2687-0517 15 4 2022 1-107

RAR RUS 7-21 ITMO University Prokofiev Kirill baterflyrity@yandex.ru

St. Petersburg, Russian Federation

0000-0002-1128-2942 ITMO University Ivanov Sergey sergei.v.ivanov@gmail.com

St. Petersburg, Russian Federation

SAR images generation from optical imagery Technological level boost in the industry has led to widespread utilization of neural networks and machine learning that require large datasets. Data processing in unmanned aerial vehicles faces several difficulties such as high equipment cost, no-fly areas and complicated imagery permission acquisition. Public datasets can be deficient and impractical. Researches are forced to either work without initial data or hire surveyors and 3D-designers. The paper solves this problem by generation of synthetic aperture radar images from publicly available optical satellite maps available anywhere on Earth. Two methods are discovered: physical processes of radio wave propagation modeling and color image convolution into grayscale. The physical model is constructed in the first approximation as a linear propagation of an electromagnetic wave with a single point of reflection, insignificant radio waves atmosphere propagation and reflection effects are omitted. Various algorithms of convolution, linear transformation of RGB color space into grayscale are examined: YPbPr, HSV, linear regression. The physical model is discarded due to its practical inapplicability and complexity of implementation. After evaluating the convolution results according to the maximum likelihood criteria, preference is given to the YPbPr algorithm. Additional steps are proposed for more accurate generation: noise addition and space transformation. The resulting algorithm generates visually and mathematically adequate pseudo-radar images to obtain initial datasets, correlated radar and optical images. The datasets are supposed to improve neural networks. The method gives an advantage over analogues in rural and wild clutters but loses in urban areas. In further studies, we propose to combine the results with objects of interest generation. 10.18721/JCSTCS.15401 004.932 radar imagery generator SAR satellite map transformation optic to SAR translation aerial imagery processing https://infocom.spbstu.ru/article/2022.75.1/ 7-21.pdf

RAR RUS 22-36 0000-0001-5439-4277 Peter the Great St. Petersburg Polytechnic University Zhu Yuqing 1918149382@qq.com

St. Petersburg, Russian Federation

Formation of flight control for a group of unmanned aerial vehicles based on algorithm of multi-agent swarm model The problem of controlling a group of unmanned aerial vehicles (UAVs) is considered to organize the movement of a swarm along a given trajectory, which ensures the most effective achievement of the flight goal. The issues of choosing a mathematical model of the spatial motion of a group of UAVs, suitable for solving the problem of synthesis of coordinated control of the entire set of aircrafts, are discussed. Taking into account the specifics of the requirements for the space-time position of individual UAVs (agents) in a group, it is proposed to use a model with a leader. A group of agents has a virtual leader who plans the route of the group in accordance with a given task and tracks a specific goal of movement. The virtual leader calculates its own motion control with a trajectory-tracking or target-tracking algorithm to move along the desired trajectory. In this case, the guidance signal can allow individual UAVs to gather at the position of the virtual leader and correspond to the velocity vector of the virtual leader in order to communicate the topology of the multi-agent system and ensure swarm formation. 10.18721/JCSTCS.15402 681.5 control system unmanned aerial vehicle mathematical model of the group coordination swarm management https://infocom.spbstu.ru/article/2022.75.2/ 22-36.pdf

RAR RUS 37-50 Peter the Great St. Petersburg Polytechnic University Vlasenko Nataliia nt.vlasenko@yandex.ru

St. Petersburg, Russian Federation

Peter the Great St. Petersburg Polytechnic University Dusaeva Anelya an.dusaeva@gmail.com

St. Petersburg, Russian Federation

Nikiforov Igor igor.nikiforov@gmail.com Peter the Great St. Petersburg Polytechnic University Prelovskii Dmitrii dimaprelovski@gmail.com

St. Petersburg, Russian Federation

Technique for automating charging of an electric vehicle based on a Raspberry Pi controller using neural networks The expansion of Russian market of electric and autonomous vehicles leads to an increase in demand for automation of contactless charging (without driver participation). The article proposes a method of contactless charging of electric vehicles, which involves automatically determining the type of car charging connector, selecting the appropriate charger and connecting it to the charging connector of an electric vehicle through the use of a robot manipulator. A feature of the technique is the determination of the type and coordinates of the location of the charging connector of the car by reading images obtained from the camera of a gas station in real time and processing them with a convolutional neural network model. A study was conducted, and a function was selected that allows optimally solving the problems of classification of charging connectors, which ensures maximum accuracy of the result. The volume of the training sample for the neural network was used in the amount of 10,000 images from a synthetic data set, which was created on the basis of three types of the most popular three-dimensional models of charging connectors on various backgrounds. The proposed technique is implemented in a prototype of a software and hardware control complex for a manipulative robot based on a Raspberry Pi controller. 10.18721/JCSTCS.15403 004.032 electric vehicle computer vision convolutional neural network robotic arm charging https://infocom.spbstu.ru/article/2022.75.3/

RAR RUS 51-63 Peter the Great St.Petersburg Polytechnic University Antonov Alexander antonov@eda-lab.ftk.spbstu.ru

Polytechnicheskaya, 29, St.Petersburg, 195251, Russian Federation

Peter the Great St. Petersburg Polytechnic University Besedin Denis 1310nero@mail.ru

St. Petersburg, Russian Federation

Filippov Aleksey filippov@eda-server.ftk.spbstu.ru Research and comparative analysis of the effectiveness of software and hardware implementations of the operation of summing transposed matrices The article is devoted to the study and comparative analysis of the software and hardware implementation of the operation of summing transposed matrices and its modified version – the operation of transposing the sum of matrices. A feature of the study is the use of high-level synthesis tools to obtain a hardware implementation. The relevance of the study is due to the widespread use of matrix operations for solving problems of various classes, the power asymptotic complexity of matrix calculations and the lack of data on the use of this toolkit in the tasks of creating hardware devices for matrix calculations. A step-by-step method of synthesis and optimization of a hardware device is proposed. A comparative study of software and hardware implementations of two computational tasks is carried out. It is shown that a large gain in the performance of hardware implementations is obtained by increasing the degree of parallelism of calculations. Additionally, conclusions are drawn about the inefficiency of attempts to achieve high clock frequencies, as well as about the increase in resources spent with increased speed due to parallelization. 10.18721/JCSTCS.15404 004.312.44 hardware implementation performance hardware costs FPGA parallel computing pipelining https://infocom.spbstu.ru/article/2022.75.4/ 51-63.pdf

RAR RUS 64-72 Yandex Pismenny Alexey pismennyy.aleks@gmail.com

Moscow, Russian Federation

National Research University “Higher School of Economics” Sokolov Evgeny sokolov.evg@gmail.com

Moscow, Russian Federation

Token-wise approach to span-based question answering Language model pre-training has led to significant success in a wide range of natural language processing problems. It was shown that modern deep contextual language models need only a small number of new parameters for fine-tuning due to the power of the base model. Nevertheless, the statement of the problem itself makes it possible to search the new approaches. Our experiments relate to the span-based question answering, one of machine reading comprehension (MRC) tasks. Recent works use loss functions that require the model to predict start and end positions of the answer in a contextual document. We propose a new loss that additionally requires the model to correctly predict whether each token is contained in the answer. Our hypothesis is that explicit using of this information can help the model to learn more dependencies from data. Our solution also includes a new span’s ranking and a no-answer examples selection scheme. We also propose approaches of accounting for information about relative positions of tokens in the dependency trees and the types of dependencies in relation to syntax-guided attention. The experiments showed that our approaches increase the quality of BERT-like models on SQuAD datasets. 10.18721/JCSTCS.15405 004.852 machine learning natural language processing question answering machine reading comprehension dependency parsing https://infocom.spbstu.ru/article/2022.75.5/

RAR RUS 73-85 Yandex LLC Fadeeva Ekaterina rediska@yandex-team.ru

St. Petersburg, Russian Federation

Yandex LLC Ershov Vasily noxoomo@yandex-team.ru

St. Petersburg, Russian Federation

Multi-channel transformer: A transformer-based model for multi-speaker speech recognition Most of the modern approaches to multi-speaker speech recognition are either not applicable in case of overlapping speech or require a lot of time to run, which can be critical, for example, in case of real-time speech recognition. In this paper, a transformer-based end-to-end model for overlapping speech recognition is presented. It is implemented by using a generalization of the standard approach to speech recognition. The introduced model achieves results comparable in quality to modern state-of-the-art models, but requires less model calls, which speeds up the inference. In addition, a procedure for generating synthetic data for model training is described. This procedure allows to compensate for the lack of real multi-speaker speech training data by creating a stream of data from the initial collection. 10.18721/JCSTCS.15406 004.8 speech recognition multi-speaker speech recognition diarization speech separation voice technologies https://infocom.spbstu.ru/article/2022.75.6/

RAR RUS 86-97 National Research University “Higher School of Economics” Tarasov Denis tarasov.denis.al@gmail.com

St. Petersburg, Russian Federation

Yandex LLC Ershov Vasily noxoomo@yandex-team.ru

St. Petersburg, Russian Federation

Fixing 1-bit Adam and 1-bit LAMB algorithms Today, various neural network models are trained using distributed learning in order to reduce the time spent. The most common way of distributed learning today is the approach, in which the data are divided into parts and sent along with the model to different devices, each device calculates updates for the model, then the updates are aggregated on the server, the server updates the weights of the model and transfers their new version to the devices. Slow network communication between devices can significantly reduce distribution efficiency. Recent studies propose one-bit versions of the Adam and LAMB algorithms, which can significantly reduce the amount of transmitted information, thus improving the scalability of training. However, it turned out that these algorithms diverge in some neural network architectures. The goal of this work is an empirical study of these algorithms, to find the solution of the discovered divergence problem and propose new aspects of testing gradient descent algorithms. 10.18721/JCSTCS.15407 004.852 machine learning deep learning gradient descent distributed training optimization https://infocom.spbstu.ru/article/2022.75.7/

RAR RUS 98-107 Dagestan State University of National Economy Kobzarenko D.N. Dmitry kobzarenko_dm@mail.ru

Makhachkala, Russian Federation

Dagestan State University of National Economy Mustafaev Arslan arslan_mustafaev@mail.ru

Makhachkala, Russian Federation

Dagestan State University of National Economy Gasanova Zarema cudakharka@yandex.ru

Makhachkala, Russian Federation

Dagestan State University of National Economy Magomedova Dinara mdc-101085@mail.ru

Makhachkala, Russian Federation

One-dimensional convolutional layers in a neural network for wind speed time series analysis Data analysis using neural networks and deep machine learning is one of the current trends in scientific research in various fields. One of the scientific tasks of this direction is the study and prediction of time series using artificial intelligence. The article discusses the results of experiments on adding one-dimensional convolutional layers to a neural network within the framework of the task of classifying meteorological time series data – wind speed. The accuracy of the forecast is shown to increase due to the inclusion of one-dimensional convolutional layers in the model. The increase in accuracy on the test data set for the problem under consideration is about 9.5 %. Several variants of architectures for building a model with one-dimensional convolutional layers and evaluating the accuracy of their classification after machine learning are given. The results obtained allow us to conclude that the use of one-dimensional convolutional layers in the neural network architecture is effective for identifying and predicting a time series of meteorological parameters. 10.18721/JCSTCS.15408 004.8 neural network deep machine learning 1D convolutional layer wind speed time series https://infocom.spbstu.ru/article/2022.75.8/