. Taxonomy of Informed Machine Learning
ARCHITECTURE OF PHYSICS-INFORMED NEURAL NETWORK
Thus, it is necessary to explore the gray box model in state space. A gray box
represents a model that learns from data, guided by information about the applied
physical properties or laws. Such models can be further used for adaptive control
and self-organization. Peculiarity PINN is to initially take into account the
underlying description of the physical interpretation of partial or ordinary
differential equations, that is, the physics of the problem, rather than trying to derive
a solution based solely on the data, that is, by approximating a set of state-value
pairs with a neural network. Therefore, the use of state space models is considered.
So, the PINNs represent the class of deep learning algorithms that can
seamlessly integrate data and abstract mathematical operators, including partial
differential equations (PDE) with or without missing physics.
According to [1, 2] to design the PINN network for signal processing and
network traffing optimisation it is necessary to implement the chain: Scientific
Knowledge-Algebraic or Differential Equations-Learning algorithms (Fig. 4).
Fig. 4. Stages of the approach based on Physics-Informed Neural Networks
Let’s explore this method from the point of view of using a simple neural
network architecture.
A regular Neural Network approximates various functions well (Fig. 5).
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin
ISSN 1681–6048 System Research & Information Technologies, 2026, № 1 160
Fig. 5. A regular Neural Network
According to the Tsybenko theorem [3], an artificial neural network of
feed-forward with one hidden layer can approximate any continious functionof
many variables using the input vector and the weights of each vector with any
accuracy, i.e.
𝑏 = 𝑔𝑏 ( 𝜃𝑎𝑏𝑎 ∙ 𝑎)
Neural networks provide access to the computational graph (Fig. 6), that
is access to all elementary operations on numbers that enter or exit the neural
network [4].
Fig. 6. The computational graph
The graph gives the property of auto differentiation. This allows to find the
derivative of a neural network based on its parameters. This is used to update the
parameters using the backpropagation algorithm (Fig. 7) through the derivative of
the error and update the weights Ꝋ. This is how a neural network is trained.
Fig. 7. The computational graph with the backpropagation algorithm
With the help of the computational graph, it is possible to differentiate the
neural network with respect to intermediate results of the сalculation of any input
Physical-informed neural network in signal processing and network traffic communications
Системні дослідження та інформаційні технології, 2026, № 1 161
parameter a, , or with respect to the input parameters X, for example, with
respect to some physical parameter X [5].
You can write an equation in the form of partial derivatives 𝑃𝐷𝐸 (𝐹, , … )(Fig. 8), which can be used in physics (these can be coordinates or
time of a process or other physical parameters).
Fig. 8. The computational graph with partial derivatives
Let’s consider the learning methods of a neural network. There are three
categories of methods to train a neural network:
observational bias involves supervisor training, submiting a lot of data on
the physical parameters of the system;
inductive bias involves constructing a neural network so that it complies to
physical laws, for example using a convolutional neural network;
learning bias involves to reduce the error through the derivatives of the
parameters.
Let’s consider the Physics-Informed Neural Network architecture for network
traffic optimization.
The PINN for network traffic optimization represents a function of variables:
the time of data transfer, the volume of data transfer in packets, the number of nodes
(Fig. 9).
So, to solve the problem of digital signal processing and network traffic
optimization, you can use dynamic neurons that are described by differential or
difference equations. The behavior of such dynamic neurons is significantly
determined by their prehistory.
Fig. 9. The PINN neural network for network traffic optimization
xt (N, V, τ, Ꝋ)
The 1st
hidden
layer
The
2nd
hidden
layer
φ(xt+1)
z-1(xt+1)= xt
Ut xt+1
xt
z-1 z-1
∑
xt
Result Ut
+
Learning
Algorithm
-
wijwij
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin
ISSN 1681–6048 System Research & Information Technologies, 2026, № 1 162
The input parameters of the neural network are: the time — τ; the amount of
transmitted data in packets – V; the number of nodes — N; the output parameter is
the network bandwidth utilization – U. The time includes: time of protocol
connection and confirmation of two devices, τ con; packet transferring time, τ pass;
processing time, τ proc: τ = 2 τcon + τpass + τproc.
What is network bandwidth utilization? In the simplest terms, network
bandwidth utilization is the rate at which data can flow through the network.
Traditionally measured in Mbits per second (Mbps), higher bandwidth allows more
traffic flow from one device to another. Utilization is the percentage of a network’s
bandwidth that is currently being consumed by network traffic.
Let’s consider neural network. Neural Network has two hidden layers
(determined experimentally). The Input vector of neural network is
xt (N, V, τ, Ꝋ)
where τ = 0,1..n; Ꝋ = {wij}, i,j ϵ [1..k]; t – timestep (discrete moments in time); the
output vector of neural network is Ut.
The neuron of the neural network is a model of a nonlinear dynamic system
in state space, that is, the time factor influences the behavior of the neuron, and the
output signal Ut is determined by the input signals xt+1 and depends on the past
states of the system xt.
In this neural network, a delay element is included in the feedback circuit,
which implements the backward shift operation z-1(xt+1) = xt and provides the neuron
with the necessary dynamic properties over time. This can be seen in the inner
circuit of the neural network. Thereby, the neural network allows to calculate the
data network utilization parameter dynamically, at times t, t +1, and so on.
A dynamic neuron is described by the recurrent equation
𝑥𝑡+1 = 𝜑( 𝑤𝑖𝑗 ∗ 𝑥𝑡 + 𝑄𝑗 )
where t = 0.1..; j = 1,2..n; i = 1,2..n.
We will update the parameters of the neural network with learning algorithm
until the error does not satisfy the result. If the error is satisfactory, then we will
assume that we have found a solution to the system, that is, we have determined the
network bandwidth utilization.
As a training algorithm, consider the classic gradient descent, which is an
optimization algorithm used to minimize errors in a machine learning model.
Despite the fact that the gradient method of training a neural network and its
modified version in the form of backpropagation are the most common algorithms
in machine learning, the task of optimizing neural networks remains a rather
difficult problem today.
The gradient descent method has a number of problems for training deep
neural networks [6–8]:
the gradient method gets stuck in a fairly deep local minimum (Fig. 10).
There are solutions to sometimes work around such problems, such as momentum,
which can carry optimizers through large lifts, or batch normalization, which
smooths out the error space. But the root cause of many branching problems in
neural networks is still a local minimum;
Physical-informed neural network in signal processing and network traffic communications
Системні дослідження та інформаційні технології, 2026, № 1 163
Fig. 10. Using the gradient descent method in the problem of searching for a global
extremum
high time spent on training. Gradient descent, due to its low convergence
rate, is typically a time-consuming method, even with adaptation options for large
data sets such as batch gradient descent.
the gradient method is highly sensitive to optimizer initialization. For
example, performance may be significantly better if the optimizer is initialized near
the second local minimum rather than the first, but this is all arbitrary.
the pace of learning determines the degree of confidence and riskiness of
the optimizer. Setting the learning rate too high may result in missing the global
minimum, while setting the learning rate too low will increase execution time. To
solve this problem, the learning rate is reduced gradually, but choosing the rate of
decrease, taking into account many other variables that determine it, is quite
difficult.
gradient descent requires gradients. This means that it is vulnerable to
inherent problems like damping or exploding gradients, in addition to its inability
to handle non-differentiable functions.
Based on the analysis of gradient descent problems, it is proposed to use a
complex optimizer to train a neural network.
An optimizer is an algorithm for achieving the best, most accurate results
while increasing the learning rate. In other words, it is an algorithm used to change
parameters such as weights of neural network and learning rates slightly so that the
model is adequate and can produce accurate results quickly. To do this, the
optimizer algorithm adjusts the neural network connection weights [9,10].
We proposed to use the popular optimization algorithm in the field of deep
learning – optimizer ADAM. The name is derived from Adaptive Moment
Estimation.
The algorithm has the following advantages compared to other optimizers:
the algorithm works effectively with online and streaming data, works well
with non-stationary data, for example, it allows to optimize noise [11,12];
the algorithm allows to very accurately approximate a set of points with a
linear (Fig. 11) or nonlinear function (Fig. 12);
the algorithm has a simple implementation and computational efficiency,
provides good accuracy compared to other optimizers, and is not demanding on
computer resources [13,14];
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin
ISSN 1681–6048 System Research & Information Technologies, 2026, № 1 164
an important advantage of ADAM is that updating the wt parameter is
completely invariant to gradient scaling, the algorithm converges even if the
objective function changes over the time or the objective function is nonlinear and
contains several local extrema, which is relevant for the data transmission over a
network (Fig. 13).
Fig. 11. Linear approximation function Fig. 12. Nonlinear approximation function
Fig. 13. Function with several local extrema
Let’s consider the ADAM algorithm for optimizing network traffic.
The optimizer is called Adaptive Moment Estimation because it uses
estimation of the first moment of the gradient, m1 (the mean), and estimation of the
second moment of the gradient, m2 (the uncentered variance) to adapt the learning
rate for each weight of the neural network.
The algorithm calculates the exponential moving average of the gradient pt
and the squared gradient qt, and the parameters m1 and m2 control the decay rate of
these moving averages.
The initial value of the moving averages pt, qt and values of m1 and m2 close to
1 (which are recommended for this algorithm, since the algorithm is stable to the
initial values of the learning rate and damping coefficient), you can see this on
command initialize.
This leads to a shift in the moment estimates 𝑝 𝑞 , towards zero. This
displacement is overcome by correcting the 1st and 2nd order moments. Next, the
weight parameter wt is updated.
The algorithmic model can be presented in the following form.
Physical-informed neural network in signal processing and network traffic communications
Системні дослідження та інформаційні технології, 2026, № 1 165
where f(w) — nonlinear stochastic objective function with parameter w containing
local extremes; g(t) — gradient at time t along w, allows to get the direction to
move towards local extrema; t — timestep, t=1, 2..; m1 — the first moment vector
(mean), the proposed value for ADAM is m1=0.9; m2 — the second moment vector
(uncertained variance), the proposed value for ADAM is m2=0.999; pt — bias of
the first order moment; qt — bias of the second order moment; 𝑝 — bias correction
of the first order moment; 𝑞 — bias correction of the second order moment; α —
initial learning rate, the proposed value for ADAM is α=0.001; ϵ — parameter
preventing division by zero, does not affect learning, the proposed value for ADAM
is ϵ = 10-8.
The implementation of this algorithm allows adaptive adjustment of the
optimal learning speed of each model parameter, which leads to highly accurate
results, but storing the model parameters consumes memory twice as much as the
model itself, which is an obstacle when training large models. In practice, to support
such an algorithm with high memory consumption, it is necessary to use unloading
to the CPU, which increases the delay and slows down the learning process.
RESULTS OF COMPUTATIONAL EXPERIMENTS
This section presents a description of the experimental learning and obtained
results. A practical implementation of PINN learning with the ADAM algorithm
using Python you can see on the Figs. 14–16.
Fig. 14 shows an example of a fragment of a network signal.
Fig. 14. An example of a fragment of a network signal
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin
ISSN 1681–6048 System Research & Information Technologies, 2026, № 1 166
The Fig. 15 shows the restored at these points the initial nonlinear function f,
in our case this is the optimized network signal. From Fig. 15 it follows that we
quite accurately approximated the set of points with a nonlinear function. This was
the purpose of learning — finding unknown parameters (𝑝 , 𝑞 ) to minimize a
given loss function.
Fig. 15. The restored at points the initial nonlinear function f
Fig. 16 shows the network signal function directly without noise.
Fig. 16. The network signal function directly without noise
Let’s consider the practical example. For example, we tried to send packages
to Tfl.gov.uk from our network in Kharkiv, Ukraine. Let’s complete the following
steps:
trace using the command tracert in Windows Operation system to
determine the number of nodes;
execute a command ping to send packets without using PINN;
calculate network bandwidth utilization for data transfer without using
PINN;
execute a command ping to send packets using PINN;
calculate network bandwidth utilization for data transfer using PINN.
Calculate the result of sending 4 packet of 4096 B to Tfl.gov.uk without using
PINN.
𝑈1 = 4 × 4096 𝐵(27 + 193 + 61 + 49) 𝑚𝑠 × 100% = 0,016384 𝑀𝐵0,327 𝑠 × 100% =
= 0,05 𝑀𝐵𝑠 × 100% = 5,0%
And testing data transmission using the PINN trained with the ADAM
algorithm we obtained the following results:
Physical-informed neural network in signal processing and network traffic communications
Системні дослідження та інформаційні технології, 2026, № 1 167
𝑈2 = 4 × 4096 𝐵(17 + 12 + 15 + 35) 𝑚𝑠 × 100% = 0,016384 𝑀𝐵0,079 𝑠 × 100% =
= 0,207 𝑀𝐵𝑠 × 100% = 20,7%
So, using the PINN trained with the ADAM algorithm to trasnsmit data the
efficiency has increased. Thanks to this method, it is possible to obtain a low noise
signal in practice, due to which network traffic is optimized.
DISCUSSION
Based on the research results presented in this article, the following main
conclusions can be drawn.
Network traffic optimization can be improved at the protocol level and at the
application level.
At the protocol level you can increase the number of transmitted packets. The
more packets are transmitted, the greater efficiency of data transmission will be,
but taking into account the channel capacity and packet loss in the network.
At the application level you can use Physics-Informed Neural Network for
optimization of network bandwidth utilization. While high network utilization
indicates the network is busy, low network utilization indicates the network is idle.
Using the PINN trained with the ADAM algorithm to trasnsmit data the efficiency
is 15.7 %.
CONCLUSION
At the application level it is proposed to use also another neural network
architecture – Learning Vector Quantization (LVQ) with Encoder-Decoder
architecture, which is suitable for optimizing of network noises, texts and images
over the network.
ACKNOWLEDGMENTS
This paper is part of the DIOR project that has received funding from the European
Union’s MSCA RISE programme under grant agreement No. 10100828.
REFERENCES
1. C. Jiang, H. Zhang, Y. Ren, Z. Han, K. Chen, L. Hanzo, “Machine learning paradigms
for next-generation wireless networks,” IEEE Wireless Communications, vol. 24, no.
2, pp. 98–105, 2017. doi: https://doi.org/ 10.1109/MWC.2016.1500356WC
2. X. Zhang, T. Wang, “Elastic and reliable bandwidth reservation based on distributed traffic
monitoring and control,” IEEE Transactions on Parallel and Distributed Systems, vol. 33,
no. 12, pp. 4563–4580, Aug. 2022. doi: https://doi.org/10.1109/TPDS.2022.3196840
3. V. Paxson, F. Floyd, “Wide area traffic: The failure of Poisson modeling,” IEEE/ACM
Transactions on Networking, vol. 3, no. 3, pp. 226–244, Jun. 1995. doi:
https://doi.org/10.1109/90.392383
4. S. Huang, P. Wei, B. Hualaitu, “Bandwidth optimization of information application
system under fine integral method of fuzzy fractional order ordinary differential
equations,” Alexandria Engineering Journal, vol. 59, issue 4,
pp. 2793–2801, 2020. doi: https://doi.org/10.1016/j.aej.2020.06.015
O.V. Zolotukhin, M.S. Kudryavtseva, Y.V. Bodyanskiy, V.O. Filatov, A.V. Antilikatorov, D.V. Kalinin
ISSN 1681–6048 System Research & Information Technologies, 2026, № 1 168
5. G. Karniadakis, I. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, L. Yang, “Physics-
informed machine learning,” Nature Reviews Physics, vol. 3, no. 6, pp. 422–440,
2021. doi: https://doi.org/10.1038/s42254-021-00314-5
6. L. Von Rueden et al., “Informed machine learning – A taxonomy and survey of
integrating prior knowledge into learning systems,” IEEE Transactions on Knowledge
and Data Engineering, 2023, pp. 614–633. doi: https://
doi.org/10.1109/TKDE.2021.3079836
7. E. Bodyanskyi, O. Rudenko, Artificial neural networks: Architecture, training,
application. Kharkiv: Teletech, 2004, 369 p.
8. O. Rudenko, E. Bodiansky, Artificial neural networks. Kharkiv: Smith Company, 2006,
390 p.
9. V. Filatov, O. Zolotukhin, A. Yerokhin, M. Kudryavtseva, “The methods for
the prediction of climate control indicators in the Internet of Things systems,”
CEUR Workshop Proceedings, pp. 391–400, 2021. doi:
https://doi.org/10.5281/zenodo.14526027
10. I.E. Lagaris, A. Likas, D.I. Fotiadis, “Artificial neural networks for solving ordinary
and partial differential equations,” IEEE Transactions on Neural Networks, vol. 9, no.
5, pp. 987–1000, Sept. 1998. doi: https://doi.org/ 10.1109/72.712178
11. Y. Bodyanskiy, O. Zolotukhin, A. Yerokhin, M. Kudryavtseva, M. Yerokhin, “Fast
stacking neuro-neo-fuzzy system for inverse modeling in online mode,” International
Journal of Computing, vol. 24, no. 4, pp. 661–667, 2025. doi:
https://doi.org/10.47839/ijc.24.4.4330
12. O. Zolotukhin, V. Filatov, A. Yerokhin, M. Kudryavtseva, and V. Semenets, “An
approach to the selection of behavior patterns autonomous intelligent mobile
systems,” in IEEE 8th International Conference on Problems of Infocommunications,
Science and Technology (PIC S&T), Kharkiv, Ukraine, 2021, pp. 349–352. doi:
https://doi.org/10.1109/PICST54195.2021.9772110
13. V. Filatov, O. Zolotukhin, M. Kudryavtseva, “Intellectual data analysis
in relational information and analytical systems,” Innovative Technologies and
Scientific Solutions for Industries, no. 4, pp. 101–111, 2025. doi:
https://doi.org/10.30837/2522-9818.2025.4.101
14. J. Uduagbomen, S. Lakshminarayana, M. Leeson, T. Xu, “Physics-informed neural
network modeling of solution pulses in optical communication systems,” 2022 IEEE
Photonics Society Summer Topicals Meeting Series (SUM), Cabo San Lucas, Mexico,
2022, pp. 1–2. doi: https://doi.org/10.1109/ SUM53465.2022.9858244
Received 16.01.2025
INFORMATION ON THE ARTICLE
Oleg V. Zolotukhin, ORCID: 0000-0002-0152-7600, Kharkiv National University of
Radio Electronics, Ukraine, e-mail: oleg.zolotukhin@nure.ua
Maryna S. Kudryavtseva, ORCID: 0000-0003-0524-5528, Kharkiv National University
of Radio Electronics, Ukraine, e-mail: maryna.kudryavtseva@nure.ua
Yevgeniy V. Bodyanskiy, ORCID: 0000-0001-5418-2143, Kharkiv National University
of Radio Electronics, Ukraine, e-mail: yevgeniy.bodyanskiy@nure.ua
Valentin O. Filatov, ORCID: 0000-0002-3718-2077, Kharkiv National University of
Radio Electronics, Ukraine, e-mail: valentin.filatov@nure.ua
Andriy V. Antilikatorov, Kharkiv National University of Radio Electronics, Ukraine,
e-mail: andrii.antilikatorov@nure.ua
Dеnys V. Kalinin, Kharkiv National University of Radio Electronics, Ukraine,
e-mail: denis.kalinin@teamdev.com
Physical-informed neural network in signal processing and network traffic communications
Системні дослідження та інформаційні технології, 2026, № 1 169
НЕЙРОННА МЕРЕЖА З ФІЗИЧНОЮ ІНФОРМАЦІЄЮ ДЛЯ
ОБРОБЛЕННЯ СИГНАЛІВ ТА ЗВ’ЯЗКУ МЕРЕЖЕВОГО ТРАФІКУ/
О.В. Золотухін, М.С. Кудрявцева, Є.В. Бодянський, В.О. Філатов,
А.В. Антілікаторов, Д.В. Калінін
Анотація. Проблема оброблення сигналів та оптимізації мережевого трафіку
вирішується на апаратному рівні і є цікавою, сучасною та актуальною з точки
зору прикладного рівня. Необхідно запропонувати підхід, який поєднує методи
машинного навчання із завданнями пропускної здатності мережі та трафіком по
мережі. Для вирішення цієї задачі запропоновано використовувати концепцію
інформованого машинного навчання (IML), а саме таксономію IML, принципи
побудови систем глибокого машинного навчання на основі інформації про
фізичні властивості досліджуваної мережі передавання даних. Платформою для
розроблення є моделі глибокого машинного навчання з використанням
нейронних мереж з фізичною інформацією (PINN). Нейронна мережа
представляє клас алгоритмів глибокого навчання, які можуть інтегрувати дані з
описом фізичних процесів або без них. Як алгоритм запропоновано
використовувати популярний алгоритм у галузі глибокого навчання – ADAM
для оптимізації мережевого трафіку. Використання PINN, навченого за
алгоритмом ADAM, для передавання даних підвищило ефективність. Завдяки
такому методу на практиці вдається отримати малошумний сигнал, внаслідок
чого оптимізується мережевий трафік.
Ключові слова: оцінка адаптивного моменту, штучний інтелект, штучна
нейронна мережа, візуалізація даних, динамічний нейрон, градієнтне зниження,
інформоване машинне навчання, алгоритм навчання, використання пропускної
здатності мережі, оптимізація мережевого трафіку, нейро-нечітка логіка,
Python.
|
| id | journaliasakpiua-article-358086 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2026-04-20T01:00:29Z |
| publishDate | 2026 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/42/506e59615bf1302fded34c0b968d8a42.pdf |
| spelling | journaliasakpiua-article-3580862026-04-19T21:53:19Z Physical-informed neural network in signal processing and network traffic communications Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку Zolotukhin, Oleg Kudryavtseva, Maryna Bodyanskiy, Yevgeniy Filatov, Valentin Antilikatorov, Andriy Kalinin, Dеnys Adaptive Moment Estimation artificial intelligence artificial neural network Data Visualization dynamic neuron gradient decent Informed Machine Learning learning algorithm network bandwidth utilization network traffic optimization neuro-fuzzy logic Python оцінка адаптивного моменту штучний інтелект штучна нейронна мережа візуалізація даних динамічний нейрон градієнтне зниження інформоване машинне навчання алгоритм навчання використання пропускної здатності мережі оптимізація мережевого трафіку нейро-нечітка логіка Python Problem of signal processing and network traffic optimization is solved at the hardware level and is interesting, modern and relevant from the point of view of the application level. It is necessary to propose an approach that combines machine learning methods with network bandwidth tasks and traffic over the network. To solve this problem, it is proposed to use concept of Informed Machine Learning (IML), that is the Taxonomy of IML, the principles of constructing deep machine learning systems based on information about the physical properties of the data transmission network under study. The platform for developing is deep machine learning models using PINN neural networks. The PINN represents the class of deep learning algorithms that can integrate data with or without physical processes description. As an algorithm it is proposed to use the popular algorithm in the field of deep learning – ADAM (Adaptive Moment Estimation) for optimizing network traffic. Using the PINN trained with the ADAM algorithm to transmit data the efficiency has increased. Thanks to this method, it is possible to obtain a low noise signal in practice, due to which network traffic is optimized. Проблема оброблення сигналів та оптимізації мережевого трафіку вирішується на апаратному рівні і є цікавою, сучасною та актуальною з точки зору прикладного рівня. Необхідно запропонувати підхід, який поєднує методи машинного навчання із завданнями пропускної здатності мережі та трафіком по мережі. Для вирішення цієї задачі запропоновано використовувати концепцію інформованого машинного навчання (IML), а саме таксономію IML, принципи побудови систем глибокого машинного навчання на основі інформації про фізичні властивості досліджуваної мережі передавання даних. Платформою для розроблення є моделі глибокого машинного навчання з використанням нейронних мереж з фізичною інформацією (PINN). Нейронна мережа представляє клас алгоритмів глибокого навчання, які можуть інтегрувати дані з описом фізичних процесів або без них. Як алгоритм запропоновано використовувати популярний алгоритм у галузі глибокого навчання – ADAM для оптимізації мережевого трафіку. Використання PINN, навченого за алгоритмом ADAM, для передавання даних підвищило ефективність. Завдяки такому методу на практиці вдається отримати малошумний сигнал, внаслідок чого оптимізується мережевий трафік. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2026-03-31 Article Article Peer-reviewed Article application/pdf https://journal.iasa.kpi.ua/article/view/358086 10.20535/SRIT.2308-8893.2026.1.11 System research and information technologies; No. 1 (2026); 155-169 Системные исследования и информационные технологии; № 1 (2026); 155-169 Системні дослідження та інформаційні технології; № 1 (2026); 155-169 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/358086/344012 |
| spellingShingle | оцінка адаптивного моменту штучний інтелект штучна нейронна мережа візуалізація даних динамічний нейрон градієнтне зниження інформоване машинне навчання алгоритм навчання використання пропускної здатності мережі оптимізація мережевого трафіку нейро-нечітка логіка Python Zolotukhin, Oleg Kudryavtseva, Maryna Bodyanskiy, Yevgeniy Filatov, Valentin Antilikatorov, Andriy Kalinin, Dеnys Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку |
| title | Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку |
| title_alt | Physical-informed neural network in signal processing and network traffic communications |
| title_full | Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку |
| title_fullStr | Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку |
| title_full_unstemmed | Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку |
| title_short | Нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку |
| title_sort | нейронна мережа з фізичною інформацією для оброблення сигналів та зв’язку мережевого трафіку |
| topic | оцінка адаптивного моменту штучний інтелект штучна нейронна мережа візуалізація даних динамічний нейрон градієнтне зниження інформоване машинне навчання алгоритм навчання використання пропускної здатності мережі оптимізація мережевого трафіку нейро-нечітка логіка Python |
| topic_facet | Adaptive Moment Estimation artificial intelligence artificial neural network Data Visualization dynamic neuron gradient decent Informed Machine Learning learning algorithm network bandwidth utilization network traffic optimization neuro-fuzzy logic Python оцінка адаптивного моменту штучний інтелект штучна нейронна мережа візуалізація даних динамічний нейрон градієнтне зниження інформоване машинне навчання алгоритм навчання використання пропускної здатності мережі оптимізація мережевого трафіку нейро-нечітка логіка Python |
| url | https://journal.iasa.kpi.ua/article/view/358086 |
| work_keys_str_mv | AT zolotukhinoleg physicalinformedneuralnetworkinsignalprocessingandnetworktrafficcommunications AT kudryavtsevamaryna physicalinformedneuralnetworkinsignalprocessingandnetworktrafficcommunications AT bodyanskiyyevgeniy physicalinformedneuralnetworkinsignalprocessingandnetworktrafficcommunications AT filatovvalentin physicalinformedneuralnetworkinsignalprocessingandnetworktrafficcommunications AT antilikatorovandriy physicalinformedneuralnetworkinsignalprocessingandnetworktrafficcommunications AT kalinindenys physicalinformedneuralnetworkinsignalprocessingandnetworktrafficcommunications AT zolotukhinoleg nejronnamerežazfízičnoûínformacíêûdlâobroblennâsignalívtazvâzkumereževogotrafíku AT kudryavtsevamaryna nejronnamerežazfízičnoûínformacíêûdlâobroblennâsignalívtazvâzkumereževogotrafíku AT bodyanskiyyevgeniy nejronnamerežazfízičnoûínformacíêûdlâobroblennâsignalívtazvâzkumereževogotrafíku AT filatovvalentin nejronnamerežazfízičnoûínformacíêûdlâobroblennâsignalívtazvâzkumereževogotrafíku AT antilikatorovandriy nejronnamerežazfízičnoûínformacíêûdlâobroblennâsignalívtazvâzkumereževogotrafíku AT kalinindenys nejronnamerežazfízičnoûínformacíêûdlâobroblennâsignalívtazvâzkumereževogotrafíku |