Електронний архів наукових та освітніх матеріалів КПІ ім. Ігоря Сікорського
ELAKPI – інституційний репозитарій, що накопичує, зберігає, розповсюджує та забезпечує довготривалий, постійний та надійний доступ через Інтернет до наукових та освітніх матеріалів професорсько-викладацького складу, співробітників, студентів, аспірантів та докторантів КПІ ім. Ігоря Сікорського. За посиланням можна ознайомитися з положенням про ELAKPI.
Доступ до матеріалів ELAKPIДоступ до повних текстів матеріалів ELAKPI вільний в мережі Інтернет, крім:
Щоб отримати права на перегляд/скачування повних текстів ресурсів, доступних тільки в локальній мережі університету, зареєстровані користувачі Бібліотеки КПІ ім. Ігоря Сікорського можуть скористатися послугою Віддалений доступ до "локальних" ресурсів. | КонтактиБібліотека КПІ ім. Ігоря Сікорського, зал № 4.4, тел. +38 (044) 204-96-72, elakpi@library.kpi.ua, elakpi.ntb@gmail.com |
Фонди
Виберіть фонд, щоб переглянути його зібрання.
- Вільний доступ
- Доступ у читальній залі № 6.6 НТБ
- Припинив існування 1.07.2020
- Доступ у читальній залі № 6.6 НТБ
- Стара назва до 1.09.2025 року - Інженерно-хімічний факультет (ІХФ)
- Припинив існування 1.07.2018
- Раніше - Факультет прикладної математики (ФПМ)
- Припинив існування 1.07.2020
Нові надходження
UAeroNet: domain-specific dataset for automation of unmanned aerial vehicles
(National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute", 2025) Kochura, Yuriy; Trochun, Yevhenii; Taran, Vladyslav; Gordienko, Yuri; Rokovyi, Oleksandr; Stirenko, Sergii
This paper addresses the challenges and key principles of designing domain-specific datasets that canbe used especially for automation of unmanned aerial vehicles. Such datasets play a key role in buildingintelligent systems that enable autonomous operation and support data-driven decisions. The study presentsapproaches we used for data collection, analysis and annotation, highlighting their importance and practicalimpact on real-world application. The preparation of a domain-specific dataset for automating unmannedaerial vehicles operations (such as navigation and environmental monitoring) is a challenging task due tofrequently low image resolution, complex weather conditions, a wide range of object scales, backgroundnoise and heterogeneous terrain landscapes. Existing open datasets typically cover only a limited variety ofunmanned aerial vehicles use cases, which restricts the ability of deep learning models to perform adequatelyunder non-standard or unpredictable conditions.The object of the study is video data acquired by unmanned aerial vehicles for creating domain-specificdatasets that enable machine learning models to perform autonomous object recognition, navigation, obstacleavoidance and interaction with an environment with minimal operator involvement. The subject focuseson the collection, preparation and annotation of video data acquired by unmanned aerial vehicles. Thepurpose of the study is to develop and systematize workflow for creating specialized datasets to trainrobust models capable of autonomously recognizing objects in real-time video captured by unmanned aerialvehicles. To achieve this goal, a workflow was designed for collecting and annotating video data, raw videodata were acquired from unmanned aerial vehicles sensors and manually annotated using the ComputerVision Annotation Tool.As a result of this work, we developed a domain-specific dataset (UAeroNet) using an open-sourceannotation tool for object tracking task in real scenarios.UAeroNetconsists of 456 annotated tracks and atotal of 131 525 labeled instances that belong to 13 distinct classes.
DDOS attack detection with data imperfections using machine learning algorithms
(National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute", 2025) Dremov, Artem; Volokyta, Artem
The issue of DDoS attacks remains a prevalent one even in recent years. Modern environment is highlydynamic and is characterized by a large amount of traffic flow. Existing research covers several models,techniques and approaches to detecting DDoS traffic, which aim to optimize the detection in controlleddatasets. However, unintentional noise or data corruption may lower the efficacy of such methods. As such,determining most effective ways to detect DDoS traffic in conditions of data imperfections is necessary forreliable network performance.Therefore, the object of this research Is the usage of machine learning algorithms for detection ofincoming DDoS attacks. The purpose of this research is to determine the performance of ways to detectincoming DDoS attacks with machine learning algorithms based on detection accuracy, while simulatingimperfect data conditions. The study also examines the impact of class rebalancing on modified data.To achieve the aim of this research a variety of machine learning algorithms were implemented andtested on aCIC-DDoS2019dataset. The data is modified by removing values and introducing noise, tested,the classes are resampled and the dataset is tested again. The goal is to achieve over 90% accuracy in aclassification task of the type of DDoS attack and to determine how much the changes affect the performanceof the algorithms.The results of the testing indicated that several solutions reach the target mark and changes to thedataset in realistic conditions do not significantly affect the final result. However, all models tested showa decrease in accuracy compared to unmodified data with more complex models showing higher resilience(smaller decrease in accuracy). In addition, resampling of the data shows comparable decrease in accuracyof the models with more complex models being affected less.The results of this study may be used in development of an algorithm of repairing the corrupted dataor development of models more resistant to such data changes. Additionally, the results of this study maybe used when considering models for practical implementations of a DDoS traffic classification system.
Evaluation of the effectiveness of two approaches to building damage detection with satellite imagery
(National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute", 2025) Oliinyk, Yurii; Rumiantsev, Oleksii
This study addresses the approaches for satellite image analysis to assess infrastructure damage. Themain aim is to conduct a comprehensive comparative analysis of the effectiveness of two key machinelearning approaches: specialized semantic segmentation based on theU-Netarchitecture and generalizedvisual analysis using large vision-language models. The object of the research is the process of quantitativelybenchmarking these two distinct approaches to determine their practical applicability for multi-class damageclassification.The research material is the publicly availablexView2dataset. The methods involved two parallelexperiments. For the semantic segmentation approach, aU-Netmodel with anEfficientNet-B4encoderwas implemented and trained on 6-channel input data (”before” and ”after” images) using a combinedDiceandFocalloss function. For the vision-language models approach, the open-sourceLLaVA-1.5-7Bmodelwas evaluated in a zero-shot mode using advanced prompt engineering for an aggregative counting task.To enable a direct comparison, the standardJaccard indexwas calculated based on the aggregated objectcounts for each damage class.The results of the experiments revealed a significant performance disparity. The specializedU-Netmodeldemonstrated high effectiveness, achieving an intersection over union score of 0.6141 on the test set. Incontrast, theLLaVAmodel proved unsuitable for accurate quantitative analysis, yielding an extremely lowJaccard indexof approximately 0.063, primarily due to its systemic failure to correctly identify and countobjects (𝑅𝑒𝑐𝑎𝑙𝑙≈0.07). The scientific novelty lies in being the first study to quantitatively document thisorder-of-magnitude capability gap, confirming that for tasks requiring high-precision mapping, specializedsegmentation models remain the indispensable tool.
Comparative analysis of LCNet050 and MobileNetV3 architectures in hybrid quantum–classical neural networks for image classification
(National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute", 2025) Khmelnytskyi, Arsenii; Gordienko, Yuri
This study explores the impact of classical backbone architecture on the performance of hybrid quantum-classical neural networks in image classification tasks. Hybrid models combine the representational power of classical deep learning with the potential advantages of quantum computation. Specifically, this research employs a quanvolutional neural network architecture in which a quantum convolutional layer, based on a four-qubit Ry circuit, preprocesses input images before classical processing. Despite the growing interest in hybrid models, few studies have systematically investigated how variations in classical architecture design affect the overall performance of hybrid quantum-classical neural networks. To address this gap, we compare two lightweight convolutional backbones – MobileNetV3Small050 and LCNet050 – integrated with an identical quantum preprocessing layer. Both models are evaluated on the CIFAR-10 dataset using 5-fold stratified cross-validation. Performance is assessed using multiple metrics, including accuracy, macro- and micro-averaged area under the curve, and class-wise confusion matrices. The results indicate that the LCNet-based hybrid model consistently outperforms its MobileNet counterpart, achieving higher overall accuracy and area under the curve scores, along with improved class balance and robustness in distinguishing less-represented classes. These findings underscore the critical role of classical backbone selection in hybrid quantum-classical architectures. While the quantum layer remains fixed, the synergy between quantum preprocessing and classical feature extraction significantly affects model performance. This study contributes to a growing body of work on quantum-enhanced learning systems by demonstrating the importance of classical design choices. Future research may extend these insights to alternative datasets, deeper or transformer-based backbones, and more expressive quantum circuits.
Optimization neural network for time series processing
(National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute", 2025) Pysarchuk, Oleksii; Baran, Danylo
The article proposes the architecture of the optimization neural network and the model of test samplesynthesis for the process of extrapolation of time series parameters. In particular, the addition of an inputlayer with the introduction of an optimization scheme of nonlinear trade-offs has been implemented.Extrapolation of the behavior of the time series was carried out according to a test sample, which isformed as a data model with the selection of the trend according to the method of least squares. Thescientific novelty of the results obtained in the article is reflected in the essence of these decisions.The aim of the research is to develop an optimization network architecture and data model forextrapolation, which allows to improve the accuracy and time of predicting the behavior of the time seriesoutside the observation interval. Subject of research: architecture of an artificial neural network andmethods of extrapolation of time series. Object of research: processes of architectural synthesis of anartificial neural network and extrapolation of time series behavior outside the observation interval.The optimization layer provides mini-requirements for the approximation of training and test samples.This is especially appropriate for time series with stochastic noise and allows you to reduce the impactof random errors on time series prediction results. The use of model data for extrapolation allows you todetermine the behavior of the time series outside the observation interval. At the same time, the forecastingtime with acceptable accuracy characteristics increases. These solutions are reflected in the name of theoptimization neural network, which is proposed by the authors. The study of the effectiveness of the proposedsolutions was implemented by methods of simulation modeling on a modified artificial neural network. Theresults of the calculations proved an increase in the adequacy of data models and an increase in the accuracyof extrapolation.