457
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Discipline
Düşük iletişim yüklü olay-tetikli dağıtık kestirim
We propose a novel algorithm for distributed processing applications constrained by the available communication resources using diffusion strategies that achieves up to a 10^3 fold reduction in the communication load over the network, while delivering a comparable performance with respect to the state of the art. After the computation of the local estimates, the information is diffused among the processing elements (or nodes) non-uniformly in time by conditioning the information transfer on level-crossings of the diffused parameter, resulting in a greatly reduced communication requirement. We provide the mean and mean-square stability analyses of the proposed algorithm, and illustrate the gain in communication efficiency compared to other reduced-communication distributed estimation schemes.
Radar anten taramalarının otomatik analizi
Estimation of the radar antenna scan period and recognition of the antenna scantype is usually performed by human operators in the Electronic Warfare (EW)world. In this thesis, we propose a robust algorithm to automatize these two criticalprocesses. The proposed algorithm consists of two main parts: antenna scanperiod estimation and antenna scan type classification. The first part of the algorithminvolves estimating the period of the signal using a time-domain approach.After this step, the signal is warped to a single vector with predetermined size(N) by resampling the data according to its period. This process ensures thatthe extracted features are reliable and are solely the result of the different scantypes, since the effect of the different periods in the signal is removed. Four differentfeatures are extracted from the signal vector with an understanding of thephenomena behind the received signals. These features are used to train naiveBayes classifiers, decision-tree classifiers, artificial neural networks, and supportvector machines. We have developed an Antenna Scan Pattern Simulator (ASPS)that simulates the position of the antenna beam with respect to time and generatessynthetic data. These classifiers are trained and tested with the syntheticdata and are compared by their confusion matrices, correct classification rates,robustness to noise, and computational complexity. The effect of the value of Nand different signal-to-noise ratios (SNRs) on correct classification performanceis investigated for each classifier. Decision-tree classifier is found to be the mostsuitable classifier because of its high classification rate, robustness to noise, andcomputational ease. Real data acquired by ASELSAN Inc. is also used to validatethe algorithm. The results of the real data indicate that the algorithm isready for deployment in the field and is capable of being robust against practicalcomplications.
İnsan gözünün topoğrafik görüntülerının özelliklerini elde etmek için akilli algoritma oluşturmak
Precise diagnosis for a wide range of diseases infecting the human eye is a commitment. Therefore, developing new, smart algorithms is necessary to enhance doctors' diagnostic decisions. The recently invented Pentacam® is a measurement system that introduces topographic maps for the cornea, measures changes upon it, and helps doctors to make a precise diagnosis. This study extracts features from corneal topographic maps to improve the Pentacam® readings and further support precise diagnosing by using deep learning techniques, with an analytical view of the extracted features. A 16-layer convolutional neural network (CNN) was trained using the VGG-16 network to extract powerful features from corneal topographic maps. A sample of 732 human eyes were selected from enlarged topographic images from both genders (414 females and 318 males), divided into two groups: normal and abnormal. The patients' ages ranged from 12 to 76 years. The procedure of the study consisted of three major steps: (1) classification of the extracted features according to the refractive map type (where the estimated accuracy was 96.6%); (2) prediction of the clinical state (normal or abnormal) per individual map (where the estimated accuracy was 88.8%, 98.9%, 94.8%, and 94.5% for the sagittal map, the elevation front map, the elevation back map, and the corneal thickness map, respectively); and (3) comparison of the predicted results and clinical decision-making. The agreement between them reaches about 94.72%, which indicated the power and usefulness of the proposed algorithm.
İç ortamlarda insan-robot etkileşimi
Bu tez çalışmasında robotik alanında tasarlanan İnsan-Robot Etkileşimine (İRE) diyalog kurma ve geliştirme konusunda etkili çözümler üreten üç model önerilmiştir. Robotların iç ortamlarda insanlarla birlikte çalışması veya insanları bilgilendirmesi amacıyla kullanılmak istendiğinde karşılaşılan ilk problemin insan robot arasındaki diyaloğun nasıl geliştirilmesi gerektiği ve etkili diyaloğun nasıl sağlanması gerektiğidir. Bu amaçla İRE alanındaki çalışmalar incelenmiş ve kullanılan yöntemler araştırılmıştır. Bu çalışmalardan esinlenilerek, İRE alanında, robotik sistemlerde kullanılabilecek diyalog modelleri üzerinde geliştirmeler yapılmıştır. Farklı diyalog modelleri için sonlu durum makinesi, içerikten bağımsız gramer yöntemlerinden faydalanılmıştır. Bu kapsamda sonlu durum makineleri ve içerikten bağımsız gramer yöntemleri incelenmiştir. Bu yöntemlerle diyalog modellemeleri geliştirilmiştir. Gerçekleştirilen bu diyalog modelleri gerçek zamanlı olarak robot üzerinde denenmiştir. Önerilen modellerde görüntü işleme ve ses işleme algoritmaları ile insan robot arasında interaktif etkileşimi sağlayan araçlar kullanılmıştır. İRE ile alakalı iki adet deney ortamı oluşturulmuştur. İlk deneyde robot insanı bulduğu zaman kafasıyla karşısındaki insanının yüz hareketlerini taklit ederek kafa takibini başlatıyor. Robot başlangıç durumunda karşısındaki insandan etkileşim beklemektedir. Eğer robotun karşısındaki insan robotla belirlenen süre içerisinde etkileşime geçmez ise robot karşısındaki insana kendisini nasıl kullanmasını anlatarak etkileşimi başlatıyor. İnsan etkileşime geçer ise robot sorulan soruları yanıtlamaktadır. Etkileşim robotu programlayan programcının belirlediği Sonlu Durum Makinesine (SDM) göre tamamlanıyor. Üç farklı SDM önerilmiştir. Bunlar "İnsan kontrol modeli", "Kafa takip modeli" ve "Robot gözünden kontrol modeli" olarak adlandırılmaktadır. Oluşturulan iki deney ortamı önerilen üç model ile çaprazlanarak altı deney beş tekrar ile gerçekleştirilmiştir. Yapılan deneylerde SDM modelleri arasında başarılı etkileşimler belirlenmiş. En uygun model tespit edilmiştir. İleride yapılacak çalışmalarda insanların yaş, cinsiyet ve duygu analizi yapılarak bu oranın daha da arttırılması hedeflenmektedir.
Beyin tümörü tedavisi için kemoterapi ilaç salınım sistemi modellenmesi ve kontrolü
Beyin tümörleri, merkezi sinir sisteminin karmaşık yapısı nedeniyle erken tanı ve etkili tedavi süreçlerini zorlaştıran, yüksek mortaliteye sahip kanser türlerinden biridir. Tedavi sürecinde en sık kullanılan yöntemlerden biri olan kemoterapi, sağlıklı hücreleri de etkileyebilmekte ve hastanın bağışıklık sistemini baskılayabilmektedir. Bu nedenle kemoterapide uygulanan ilaç dozunun bireyselleştirilmiş, güvenli ve optimum şekilde kontrol edilmesi büyük önem taşımaktadır. Bu çalışmada, beyin tümörü tedavisinde kemoterapi ilaç dozajının kontrolü için Tip-2 Bulanık Mantık Denetleyici (T2BMD) tasarlanmıştır. T2BMD'nin, üyelik fonksiyonlarındaki belirsizlikleri daha başarılı bir şekilde yönetmesi sayesinde daha esnek ve hassas kontrol mekanizmaları sunduğu gösterilmiştir. Denetleyici sisteme ait parametreler doğadan esinlenen yeni nesil bir algoritma olan Yapay Sinekkuşu Algoritması (Artificial Hummingbird Algorithm, AHA) ve literatürde oldukça yaygın olan Parçacık Sürü Optimizasyonu (PSO) algoritması ile optimize edilmiştir.
Derin öğrenme kullanilarak cilt kanserinin tespiti ve siniflandirilmasi
One of the worst malignancies is skin cancer. It is expected to spread to other parts of the body if it is not identified and treated at the outset. The method for successfully treating skin cancer uses both image processing and deep learning. In this research, three different kinds of skin cancer, including melanoma, pigmented benign keratoses, and basal cell carcinoma are introduced for detection and classification using efficient methods in Machine Learning (ML) such as K-Mean clustering and Multi-class support vector machine (M-SVM) algorithm, and then Deep Learning (DL) techniques such as AlexNet and ResNet are used. Additionally, deep convolutional neural networks (CNN) effectiveness and capacity are seen. The data set contains 1176 images of skin cancer with different classes of disease, this data set is used in both ML and DL. In ML the data set is divided into training and testing sets, these sets are pass through a few steps of enhancement using a canny edge detection filter and extracting the features using the Gray-Level Co-occurrence Matrix (GLCM) method. Then, these images are segmented into three clustering using K-mean clustering algorithms. In DL two models are used AlexNet and ResNet, in these models the data sets are divided into training and testing sets. Then, an augmentation technique has been proposed, it's very useful for small data sets, and the results show changes in accuracy result. The results show an accuracy of 99.03%, and 97.02% for AlexNet, and ResNet respectively.
Elektrik akıllı şebeke stabilitesi için yapay zeka modellerinin uygulanması
This study does a comparative analysis of several models, including XGBoost, SVM, Random Forest, KNN, Logistic Regression, Decision Tree, Neural Network, SimpleRNN, LSTM, and GRU. The performance of these models is evaluated based on metrics such as accuracy, precision, recall, and F1-score. The XGBoost Classifier is considered the optimal model due to its superior combination of precision, computational effectiveness, and interpretability. The findings presented in this study underscore the considerable capacity of machine learning in facilitating predictive analytics within the context of smart grids. Moreover, these results establish a robust basis for further research endeavours in this domain. Nevertheless, the literature highlights some challenges, such as the intricate nature of models, their limited interpretability, and the substantial computational demands they impose. These issues underscore the necessity for more research and enhancements in this field. As the study concludes, this paper offers strategic recommendations for effectively integrating the findings into actual applications of smart grids. Additionally, it outlines potential avenues for future research in this field.
Kullanilan iot cihazlarinda kullanici gizliliğimakine ve derin öğrenme yaklaşimlar
The swift expansion of Internet of Things (IoT) technology has sparked concerns regarding the privacy of users, since these devices often collect and transmit vast amounts of personal information. To address these issues, this thesis will look at the use of machine and deep learning technologies to improve user privacy on IoT devices. First, the research will examine the existing state of user privacy on IoT devices, as well as the issues associated with protecting privacy. This will include a discussion of the many types of data gathered and communicated by IoT devices, as well as the numerous ways in which this data might be exploited or hacked. The research will also look at current legislative frameworks and best practices in the sector for preserving user privacy on IoT devices. The thesis will then investigate the application of machine learning approaches to improve user privacy on IoT devices. This will take a look at the many machine learning techniques that may be used for this, such as decision tree algorithms and ANNs. The research will also look at the possible benefits and drawbacks of utilizing these algorithms for privacy protection, such as the trade-offs between privacy and other objectives like performance or accuracy. Besides machine learning, the project will look into the use of deep learning technologies for improving user privacy on IoT devices. Deep learning models, a specific category of machine learning techniques, have demonstrated significant potential across various applications. The research will examine the potential benefits and challenges of applying deep learning algorithms for privacy protection on IoT devices, as well as the present related works in this field. Finally, the dissertation will conclude with a discussion of the potential future direction of research in this area, including the potential for integrating machine and deep learning approaches with other privacy-enhancing technologies and the potential for additional regulatory or industry-led efforts to improve user privacy on IoT devices. This dissertation intends to offer a complete assessment of the present status of user privacy on IoT devices, as well as the possibilities for enhancing privacy via the application of machine and deep learning technologies. The study intends to contribute to continuing efforts to secure user privacy in the rapidly developing realm of IoT by addressing these challenges.
Değiştirilmiş SVM iş akışı kullanılarak meme kanserinin yerleştirilmesi ve tespiti
Mammography is the most effective method in the early detection of breast cancer, which can detect up to 90% of cases. mammography was used in only 24% of cases in 2017, which is below the 70% expected by the World Health Organization (WHO). This explains one of the causes of late diagnosis and the increase in mortality. According to WHO data, the average time to start treatment is 120 days after the first appointments. This situation is due to the delay in scheduling consultations, in addition to the fact that many times excessive tests are requested, which slow down the diagnostic process. In this paper, we Propose a medical decision support model in breast cancer to help diagnose more quickly, using prototype selection associated with (Support Vector Classifier (SVC), K-Nearest Neighbor (KNN), Random Forest (RF), Logistic Regression (LR)). In this way, we expect to obtain robust results with agility and efficiency.
Eliptik eğri uygulaması nesnelerin interneti için kriptografi(IOT) güvenlik
The Internet of Things, or IoT, is essential to the fields of industry, healthcare, and information technology, among others. It is made up of numerous interconnected things that are in communication with one another. Because approved objects, in addition to users, may access data regarding the Internet of Things. For IoT applications and technology to be widely adopted, security is a necessary component. To improve the security of the IoT data, this study suggests an Elliptic Curve Cryptography approach. Two keys are used in Elliptic Curve Cryptography (ECC): a public key and a private key. The user uses the private key for encryption, and the public key is used for user identification during authentication. Similar to this, using the private key, the sender encrypts; if secrecy is desired, using the public key, one can decrypt the communication. Selecting the private key is a problem with every public key. Random selection of small values raises concerns about the overall algorithm's security. considering the values. This study suggests using the Cuckoo Search Algorithm to select values at random.
MFF-LSTM: Çok Ölçekli Özellik Füzyon Tabanlı Uzun Kısa TasarlamaYalan Haber Tespit Sistemi için Farklı Özelliklere Sahip Dönem Belleği
The rise in the usage of media has led to a significant increase in the raise of false information, making it imperative to combat this issue and reduce our reliance on such unreliable sources. Fake news can mislead people, spread rumors, and even impact the positions of political leaders. Detecting fake news has become crucial in this digital era, with direct messaging platforms and social media playing a major role in its proliferation. Various innovative techniques have been suggested to determine fake news, making the endeavor both intriguing and challenging. Hence, this synopsis aims to develop the adaptive learning model with multiscale feature fusion for fake news detection. The proposed system constitutes "text collection, text pre-processing, feature extraction and detection". Initially, the text input is collected from the benchmark datasets, which is then followed by the text stage of pre-processing. Here, the pre-processed text is obtained that is fed into the feature extraction phases. The three feature extraction techniques such as "Bidirectional Encoder Representations from Transformers (BERT), Term Frequency-Inverse Document Frequency (TF-IDF) and GloVe Embedding" are employed to provide the feature set 1, 2 and 3. Finally, these resultant features are given to "Multiscale Feature Fusion based Long Short Term Memory (MFF-LSTM)", where the features are fused together in multiscale manner and detection is taken place by LSTM. Therefore, the system evaluation is done by considering the distinct measures and compared among traditional approaches. Hence, the recommended model attains the desired results to detect the fake news that helps to evade the exploration of any false information.
Makine öğrenmeyi kullanarak COVID-19 aşı görüşlerinin duyarlılık analizi için sınıflandırma algoritmalarının karşılaştırmalı bir çalışması
The task of analyzing textual data and classifying them into positive, negative, or neutral emotions within the domain of natural language processing is a multifaceted undertaking. The primary objective of this study is to employ machine learning algorithms in order to classify opinions pertaining to the coronavirus disease and vaccines. In this study, four algorithms were employed, namely Random Forest (RF), Gradient Boosting Classifier (GBC), Logistic Regression (LR), and Decision Tree (DT). The RF and GBC algorithms demonstrated a commendable accuracy rate of 89%, while the LR and DT algorithms yielded a slightly lower accuracy rate of 87%. The findings derived from this research can provide valuable guidance to policymakers in effectively addressing potential barriers that may impede the successful execution of vaccination campaigns. The analysis of the Kaggle data, which encompasses a wide range of commentaries related to the pandemic and vaccines, underscores the urgent need for prompt measures to attain herd immunity against Covid-19. This imperative objective holds significant importance in effectively managing the transmission of the virus and mitigating its adverse consequences on the well-being of the general population. The task at hand necessitates the acknowledgment and resolution of public apprehensions, as well as the establishment of trust and assurance in the vaccination initiative. This study presents an analysis of the machine learning techniques employed and conducts a comparative evaluation of their significance. The forthcoming research endeavors to create an application that will be capable of categorizing sentiments and opinions pertaining to diseases and vaccines. It is imperative for governments and organizations to comprehend the obstacles linked to the worldwide COVID-19 vaccination endeavor in order to develop efficacious strategies. Nevertheless, it is crucial to acknowledge that the scope of the analysis was restricted to tweets written in the English language. This limitation may potentially undermine the credibility and generalizability of the findings pertaining to overall sentiment. Additional investigation could be conducted to examine more extensive Twitter datasets employing deep learning models in order to gain a deeper comprehension of the general public's attitudes towards COVID-19 vaccines.
5G iletişim sistemi için yeniden yapılandırılabılır akıllı meta yüzey
For 5G mobile communication networks, a novel antenna array is built using a printed monopole based on a metamaterial (MTM). For 5G communication networks, the architecture and functionalities of a Multiple-Input Multiple-Output (MIMO) antenna array are anticipated. Four components make up the MIMO antennae array, which is stacked in an octagon 3D array arrangement consecutively. The components are arranged to create a particular pattern. The array is constructed on a 25 x 25 x 30 mm³ FR4 substrate. Because of its RF characteristics, this substrate option is frequently used in antenna designs. A split ring resonator (SRR) is connected to each antennae element, which is a printed circuit monopole, via a T-resonator (TR). Resonances at 3.6 GHz and 4.8 GHz are provided by this design, making it appropriate for sub-6GHz 5G networks. To lessen mutual coupling, electromagnetic band gap (EBG) spacers are utilised between antenna elements. The antennae array's concerts are improved by the EBG spacers' moonquake fractal-shaped array design. The moonquake fractal EBG spacers are placed to achieve a -20dB mutual coupling reduction between antennae elements. This reduction in mutual coupling is important for maintaining good isolation and minimizing interference between elements. Four PIN diodes are connected to each antennae element to control the antennae's concerts . By using these diodes, different switching scenarios are tested to understand their impact on antennae behaviour. The projected array's
Mülteci ve göçmenlerin etkili tespiti Türkiye'nin çevresinde insanlar evrimsal sinir kullanıyor ağ
In the years following the outbreak of the Syrian Civil War in 2011, non-governmental organizations (NGOs) played a crucial role in assisting refugees and distributing relief to people in need so that
Derin özelliklere sahip enerji verimli bir SDN-ıot mimarisi IoT destekli akıllı şehir için öğrenme tabanlı trafik tahmini
Thesis proposes an energy-efficient SDN-IoT architecture tailored for IoT-enabled smart cities. This architecture addresses the challenges of resource optimization and energy consumption management in the context of diverse IoT devices and dynamic traffic patterns. The architecture provides a framework for efficient network management and facilitates sustainable operation in smart city environments. A modified SVM (Support Vector Machine) algorithm is introduced and integrated into the proposed architecture for traffic prediction. By enhancing the traditional SVM algorithm with specific modifications tailored for IoT-generated data, the thesis contributes to improving the accuracy and reliability of traffic prediction in smart city networks. The modified SVM algorithm can effectively capture complex patterns and variations, enabling precise anticipation of traffic demands. The research presents a comprehensive evaluation of the proposed architecture's performance. Through extensive simulations and practical deployments in a smart city testbed environment, the paper assesses the energy efficiency, resource utilization, and predictive accuracy achieved by the architecture
Koronavirüs (COVİD-19) sürecinde 5g sağlık sistemi
This study's primary goal is to identify knowledge gaps and the most often used information mining techniques in the body of current literature. Novel coronavirus pneumonia has become a significant public health concern, particularly COVID-19. In the academic community, information extraction techniques have been used to uncover concealed knowledge as the possibility of a widespread raises open concerns for wellness. A methodical investigation was carried out using the use of the PubMed, Scopus, and Internet of Science databases. There were 515 million COVID-19 cases and about 6 million deaths worldwide as of May 2022. According to estimates from the World Health Organization, 115,000 healthcare workers died from COVID-19 between January 2020 and May 2021. Patients with impaired immune systems and nano-oncology for the Web of Nano Technologies (IoNT).. There are still issues that need to be clarified and addressed in order to plan clinical execution. These cover the following topics: foundation and breadth; health risks; security and comprehension of information security; artificial intelligence; blockchain; IoT and 5G utilization; authorization; ongoing acknowledgement; and final client education on these technologies.
Dünyadaki siber suçları azaltmak için gelişmiş yapay tabanlı siber güvenlik ağı
In the current digital epoch, cybersecurity emerges as a paramount concern, given the relentless advancement of cyber threats that pose formidable challenges to global digital infrastructures. This thesis embarks on a comprehensive exploration, centering on advanced neural network models, particularly emphasizing the Convolutional Neural Network (CNN), to address the imperative need for resilient cyber defense mechanisms. Employing meticulous experimentation and analysis utilizing the 'Cyber Security Indexes' dataset, this study meticulously evaluates the performance of these models across a spectrum of cyber-attack types. The findings illuminate the CNN model's robustness and efficacy, portraying its potential in fortifying cybersecurity measures and countering the evolving landscape of threats. Throughout this exploration, with a focus on achieving a 96% accuracy threshold, the study outlines the implications of these advancements in fostering a more secure digital landscape on a global scale. The comprehensive insights drawn from this research collectively underscore the pivotal role of the CNN model in fortifying cybersecurity defenses, offering a beacon of hope in mitigating the escalating cyber threats prevalent in today's digital milieu.
Bir roman dört yönlü DNN modeli için siber saldırı tespiti ve önlenmesi
In an era marked by escalating cyber threats, traditional security measures struggle to contend with the surging prevalence of cyber-attacks. To address this challenge, we present a groundbreaking solution in the form of the Quad Directional Recurrent Neural Network (Quad-RNN). This novel architecture, featuring four directions of input and output, amalgamates the strengths of Bidirectional RNNs (BRNNs) and Simple RNNs. Our evaluation, conducted on the NSL-KDD and DDoS datasets, establishes the superiority of Quad-RNN over BRNN and Simple RNN counterparts. Demonstrating enhanced accuracy, precision, recall, and F1 score, the Quad-RNN architecture notably diminishes false positives. This research heralds a pivotal advancement in the realm of cyber-attack detection and prevention, addressing the imperative for resilient and adaptive security measures in the face of evolving threats.
Göğüs röntgeni görüntülerinde çocuk pnömonisinin otomatik olarak tespitinde derin öğrenme ve makine öğrenme yöntemlerinin etkinliğinin değerlendirilmesi
Accurate identification and classification of images of patients with pneumonia is vital for effective diagnosis and treatment. Advanced learning techniques such as CNN LSTM (Convolutional Neural Network Long Short-Term Memory) have proven 97% accuracy on this task. A labeled dataset of 5,856 images of pediatric pneumonia patients from Kaggle was used to evaluate these models. The dataset was divided into training (70%), testing (15%), and validation (15%) sets. It is worth noting that the ResNet 50 -CNN model and the CNN-ExtraTrees model achieved an accuracy rate of 98% and 94%, respectively. These models provide resources for clinicians and radiologists in their decision-making processes. Extracting features from images plays a role in pattern prediction using the optimizer as part of the modeling process. To improve performance, modifications were made to the RMSprop parameters. Model performance evaluation included metrics such as recall, precision, F1 score, and overall accuracy evaluation. These measures provide insight into the effectiveness of models while also serving as reference points for investigations. The exceptional accuracy rates achieved by these models confirm their ability to accurately classify patients with pneumonia, which has implications for improving patient outcomes. The evaluation criteria used in this research provide an examination of the advantages and limitations of models that highlights the importance of incorporating these metrics to ensure accurate performance in scenarios. Keywords: Medical Image, Pediatric Pneumonia, Resnet50-CNN, CNN-Extratrees, Convolutional Neural Network, Long Short-Term Memory
Ekonomik analiz ile petrol rafinerisi iş akışı dönüşümü için modern web tabanlı karar destek sistemi
The refinery industry operates within a dynamic and competitive environment, requiring decision-makers to navigate complex challenges and uncertainties. This project focused on the development of a sophisticated Decision Support System (DSS) using advanced technologies such as NET 6 and C#. The primary objective was to empower decision-makers in the refinery industry with a robust tool to analyze extensive datasets, gain valuable insights, and make informed strategic choices. The project followed a systematic development process, involving the transformation of traditional Excel-based models into an automated application. By harnessing the capabilities of NET 6 and C#, the DSS facilitated seamless integration with backend systems, efficient data processing, and complex calculations. The frontend of the DSS was built using ReactJS, resulting in a user-friendly and visually appealing interface. Decision-makers can effortlessly navigate through the system, visualize data, and analyze key metrics, enabling them to make data-driven decisions with confidence. The implementation of the DSS offers significant advantages to the refinery industry. It reduces dependency on manual processes, automates data analysis, and provides real-time access to critical information. Decision-makers can explore various scenarios, evaluate financial implications, and optimize resource allocation. By streamlining the decision-making process, the DSS saves time, enhances operational efficiency, and empowers decision-makers to stay ahead in the competitive refinery industry. This project vii underscores the transformative potential of advanced technologies in leveraging data for informed decision-making, fostering a competitive edge in the refinery industry landscape. Keywords: Oil, Decision Support System, Web-Based Decision, Economic Analysis, Decision Support System (DSS).
Makine öğrenmesi yöntemleri ile çevrimiçi kredi kartı işlemlerinde Fraud analizi tahmini
Kredi kartları, dünya genelinde yaygın kullanımı ve sağlam altyapısı sayesinde hızla insanların günlük hayatlarına entegre olmuş ve güvenle kullanılan ödeme araçlarından biri haline gelmiştir. Ancak, kredi kartı sayılarının artması ve işlem hacminin hızla büyümesi, dolandırıcıları cezbetmiş ve haksız kazanç elde etme amacıyla çeşitli dolandırıcılık yöntemlerini ortaya çıkarmıştır. Günümüzde kredi kartı bilgilerine ulaşmanın kolaylaşması, kredi kartı dolandırıcılarının faaliyetlerini kolaylaştırmaktadır. Gelişen teknoloji ile hesap hareketleri zaman içinde analiz edilebilmekte ve kötü niyetli verilerin kullanımı izlenebilmektedir. Bu çalışmada, Kaggle veritabanından elde edilen Kredi Kartı Dolandırıcılık Teşhis veri seti kullanılarak bu çalışma, kredi kartı dolandırıcılığı tespiti için topluluk tabanlı XGBoost modeli ile diğer geleneksel makine öğrenmesi modellerini karşılaştırarak önemli bulgular ortaya koymaktadır. XGBoost, Random Forest ve CatBoost modellerinin, kesinlik, geri çağırma gibi performans ölçütleri üzerinde daha iyi performans sergilediği belirtilmektedir. Bu sonuçlar, finansal kurumların günlük operasyonlarında karşılaştıkları dolandırıcılık risklerini azaltma potansiyeline işaret etmektedir. Çalışmanın dikkate değer bir noktası, XGBoost, Random Forest ve CatBoost'un dengesiz veri setleriyle başa çıkma yeteneğinin vurgulanmasıdır. Geleneksel modellerin zayıf performans gösterdiği bu durumlarda, XGBoost, RandomForest ve CatBoost'un %99'a kadar tahmin doğruluğu sağladığı ve diğer modellere göre daha iyi bir performans sunduğu gözlemlenmektedir. Anahtar Kelimeler: Çok Katmanlı Sinir Ağları, Veri Madenciliği, Naive Bayes Yönetmi, Sahtekarlık tespiti.
Şebeke bağlı PV sistemlerde maksimum güç noktası takibi ve aktif/reaktif güç kontrolü için sinir ağlarının tasarlanması ve simülasyonu
This thesis explores the application of Neural Network (NN) in the management of gridconnected Photo Voltaic (PV) system. This study aims to improve the performance and reliability of PV systems using NN capabilities. The controller modelling using MATLAB/Simulink, the first section of the thesis designs and implements a Maximum Power Point Tracker (MPPT) control system based on NN, designed to dynamically adapt to variable solar radiation intensity due to weather conditions. it also raises the voltage from 260 V to 350 V despite fluctuations in radiation intensity. The result obtained indicates that the boost converter effectively raised the PV voltage and maintain it at a certain level while working under the guidance of the NN. The final section discusses active and reactive power control. The algorithm provides local reactive power compensation making it economically viable. We use five different performance scenarios of the proposed control method are tried, and the NN controller shows remarkable flexibility and quickly adapting to fluctuations in load and radiation. The inverter voltage of 230V was equivalent to the mains voltage due to their parallel connection. The Total Harmonic Distortion (THD) of the grid current under all operating conditions was measured at less than 1.86%. In addition, it has been observed that the success rate of NN is more than 99%. This thesis provides insight into the potential of NN in renewable energy applications by using NN, contributing to the development of a more sustainable and stable energy grid, and obtaining an integrated system that can be integrated with the grid.
Gerçek zamanlı hizmetler sunmak için sis-bulut ortamında ıot kaynak kullanımının artırılması
When resources and computers are made available on demand over the internet, this is called fog-cloud computing. This makes it easy to offer people a variety of integrated computing services without being limited by local resources. This means giving people more than just places to store and back up their data and ways to sync their own files. but it also has processing power and a simple software interface that lets the user control when it's connected to the network. This makes things easier by ignoring many details and internal processes. When it comes to the cloud, scheduling algorithms are necessary to provide services that meet goals like high performance, low prices, minimal energy use, and so on. It is an NP-hard problem to come up with scheduling methods that meet more than one of these goals. We introduce some new heuristic scheduling algorithms in this thesis that allow for multi-objective optimization. We then compare their effectiveness with some well-known scheduling algorithms to study how well they work. The first algorithm, the BDA, looks at the Make-span and the period it takes to complete jobs by the due date. Our algorithm takes into account a task's due date by giving the most weight for the job with the earlier due date and send it to the resource that can complete it in the shortest amount of time to achieve the shortest Makespan. Fog Max-Cloud Min is the name of the first part of our method. Ant Colony Optimization is the name of the second part. We looked at how well our suggested algorithm methods met deadlines and how long the system took to make. Compared to the tools we have now. The study results show that our algorithm is better at getting things done with the lowest Makespan and the best deadline satisfaction. In the second algorithm, We develop improvements to the performance and cost (PC) algorithm in order to give more weight Considering the significant expenses involved, reduce energy consumption, and reduce the duration of create a product. In this work, we describe an approach that is a combination of the PCA and GWO techniques. This algorithm is called the Performance and Cost-Gray Wolf Optimization (PC-GWO) algorithm. The results of the test indicate that the PC-GWO The algorithm decreases the average total energy usage by 12.17 percent, 11.5 percent, and 7.19 percent, as well as the Makespan by 16.72 percent, 16.38 percent, and 14.10 percent. When compared to the GWO algorithm, the PCA method, and the PSO algorithm, it also improves the best average resource consumption by 13.2 percent, 12.05%, and 10.9 percent respectively.
Vıt-skınnet: Segmentasyon mekanizması ile etkili cilt kanserinin tespiti için yeni bir vizyon transformatör tabanlı ShufflenetV3 modeli
Melanoma, a deadly type of skin cancer, claims hundreds of lives every year. Skin cancer typically affects areas of skin that are exposed to sunlight on a frequent basis, such as the "legs, face, arms, and neck". By visually examining lesions with pigment on the skin, melanoma can be identified early and treated with a straightforward method of removal of the malignant cells. However, the examination of the skin by the naked eye alone has a restricted and inconsistent accuracy owing to the scarcity of dermatologists. This also leads the patients to undergo multiple biopsies and thus hampers the course of treatment. The automatic identification of lesions from the dermoscopy images involves several obstacles due to the intricate lesion characteristics and as a result of the detection backdrop. There is a dearth of research on major intra-class variations and inter-class similarities of lesion characteristics, and the prior solutions primarily concentrate on employing larger and more complicated systems for detecting the presence of skin cancer with much-enhanced detection accuracy. In order to address various issues with the traditional methods, it is therefore increasingly crucial to create a successful structure for the identification of skin cancer through the use of deep learning approaches. The implemented skin cancer detection and classification model has three crucial procedures to perform. The collection of dermoscopy images, the segmentation process for segmenting the images, and the detection phase are the three main stages of the implemented skin cancer detection system that is put into practice. First, the benchmark images are utilized to provide the images that are needed for the tests. Once the images are gathered, then the segmentation phase is executed. The segmentation step receives the gathered images as its input. Then, the Residual DenseUNet++ (ResDenseUNet++) is used to carry out an effective segmentation. At this point, the resultant segmented image is provided by the developed ResDenseUNet++, which is then considered for the later stage's inputs. Additionally, the segmented image is then given to the feature extraction process by which the gradient filter image and the texture pattern image are obtained. Additionally, the obtained image results are then inputted for the phase of skin cancer detection. During the detection phase, the newly developed ShufflenetV3 based on Vision Transformer (ViT-ShufflenetV3) is implemented and used to effectively recognize skin cancer. In many experimental validations, the recommended skin cancer identification system has ensured a more precise classification outcome regarding skin cancer than other traditional models.
Çok etikeli dengesizlik verileri ile litolojiyi sınıflandırmak için derin öğrenme yöntemlerinin kullanılması
To anticipate and detect lithologies in a variety of surveys, geologists can save operational expenses and uptime by utilising deep learning methodologies and applications. Accurate data processing and scientific research using data gathered in different geological areas are made possible by this. The four lithologies data in the present research were analysed and classified using multi-class imbalance issues and high dimensionality. One of the biggest issues facing modern data analysis is the imbalance in data classification. Particularly when combined with other challenging factors like the existence of overlapping class distributions, and data imbalance can have a significant impact on the accuracy of classification. When there are several classes involved, mutual imbalance relationships between them exacerbate the situation, making its influence more evident. Furthermore, the high dimensionality issue may result in overfitting and increased computational complexity, both of which may impair classification efficiency. Recursive Feature Elimination (RFE) is used to find the most valuable predictive features, while Synthetic Minority Oversampling (SMOTE) is used to resample the data. Using hybrid multi-class DL system unbalanced learning approach is our solution to solving these issues. Finally, by offering precise categorization and quick responses about the interpretation of data collected in many study regions, we think that our innovations might contribute to the advancement of geological research.
IoT platformunda sürücü dikkat dağılmasını algılamak için LSTM ile evrimsel sinir ağının çeşitleriyle optimum özellikli ayar modeli
Nowadays, traffic accidents are caused due to the distracted behaviors of drivers that have been noticed with the emergence of smartphones. Due to distracted drivers, more accidents have been reported in recent years. Therefore, there is a need to recognize whether the driver is in a distracted driving state, so essential alerts can be given to the driver to avoid possible safety risks. For supporting safe driving, several approaches for identifying distraction have been suggested based on specific gaze behavior and driving contexts. Thus, in this paper, a new Internet of Things (IoT)-assisted driver distraction detection model is suggested. Initially, the images from IoT devices are gathered for feature tuning. The set of Convolutional Neural Network (CNN) methods like ResNet, LeNet, VGG 16, AlexNet GoogleNet, Inception-ResNet, DenseNet, Xception, and mobilenet are used, in which the best model is selected using Self Adaptive Grass Fibrous Root Optimization (SA-GFRO) algorithm. The optimal feature tuning CNN model processes the input images for obtaining the optimal features. These optimal features are fed into the Long Short-Term Memory (LSTM) for getting the classified distraction behaviors of the drivers. From the validation of the outcomes, the accuracy of the proposed technique is 95.89%. Accordingly, the accuracy of the existing techniques like SMO-LSTM, PSO-LSTM, JA-LSTM, and GFRO-LSTM is attained as 92.62%, 91.08%, 90.99%, and 89.87%, respectively for dataset 1. Thus, the suggested model achieves better classification accuracy while detecting distracted behaviors of drivers and this model can support the drivers to continue with safe driving habits.
Derin takviyeli öğrenme ile atarı oyunlarını güçlendirmek
Deep Reinforcement Learning (DRL) has made significant strides in the domain of gaming, optimizing artificial intelligence agents to navigate intricate game environments. This study embarks on an exploration of the Snake Optimization Algorithm (SOA) and the Energy Valley Optimization (EVO), it synergizes into a unified approach, aptly named the Energy Serpent Optimizer (ESO), benchmarking their efficacy within a maze-like game setting. Within this environment, an AI agent is set to maneuver through complex pathways, while engaging with diverse challenges such as snakes, skulls, and other animated characters. The overarching objective for the agent is to adeptly navigate the maze, sidestep potential threats, and engage with specific game elements to amass points. A comparative analysis of ESO revealed notable differences in their execution time efficiencies. The SOA emerged as the more time-efficient algorithm, clocking in at a mere 0.43 seconds, This discernible time gap accentuates the superior efficiency of ESO in this particular setting. Additionally, the ESO was put to the test in the said game environment, where the optimization process entailed evolving a set of hyperparameter configurations via genetic algorithms. The goal was to iterate and adapt these configurations to maximize the AI agent's in-game reward, while ensuring the process is time-efficient. Impressively, the AI agent, under the guidance of ESO, achieved a remarkable reward score of 1100.0 in a span of 32 seconds.
Makine öğrenimi aracılığıyla kendi kendine sürüşlü araçlarda şerit tespiti ve direksiyon kontrolünün geliştirilmesi
Auto lane keeping, an increasingly prevalent driver assistance technology in modern vehicles, facilitates the accurate positioning of the vehicle within road lanes, a crucial aspect for subsequent lane deviation and trajectory planning in fully autonomous vehicles. Traditional lane detection methods have historically relied on sophisticated hand-crafted features and heuristics, which, while computationally efficient, face scalability challenges due to the diverse and dynamic nature of road scenes. However, recent advancements in machine learning, particularly with Convolutional Neural Networks (CNNs), have revolutionized this field by replacing hand-crafted feature detectors with deep networks capable of learning pixel-wise lane segmentations. In this thesis, we aim to address the lane detection problem using a variety of methods. To delve deeper, we utilize a dataset comprising highway lane images to conduct a comparative analysis of two distinct methods. Initially, we employ the traditional edge-detection method, featuring hand-crafted features. Subsequently, we explore various Deep Convolutional Network (CNN) architectures tailored to tackle the lane detection challenge. Our investigation culminates in a comparative assessment of these methods, leveraging images derived from their respective outputs.
Çeşitli trafik kaynaklarını kullanan performans bilgisayar ağına dayalı sanal özel ağ sisteminin geliştirilmesi
This study investigates the effects of implementing a Virtual Private Network (VPN on network performance, with a specific emphasis on throughput and time delay. The study entails the manipulation of connection types within three separate protocols, namely HTTP, FTP, and CBR, in order to examine the fluctuations in network performance metrics. The findings suggest that the integration of VPN has a minimal impact on the throughput of the Constant Bit Rate (CBR) protocol, whereas the File Transfer Protocol (FTP) and Hypertext Transfer Protocol (HTTP) protocols exhibit a decrease in throughput. Furthermore, the implementation of the VPN network results in a notable augmentation in the average time delay experienced across all protocols. The aforementioned findings provide significant contributions to the understanding of the intricate correlation between the implementation of VPNs and the performance of computer networks. These insights illuminate the complexities involved in ensuring security while simultaneously considering the potential compromises in network dynamics.
AI tekniği ve IoTs kullanılarak EEG sinyal tanıma ile epilepsi hastalığının tespiti
This research introduces a novel approach for detecting epileptic seizures, leveraging advancements in IoT, moderate signal strength, and advanced deep learning autoencoders. The core aim is to synergize signal function performance with enhanced feature extraction capabilities of a deep learning autoencoder, thereby enabling the technology to identify optimal characteristics more efficiently and swiftly than existing traditional methods. This new method will undergo comparative analysis against various existing computational tools in the same domain. Additionally, it will be benchmarked against established studies in this field. The efficacy of this framework is underscored by its impressive 99.00% accuracy, positioning it favorably among leading research in epilepsy detection and EEG signal classification.
Deep kullanılarak el işaret dilinin sınıflandırılması öğrenme
People with hearing disabilities face many problems, which impede many of their social life issues in all areas of communication, so effective communication is crucial in the development of a nation. It promotes understanding and inclusivity among all members of the community, including those who are deaf. Good communication is key to building and maintaining a strong, cohesive society. I used 29,000 sign language images, each class contained 1,000 images. I built a model from scratch (ASL.model) and compared it with pre-existing models (Xception, Inception, ResNet, VGG16 and MobileNet). The study of intelligent computers that can carry out activities without direct human guidance is known as artificial intelligence (AI), a fast-growing topic within computer science. These tasks may include learning, decision making, and problem solving, and they are often accomplished through the use of algorithms, data, and machine learning techniques. To find the most appropriate classification features to be used for classification, deep learning techniques will be employed in this thesis to create a model for classifying sign language utilizing photographs obtained from the Kaggle depository as a training data set. Deep learning is now widely employed across a variety of industries due to its accuracy and efficiency, particularly for vast yet complicated data, such as photos, sounds, or text, where deep learning algorithms are taught using massive, labeled data sets. In this thesis, we used deep learning to classify a total of twenty-nine sign language-related classes. We put forth a fresh framework for categorizing sign language. The suggested model was put into practice, trained, verified, and tested. The model passed the test with a 99.97% success rate. To assess the effectiveness of our suggested model (ASL.model) with that of these methods, we also employed five pre-trained models (Xception, Inception, ResNet, VGG16, and MobileNet) of assisting the performance of deep learning algorithms that use Convolutional Neural Networks (CNNs). The five pre-trained models had F1-score levels of 100%, 99.51%, 99.87%, 100%, and 99.68%, respectively. The model Xception and VGG16 outperformed all others in terms of testing accuracy but when the testing time was smaller than 1.45 seconds, the suggested model outperformed all others in terms of time.
ML kullanarak gelişmiş sürücü destek sistemleri (ADAS) ile geliştirilmiş otonom bir aracın tasarımı ve uygulanması
This thesis discusses the design and implementation of an autonomous vehicle enhanced with advance driver assistance systems (ADAS) using machine learning. This vehicle can be classified as an educational platform suitable for researchers and specialists in the field of autonomous vehicles. Its structure can be easily modified to meet the needs of researchers, and it can be reprogrammed with ease. The work details the construction of the vehicle, including the chassis structure, suspension system, steering system, brakes, and the anti-lock braking system (ABS). Control of the vehicle is achieved through a mobile phone using a control program developed with MIT App Inventor, allowing wireless Bluetooth communication for driving. The thesis also covers the vehicle's key tasks, such as path planning and navigation using a specialized algorithm for selecting the shortest path to the destination. The vehicle is equipped with ultrasonic sensors distributed around it to detect both stationary and moving obstacles. Additionally, a LIDAR sensor is used for obstacle detection. A machine learning model is created to sense obstacles, trained on data collected from various sensors and scenario, and used to implement autonomous driving in simulation and augmented reality. The results demonstrate the vehicle's ability to navigate obstacles during its journey. Finally, the vehicle can recognize different traffic signs, trained using machine learning on a dataset of over 50,000 samples of 43 classes of German traffic signs. The model is tested for visualization and through the vehicle's camera, enabling it to recognize and respond to all traffic signs appropriately.
Elektrikli araçlarda batarya yönetim sistemlerinin kullanımı
The rise in popularity of electric vehicles (EV) reshaped trend that is clearly boosting the automative market. This increases main three reasons are changes of customer preferences, innovativeness and protective laws from governments. Lately, governments from all around the world are legislating new laws, such as decreased EV taxes, to protect our ecosystem. This kind of situation increased the sales of EVs. Keywords: BMS, Electrical Vehicles, Battery, SOC, SOH, SOT
MXene/MBene yüzey tipi FexB filmlerin emı kalkanlama verimliliklerinin yoğunluk fonksiyon teorisi ve Monte Carlo metodu yaklaşımıyla araştırılması
Bu tez çalışmasında MXenes ailesinin en yeni üyesi olan MBenes kapsamındaki FeXB'un (FeB, Fe2B) yapısal, elektronik, manyetik ve optik özellikleri ab-initio yöntemler kullanılarak incelenmiştir. Ardından bu malzemelerle elde edilen 2D film tabakaların elektromanyetik kalkanlama etkililikleri fotonun polarizasyon durumu gözetilerek detaylı şekilde incelenmiştir. Çalışmada isim karışıklığını önlemek için; I41/amd uzay grubunda bulunan FeB için FeB(1), cmcm uzay grubunda bulunan FeB için FeB(2), I41/mcm uzay grubunda bulunan Fe2B için Fe2B(1), P4/mmm uzay grubunda bulunan Fe2B için Fe2B(2) denilmiştir. Bu 4 malzemenin analizi yapılırken kuantum mekaniksel alt yapıyı kullanan yoğunluk fonksiyon teorisi (DFT) kullanılmıştır. FeXB nanoparçacıklar (FeB(1), FeB(2), Fe2B(1), Fe2B(2)) kullanılarak elde edilen dört farklı tek katmanlı film yapının da ultraviyole bölgede 20 dB'in üzerinde kalkanlama performansı gösterdiği görülmüştür. Elektromanyetik radyasyondan korunmanın en etkili yolu elektriksel iletkenliği yüksek aynı zamanda yüksek manyetik geçirgenliğe sahip kalkanlama malzemesi kullanmaktır. Bor tabanına demir katkılanarak elde edilmiş olan FexB yapıları birçok üstün özelliğinin yani sıra, göstermiş oldukları metalik davranışları ve ortaya koydukları manyetik özellikleri sayesinde yeni nesil kalkanlama malzemesi olmaya da aday olduklarını göstermişlerdir. Tek katmanlı atomik yapılarında bile >20 dB dolaylarındaki kalkanlama performansları heyecan vericidir.
Ana Arap lehçeleri arasında ayrım yapmak için bir sistem tasarlama
Dialect identification is one of the more recent areas of interest for scholars. This study focuses on distinguishing between well-known Arabic dialects through conversations or speeches. What distinguishes this study from others is the way it approaches the topic, as speech was treated as if it were a sound, such as the sound of birds or even music. In other words, the networks were trained on everyday speech without delving into the details of the language, which represents a strong challenge for researchers. With this approach, all the difficulties and challenges facing researchers are overcome. Here, three Arabic speech patterns are taken into consideration: Levantine Dialect (LD), Egyptian Dialect (ED), and Arabian Peninsula Dialect (APD). To distinguish between the three talking styles, we suggest three models for artificial neural networks. The Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and Deep Recurrent Neural Network (DRNN) networks serve as the foundation for these models. This work presents a thorough analysis of the three suggested models, along with comparisons between them. Spoken Arabic Regional Archive (SARA) dataset is employed. It has been prepared and split up into three sections. The Original SARA (OSARA), Filtered SARA (FSARA), and Mixed SARA (MSARA), which combines the OSARA and FSARA, are these. The suggested DRNN model using the MSARA group of the used dataset has the highest accuracy, 90.70%, according to the results.
ITIL'ın müşteri analizindeki etkisi
The Information Technology Infrastructure Library (ITIL) is a widely recognized framework that provides best practices for IT service management (ITSM). Its principles are designed to enhance the quality of IT services, align IT operations with business objectives, and ultimately improve customer satisfaction. The purpose of this thesis is to explore the impact of ITIL principles on customer analysis, focusing on how these principles can be leveraged to better understand and meet customer needs within IT service contexts. The framework encourages the alignment of IT services with customer expectations and business goals, thereby fostering a culture of continuous improvement and service excellence. For instance, the implementation of ITIL processes has been shown to enhance service quality, increase reliability, and improve overall customer satisfaction. This alignment is not merely operational; it requires a strategic approach to understanding customer needs and integrating those insights into service design and delivery. Moreover, the ITIL framework facilitates the standardization of IT services, which can lead to more predictable and efficient service delivery. By adopting ITIL principles, organizations can create structured processes that allow for better incident management and service operation, ultimately leading to enhanced customer experiences. The focus on service operation within ITIL is particularly important, as it is during this phase that customers directly perceive the quality of IT services. The integration of ITIL with other governance frameworks, such as COBIT, further strengthens its impact on customer analysis. This integration allows organizations to align IT governance with customer-centric strategies, ensuring that IT services not only meet internal operational goals but also deliver value to customers. In conclusion, the principles of ITIL provide a robust framework for enhancing customer analysis within IT service management. By focusing on service quality, aligning IT operations with customer expectations, and fostering a culture of continuous improvement, organizations can significantly improve their service delivery and customer satisfaction. This thesis aims to delve deeper into these aspects, providing empirical evidence and case studies to illustrate the transformative impact of ITIL principles on customer analysis. Keywords: ITIL, Ai, Customer services, Software, Management
CRM yazılım sistemlerinde yeni trendler: Yapay zeka ve makine öğrenimi entegrasyonlarının müşteri analitiği üzerindeki etkisi
In this era of paced digitalization significant changes have surfaced. Nowadays companies are not just gathering customer information. Also concentrating on analyzing this data in an efficient manner to improve the overall customer satisfaction. The incorporation of machine learning and artificial intelligence, in CRM systems has revolutionized the way customer analysis is conducted making it a sophisticated and multifaceted procedure. Traditional CRM systems were originally created to store customer information and conduct analyses; however., AI and machine learning powered CRM systems offer sophisticated analytical features, like forecasting customer actions and creating customized marketing plans. These advanced technologies, fundamentally utilizing big data and data mining methods to analyze customer data, offer businesses strategic advantages in competitive markets. The contributions of machine learning and AI to CRM software play a crucial role, especially in customer segmentation, customer lifecycle analysis, and the implementation of upselling and cross-selling strategies. Moreover, these technologies allow for faster and more efficient analysis of customer complaints and feedback, enabling businesses to develop proactive solutions aimed at increasing customer satisfaction. Within the framework of machine learning and AI, the data analysis capacity of CRM software has been enhanced, providing deep insights into customer-business interactions and automating business processes (BPA), particularly in marketing. The primary objective of this study is to examine in detail the impact of AI and machine learning integrations on customer analysis in CRM software. Specifically, the study will explore the role of AI and machine learning in increasing customer loyalty, predicting customer behavior, and personalizing customer interactions. The contribution of these technologies to the development of customer-centric strategies by businesses will be assessed. By analyzing the effects of machine learning and AI on customer segmentation, customer lifecycle management, and data-driven marketing strategies, this study will investigate how these integrations have driven transformative changes in CRM software.
Derin öğrenme kullanarak retinal göz hastalığı tespiti
The global prevalence of retinal abnormalities affects millions of individuals, highlighting the urgency of early detection and intervention to prevent the advancement of these conditions, ultimately mitigating the risk of avoidable blindness. In this thesis, we delve into the critical realm of retinal disease detection using deep learning techniques. Leveraging a diverse ensemble of state-of-the-art neural network architectures, including MobileNetV2, ResNet50, InceptionV3, and DenseNet, we conduct a comprehensive evaluation of their performance in classifying retinal scans. Our meticulous preprocessing steps, resize images, convert grayscale to RGB and rigorous training cycles lay the foundation for an advanced model. Notably, our results reveal that ResNet50 outperforms other models, achieving an accuracy of 0.89, setting a new benchmark in retinal scan analysis. This research contributes to the vital field of early retinal disease detection, offering the potential to enhance clinical diagnosis and patient outcomes
Kablosuz sensör ağında optimize edilmış özellik seçimi ile hibrit makine öğrenmesi tekniğine dayalı gelişmiş anomali tespit sistemi tasarımı
At present, the Wireless Sensor Network (WSN) plays a pivotal role in the wireless communication system that works based on a large number of sensor nodes. The development of the WSN has become more popular and the nature of versatility resulted in more security concerns and making it hard for the investigation to prevent the anomaly in it. One of the essential and challenging tasks in WSN is the security concern. Detecting the anomaly present in the network becomes the major challenge to ensure the security of WSN. In general, WSNs are affected by a different type of threats that tends the nodes to get damaged and form the wrong determination. Therefore, it is necessary to identify anomalous to minimize the false alarm. Moreover, the quality of the data gathered by the sensor nodes is mainly affected by the anomalies that are produced because of different reasons like reading errors, malicious attacks, failures, and unusual events. Hence it is significant to process the anomaly detection to ensure the quality of the sensor data before it is used to make decisions. In WSN, anomaly detection is the significant process to determine the anomaly or unusual event. However, timely anomaly identification is more complex to function to execute reliably in real-time. For the secure and reliable operation in the WSN, effective anomaly detection is more necessary. However, the standard anomaly detection techniques often fail to adequately secure the privacy of the data and identify the complex, particular, and unique breaches in the WSN. Also, the present anomaly detection models only process under the stationary environment and need to keep all the training data in the node. To address these limitations, a novel Hybrid Machine Learning Technique (HMLT) is introduced to effectively detect the presence of anomalies in WSN. The advanced hybrid technique is designed to enhance the detection performance and safeguard privacy. Initially, the required data from the WSN is collected from the available data resource. Further, the significant feature from the raw data is selected using the optimal feature selection process. Here the Secretary Bird Optimization Algorithm (SBOA) is used to achieve the optimal feature selection. Finally, the HMLT is implemented to perform the detection task, in which the hybrid classifier is the combination of the Deep Belief Network (DBN) along with the Bayesian Learning (BL). The model is specifically developed to identify the occurrence of anomalies in WSN using the HMLT for a given dataset. Extensive comparative analysis is performed to analyze the detection capability of the designed approach along with the conventional model. The resulting outcome defines that the proposed approach performs greater in detecting the anomaly than other standard modelsAt present, the Wireless Sensor Network (WSN) plays a pivotal role in the wireless communication system that works based on a large number of sensor nodes. The development of the WSN has become more popular and the nature of versatility resulted in more security concerns and making it hard for the investigation to prevent the anomaly in it. One of the essential and challenging tasks in WSN is the security concern. Detecting the anomaly present in the network becomes the major challenge to ensure the security of WSN. In general, WSNs are affected by a different type of threats that tends the nodes to get damaged and form the wrong determination. Therefore, it is necessary to identify anomalous to minimize the false alarm. Moreover, the quality of the data gathered by the sensor nodes is mainly affected by the anomalies that are produced because of different reasons like reading errors, malicious attacks, failures, and unusual events. Hence it is significant to process the anomaly detection to ensure the quality of the sensor data before it is used to make decisions. In WSN, anomaly detection is the significant process to determine the anomaly or unusual event. However, timely anomaly identification is more complex to function to execute reliably in real-time. For the secure and reliable operation in the WSN, effective anomaly detection is more necessary. However, the standard anomaly detection techniques often fail to adequately secure the privacy of the data and identify the complex, particular, and unique breaches in the WSN. Also, the present anomaly detection models only process under the stationary environment and need to keep all the training data in the node. To address these limitations, a novel Hybrid Machine Learning Technique (HMLT) is introduced to effectively detect the presence of anomalies in WSN. The advanced hybrid technique is designed to enhance the detection performance and safeguard privacy. Initially, the required data from the WSN is collected from the available data resource. Further, the significant feature from the raw data is selected using the optimal feature selection process. Here the Secretary Bird Optimization Algorithm (SBOA) is used to achieve the optimal feature selection. Finally, the HMLT is implemented to perform the detection task, in which the hybrid classifier is the combination of the Deep Belief Network (DBN) along with the Bayesian Learning (BL). The model is specifically developed to identify the occurrence of anomalies in WSN using the HMLT for a given dataset. Extensive comparative analysis is performed to analyze the detection capability of the designed approach along with the conventional model. The resulting outcome defines that the proposed approach performs greater in detecting the anomaly than other standard models
Sağlık 4.0 bütüncül yaklaşım aşı talep tahmin örneği
Bu çalışmada Sağlık 4.0 açıklanarak, mevcut yöntemlerden farklı olarak, aşı kıtlığı ve fazlalığını önlemek için uygun aşı stokunu tahmin etmek amacıyla makine öğrenmesi yöntemleri uygulanmış ve başarıları karşılaştırılmıştır. Aşı talebi, artan nüfus hareketliliği ve salgınların yaygınlığı nedeniyle çeşitli ülkelerde önemli ölçüde artmaktadır. Bu çalışma hem aşı kıtlığını hem de aşırı arzı önlemeyi amaçlayarak en uygun aşı stok seviyelerini tahmin etmek için makine öğrenimi yöntemleri kullanmıştır ve bu tahminlerin etkinliğini karşılaştırmayı hedeflemiştir. Tahmin modellerinde kullanılan veriler, Hudut ve Sahiller Sağlığı Genel Müdürlüğü'nden temin edilmiştir. Bu çalışma, 2003 ile 2023 yılları arasında toplanan 21 yıllık geriye dönük bir veri setini analiz edilmektedir; bu veri seti aylık aşılanma kapsamı verilerini içermektedir. Yıllık aşı talebini tahmin etmek için literatürde yaygın olarak kullanılan dört farklı yöntem uygulanmaktadır. Bunlar arasında en yaygın kullanılan yöntem Otoregresif Entegre Hareketli Ortalama (ARIMA) olduğu tespit edilmektedir. Ayrıca, Mevsimsel Otoregresif Entegre Hareketli Ortalama (SARIMA), Doğrusal Regresyon ve XGBoost modelleri de kullanılmaktadır. COVID-19 pandemisi gibi belirli olaylar, veri setindeki kalıpları bozmuş-tur. Budama testlerinde, ham veri setindeki veri frekansındaki değişiklikler analiz edilmektedir. Modeller, Kök Ortalama Kare Hatası (RMSE) ve Ortalama Mutlak Hata (MAE) kullanılarak değerlendirilmektedir. Tüm veri seti, durağanlık sağlamak için dönüştürülmekte-dir. Mevsimselliğin ve beyaz gürültünün kaldırılmasının ardından modeller yeniden değer-lendirilmektedir. En doğru tahminleri veren modellere çapraz doğrulama uygulanmaktadır. Optimize edilmiş modelden elde edilen tahmin sonuçları, Değer Risk (VaR) modeline girdi olarak kullanılmaktadır. Gerçek, kestirim ve ortalama aşılanma sayıları, SARIMA, Doğrusal Regresyon ve XGBoost tahminlerine dayalı olarak %95 ve %99 güven aralıkları (kritik stok aralığı) ile sunulmaktadır. Aşı tahmin aralığı dengesi nedeniyle, XGBoost'un çıktıları Değer Risk (VaR) modeline girdi olarak alınmakta ve önümüzdeki günlerde ortaya çıkabilecek güvenli aşı stoku ile ilgili maliyet riski değerlendirilmektedir. Çalışma boyunca, modellerin etkili bir şekilde öğrenmeye devam edebileceği koşullar ile bu modellerin seçilme gerekçesi izlenebilmektedir.
Derin öğrenme yaklaşımı ve dörtlü güvenlik katmanlarına dayalı geliştirilmiş bir görüntü steganografi şeması
This thesis intends a fresh approach for the improvement of dataset secreting in images via the ACO algorithm. Digital image steganography requires a balance between two essential goals: this provides the maximum optimization of the concealment of data; it would also reduce the possibility of the image from existence noticed via optimizing the quality of the original image to be concealed. Many current steganographic methods tend to sacrifice the goals above and below at some point. The thesis introduces the "ACO-LSB" approach, which is designed to enhance embedding capacity using a gray-scale shelter image to hold secret dataset through adding an extra bit-pair in byte (b) to make a checksum of the integrity of the image or a check sum of the hidden message. The method encrypts secret information as the pairs of bits and embeds into the uncompressed images in grey scale. The algorithm used in the ACO is the adaptive scanning to find pixel locations and increase the data embedding capacity while decreasing the strong impact on the image quality. Otherwise, the specific pheromone values are changed in a cyclical fashion to avoid problems with stagnation in the context of the overall optimization process – the values should be ideal for proper selection of pixels. The performance results of the ACO-LSB method are outstanding and this research confirms that they enabled enhancement in the subsequent image embodiment, with up to a 30% increase in embedding capacity compared to traditional methods. Technology achieves an average maximum Peak Signal-to-Noise Ratio (PSNR) of (40.5) dB and Structural Similarity Index (SSIM) of (0.98). Furthermore, Methodology shows strong resistance to detection, reducing detection rates by 20%. The model was implemented using MATLAB R2023a and tested on a publicly available dataset of 1000 gray-scale pictures, providing strong evidence of its effectiveness.
Geliştirilmiş bir ağ girişim tespit sistemi için ağ trafik özelliklerinin ana bilgisayar trafik özellikleriyle birleştirilmesi
Network security is a key concern in today's linked world as cyber threats grow more sophisticated and ubiquitous. Traditional Network Intrusion Detection Systems (NIDS) generally fall short owing to their dependence on predetermined signatures and restricted detection scope, exposing substantial gaps in efficiently recognizing new and unanticipated intrusions. This research tackles these difficulties by merging network and host traffic data with sophisticated deep learning algorithms to boost NIDS performance. Utilizing the Network Intrusion Detection dataset, which comprises multiple intrusion scenarios replicated in a military network context, our technique involves painstaking data collection, preprocessing, and feature extraction. We employed a convolutional neural network (CNN) to assess these data, applying rigorous feature selection and dimensionality reduction to enhance model performance. The findings reveal that our deep learning-based NIDS achieves an amazing detection accuracy of 98.5%, exceeding current approaches and successfully resolving real-world cybersecurity problems. This complete approach not only develops NIDS technology but also provides a practical solution for boosting network security across many applications, therefore contributing to the development of intrusion detection systems.
Akilli sayaç veri simülasyonu ve analizi yoluyla libya'nin el khums şehrinde elektrik enerjisi tüketiminin optimizasyonu
The challenges facing Al Khums in the energy sector are significant, including aging infrastructure, frequent power outages, and heavy reliance on fossil fuels, exacerbated by political unrest and unreliable demand management in the energy sector. Accordingly, this study proposes smart meter technology as a model solution to address these challenges, improve electricity consumption, enhance grid reliability, and support renewable energy integration. This is achieved by simulating and analyzing smart meter data using modern techniques such as Python simulation. The research aims to achieve real-time monitoring capabilities, advanced analytics, and demand management strategies to address energy efficiency shortcomings in Al Khums. The research tools utilized Python-based simulations and data extracted from traditional electronic meters to simulate smart meter functions and then compare performance metrics such as load forecasting accuracy, anomaly detection, and energy loss reduction. The main findings concluded that smart meters enable accurate insights into consumption patterns, reduce technical and non-technical losses through rapid anomaly detection, improve load management, and identify peak demand periods associated with weather and seasonal temperatures. Additionally, smart meters facilitate dynamic pricing models and demand response programs, empowering consumers to adjust usage and lowering peak demand. The implementation of smart meters in Al Khums is shown to enhance grid stability, reduce carbon emissions, and align with Libya's renewable energy goals. However, challenges such as high initial costs, cybersecurity risks, and public awareness gaps require targeted interventions. Recommendations include pilot programs, regulatory frameworks for data privacy, and infrastructure upgrades to support bidirectional communication. This study contributes a localized framework for smart grid adoption in regions with similar socio-political and infrastructural constraints, emphasizing the role of data-driven solutions in achieving sustainable energy transitions. The results underscore the viability of smart metering as a cornerstone for modernizing Libya's energy sector, offering actionable strategies for policymakers and utilities to improve efficiency, equity, and environmental outcomes.
Gözetimli öğrenmeye dayalı nesnelerin interneti uygulamaları için girişim tespit sistemi
The Internet of Things (IoT) has evolved dramatically of recent years, and development of commercial and personal applications that can track a person's daily routine, attracting hackers to exploit the vulnerability to steal or modify collected data or disrupt system functions. To prevent hackers from compromising IoT devices, it is important to develop a method that captures and examines network traffic to identify and categorize malicious behavior that attackers may use. Therefore, this work uses a new IoT intrusion detection system designed to determine whether IoT network traffic is normal or potentially anomalous. The algorithm also determines the kind of anomaly if network traffic is deemed to be possibly abnormal. The performance of the proposed system was tested after selecting a dataset called IoTID20, which consists of 625,783 IoT network traffic packets, with 83 distinct features for each packet. The dataset is classified into three classes. The first class (Class I) classifies network traffic into normal or potential anomaly packets. The second class (Class II) classifies each anomalous network traffic into four main classes: 1. Mirai attacks. 2. Denial of service attacks (DoS), 3. Scanning. 4. Man-in-the-middle (MITM). The third class (Class III) classifies the anomalous network traffic into subclasses of the main classes. To identify the class of an IoT network packet, this system goes through four main stages: 1. Feature Preprocessing. 2.Feature selection. 3.Hyperparameter optimization. 4.Classification. In the feature preprocessing stage, the proposed system automatically removes the feature ID column and empty row values and handles useless feature distributions. Therefore, to find salient characteristics to represent each IoT network traffic, the feature extraction step uses correlation coefficient, particle swarm optimization (PSO), and grey wolf optimization (GWO). In this step, features 17, 16, and 22 are selected to represent network traffic for Category 1, Category 2, and Category 3, respectively. In addition, in this stege, four features are selected because they are the common features chosen from all the feature selection algorithms that make our system consume limited computing resources. In the penultimate stage, the Decision Tree (DT) hyperparameter is determined utilizing a Coronavirus Herd Immunity Optimizer (CHIO) to boost the effectiveness of the suggested system. Finally, IoT network traffic is classified using a decision tree algorithm through three phases. In the first phase, it is classified as normal or anomalous. In the case of an anomaly, the system identified the main and subcategories of the anomaly in the second and third phases. Many machine-learning techniques are used to train and test our proposed system; however, the decision tree outperforms other machine-learning algorithms, achieving 99.96%, 99.56%, and 77.6% accuracy for Phases 1, 2, and 3.
Yüksek hızlı görünür ışık iletişimleri için filtreli-OFDM sistemlerinin karmaşıklığının azaltılması
Contemporary wireless communication infrastructures face inherent limitations in radio frequency (RF) bandwidth allocation. As global requirements for enhanced data throughput intensify, systemic challenges, such as spectral congestion and diminished data transmission efficiency, assume critical importance. Within the framework of emerging 5G and 6G architectures, scholarly and industrial focus has pivoted toward leveraging the optical spectrum as a viable alternative to circumvent these constraints. Notably, visible light communication (VLC) systems have recently emerged as an innovative paradigm for augmenting traditional RF-dependent wireless networks. Among these advancements, the Filtered-Orthogonal Frequency Division Multiplexing (F-OFDM) modulation scheme has positioned itself as a promising solution for mitigating OFDM's inherent limitations of excessive out-of-band radiation and consequent spectral inefficiency. In direct detection and intensity modulation (DD/IM), LEDs require real-valued and non-negative signals, achieved through the inverse fast Fourier transform (IFFT) with Hermitian symmetry (HS). While this enables noncomplex-valued signals, it halves the electrical bandwidth. Larger IFFT/FFT sizes increase processing demands, chip area, energy consumption, and hardware costs. The HS constraint complicates the design by requiring duplicate FFT/IFFT blocks, raising computational complexity. Additionally, increasing subcarriers worsens the peak-to-average power ratio (PAPR), adding further challenges. Legacy optical OFDM methods employed to transform bipolar OFDM signals into unipolar formats include DC-biased (DCO-OFDM) and asymmetrically clipped (ACO-OFDM) techniques. However, these adaptations entail compromises: DCO-OFDM elevates power requirements, whereas ACO-OFDM diminishes spectral efficiency. This thesis primarily concentrates on optimizing spectral efficiency and reducing the complexity of VLC systems. Therefore, the effort offered in this thesis introduces a different methodology to eliminating the reliance on HS. Two methods, NHS-DCO-F-OFDM and NHS-Flip-F-OFDM, are proposed to generate real F-OFDM signals by rearranging the real and imaginary components of complex F-OFDM signals in the time domain. These methods reduce system complexity by 50%, avoid PAPR issues, and enhance power efficiency contrasted to established HS-based F-OFDM. The NHS-filtered OFDM approach also improves bandwidth efficacy and cuts out-of-band emission power by 115 dB, offering better spectral localization than conventional OFDM.
Hafif şifreleme entegrasyonu Güvenlik için bir izleme ve değerlendirme çerçevesi IoT teknolojisiyle desteklenen verimli bir bulut tabanlı sistem
In the present generation, security issues are a major issue experienced by everyone due to the increase in digital technologies. With the introduction of digital technologies, data is stored and transmitted through platforms where security issues arise. Many cryptographic techniques are available to protect data in transmission and storage through digital technologies to overcome security issues. The methods are utilized to encrypt the data with passwords or PIN codes, which hackers can easily identify. In this case, for a safe process, a hybrid encryption method is proposed in this thesis for security data purposes, integrated with lightweight and hidden ciphertext policy attributes with the assistance of the cloud IoT environment. The lightweight encryption provides proxy services for data authentication. The analysis of the security has provided a positive impact on the encryption of the integrated lightweight. The system offers better positive results with hybrid encryption for data authentication for security purposes. The third part of the framework introduces a hybrid encryption system that combines a hidden hierarchical policy with a lightweight Ciphertext Policy Attribute-Based Encryption (CP-ABE) enforcement model. This is to ensure fine grained access control, minimal computational burden, and efficient encryption-decryption suited for IoT devices. It also incorporates QGE-HMAC for secure message authentication and ELP-DSA for signature validation, providing strong protection for encryption key integrity and end-to-end data authenticity. The final part presents a reliable cloud-enabled IoT network monitoring system powered by JPAO-GRU. This advanced gated recurrent unit model integrates Aranda-Ordaz activation and Jeffreys prior regularization for accurate anomaly detection. Additionally, the framework includes HTT-Fuzzy logic for outdated device detection, QASOA for intelligent load balancing, and LCA-BST for low-latency, Hierarchical device monitoring. Comprehensive experimental validation shows that the The proposed dual approach enhances security, integrity, and network efficiency, achieving high accuracy (99.12%) in intrusion detection, 99.18% encryption security, and substantial reductions in encryption time, latency, and energy consumption. This research contributes a A unified and scalable solution for addressing the growing security, authentication, and trust challenges in next-generation cloud-IoT systems.
Kara, deniz altı ve serbest uzayda kullanılan optik haberleşme sistemlerinin analizi, entegrasyonu ve optimizasyonu
The history of optical communication can be traced back to simple signalling techniques like smoke and flame beacons used by ancient cultures to transmit messages over long distances. Significant advancements occurred in the 18th and 19th centuries with the invention and widespread use of semaphore systems and later optical telegraphs, which enhanced visual signalling capabilities. These systems laid the foundation for structured long-range communication. However, practical low-attenuation optical fibers and lasers were developed in the 20th century before optical communication could become an efficient means for high-speed, long-distance data transmission. Modern telecommunication infrastructures worldwide increasingly rely on extremely long optical fiber cables due to several key advantages. These include the high bandwidth, low signal loss, and overall transmission efficiency of fiber optics. Additionally, fiber optics offer immunity to electromagnetic interference and other forms of crosstalk, such as radiation crosstalk and long-distance crosstalk. Furthermore, they provide high security, data security, and data integrity. These advantages make fiber optics suitable for both urban and remote deployment. The study employs the OptiSystem 18 simulation platform to comprehensively assess the performance of optical communication systems under three distinct transmission environments: fiber, underwater, and free-space optical channels. An extremely critical system parameter, laser power, operational wavelength, modulation schemes, pulse bit patterns, and transmission frequency, were systematically varied one at a time to observe their individual effects and the combined effects on system performance. The modelling was conducted with utmost precision. The simulations have offered valuable insights into optimal configurations for maximizing signal integrity, minimizing attenuation, and enhancing overall transmission quality under various environmental conditions. The results obtained have significantly contributed to a comprehensive understanding of the performance of optical links under various operating conditions. This knowledge has, in turn, led to the generation of valuable recommendations for enhancing systems. The acquired insights will be instrumental in designing robust, user-friendly, and low-power networks in critical sectors and numerous applications and locations.
Derin öğrenme tekniklerinin hibrit yöntemini kullanarak farklı saldırıları tespit etmek için SDN ortamına yönelik siber güvenlik sistemi
As contemporary networks come under attack with sophisticated and mass-scale cyber-threats, traditional intrusion detection systems (IDS) are unable to cope with changing threats. Such limitations are the focus of this research, and two new deep learning-based frameworks are introduced, which are best suited for fast and effective intrusion detection in Software-Defined Network (SDN) environments: (1) the Adversarial Learning-based Multi-Branched Hybrid Architecture and (2) the Multi-Branched Hybrid Perceptron Network (MBHPN). Together, these models offer an intelligent and scalable IDS platform that can detect advanced threats, in particular Distributed Denial of Service (DDoS) attacks, through adaptive learning, real-time integration, and context analysis. The first architecture, as explained in Chapter 3, combines Convolutional Neural Networks (CNN), Capsule Networks, and Long Short-Term Memory (LSTM) layers in a multi-branch architecture. CNNs encode spatial dependencies, Capsule Networks preserve hierarchical relations among features, and LSTMs describe sequential behaviors of traffic flows. The framework also incorporates Dynamic Adversarial Learning, where adversarial samples are produced to mimic attacks, thus enhancing model robustness. Each branch produces a unique feature representation, which is combined using attention-weighted methods to create an integrated, discriminative feature space tailored for intrusion detection. In Chapter 4, this architecture is enriched by the addition of MBHPN, which is directly optimized for DDoS attack detection. This network combines three deep learning branches: MLP, DenseNet-like, and ResNet-like units. It adds Dynamic Feature Adaptation (DFA) to down-regulate noisy features and up-regulate pertinent traffic signatures. Multi-instance Learning (MIL) to process aggregated traffic flows instead of standalone instances, and thus greatly enhances contextual perception. These improvements make MBHPN robust against class imbalance, evasion attacks, and changing attack patterns prevalent in actual networks. The models are trained and tested on three intrusion detection benchmark datasets: UNSW-NB15, CICIDS2017, and CSE-CIC-IDS2018, which contain varied traffic patterns and attack conditions. Proven through extensive experimentation, the models outperform the current state of affairs. On the UNSW-NB15 dataset, MBHPN reports 99.31% accuracy, 98.02% precision, 98.87% recall, and an F1-score of 98.44% with a false positive ratio (FPR) decreased to 0.61% from the original 1.12% FPR of the base model. The time taken for inference reduced from 7.8 ms to 5.3 ms, which reflects a 32% boost in speed from SDNetc integration.
Makine öğrenmesi destekli 2D malzemelerin karakterizasyonu
Bu tez çalışmasında, Atomik Kuvvet Mikroskobu (AKM) ile elde edilen ardışık görüntüler kullanılarak yüzeydeki konum kaymalarının yön ve hız bakımından analiz edilmesi amaçlanmıştır. Bu doğrultuda hem klasik makine öğrenmesi algoritmaları hem de derin öğrenme yaklaşımları kullanılarak yön tahmini ve kayma hızı tahmini yapılmıştır. İlk aşamada, X ve Y eksenlerindeki mikrometre cinsinden kayma miktarları giriş değişkeni olarak kullanılarak yön sınıflandırması gerçekleştirilmiştir. Bu sınıflandırma için Random Forest ve Destek Vektör Makineleri (SVM) algoritmaları uygulanmıştır. Random Forest modeli %98.6 test doğruluğu ile yüksek bir başarı gösterirken, SVM algoritması %94.6 doğruluk sağlamıştır. Random Forest modeli, sınıf bazında da daha dengeli F1-score sonuçları vermiştir. İkinci aşamada ise, ardışık görüntü çiftleri arasındaki zaman farkı ve konum değişimleri kullanılarak kayma hızları hesaplanmış, bu veriler zaman serisi analizi yapabilen LSTM (Uzun Kısa Süreli Bellek) ağı ile modellenmiştir. Model mimarisi iki LSTM katmanından oluşmakta olup sırasıyla 64 ve 32 nöron içermekte ve her iki katmanda da tanh aktivasyon fonksiyonu kullanılmaktadır. Çıkış katmanı ise tek nöronlu yoğun (dense) yapıdadır. Farklı sayılarda eğitim veri çifti ile yapılan denemelerde modelin performansı sistematik olarak değerlendirilmiş ve her konfigürasyonda model 50 epok boyunca eğitilmiştir. Modelin tahmin başarısı; Ortalama Kare Hata (MSE), Determinasyon Katsayısı (R²), Kök Ortalama Kare Hata (RMSE) ve Ortalama Mutlak Hata (MAE) metrikleri üzerinden analiz edilmiştir. Test verisi üzerinde elde edilen sonuçlar, modelin düşük hata oranlarıyla tahminler yapabildiğini ve R² değerinin yüksek seviyelerde olduğunu göstermektedir. Bu da LSTM mimarisinin, zaman bağımlı mikroskobik verilerle çalışırken kayma hızlarını başarıyla tahmin edebildiğini göstermektedir.
Yapay zeka teknolojilerine dayanan 5G ile nitelikler arasındaki saldırı tespitine yönelik siber güvenlik önlemleri
With the emerging cybersecurity arena, especially in the context of next-generation communication networks like 5G networks, precise identification of Distributed Denial of Service (DDoS) attacks is an urgent challenge. The large dimensionality, high-speed data transmission, and high network diversity typical of 5G environments have a propensity to make conventional machine learning models inadequate. To address these problems, this research suggests a new deep learning approach that uses symmetry to help detect unusual activities in today's fast and complex networks. The system architecture under examination is a Tree Convolutional Neural Network (Tree-CNN) that is particularly capable of understanding hierarchical and symmetrical interdependencies among network traffic, prevalent in 5G communications, as their architecture is distributed and layered. Supporting it is a deep autoencoder module that is employed to be capable of extracting latent symmetrical patterns, noise reduction, and improving the discriminative representation of anomaly behaviour. Model learning performance is further enhanced by the addition of a leader-influenced velocity-based spiral optimization algorithm, a novel metaheuristic, which achieves an effective exploration-exploitation trade-off. The aim here is to optimize Tree-CNN parameters, deep autoencoders, and classification thresholds, which results in improved detection accuracy but at the cost of computational practicality. In the scenario of 5G networks—where there is a sense of urgency in real-time processing and adaptive threat response—the necessity for an accuracy-speed trade-off arises. Performance measurements were performed on three benchmark datasets: UNSW-NB15, CIC-IDS 2017, and CIC-IDS 2018, which represent various traffic patterns as one would see in 5G networks, such as bursty rates of traffic and multi-modal inputs. The novel framework achieved exceptional accuracy levels of 96.02% in UNSW-NB15, 99.99% in CIC-IDS 2017, and 99.96% in CIC-IDS 2018, along with nearly perfect precision and recall rates. These results justify the capability of the system to well-detect threats with very low false negatives and positives. While design imposes a medium computational overhead, this is offset by a dramatic enhancement in detection resilience as well as scalability. What is important here is that the model demonstrates its effectiveness in 5G-supported infrastructures, where high-data-rate streams, edge processing, and latency-prone services require high-performance but guaranteed security mechanisms. The proposed symmetry-aware hybrid detection model not only promotes state-of-the-art future 5G networks. It emphasizes the significance of symmetrical pattern detection, hierarchical feature learning, and adaptive optimization as essential components in the construction of next-generation smart security systems.