Elektronik ve Bilgisayar Mühendisliği Anabilim Dalı

ECG arrythmia classification using Deep Neural Network

For assessing heart electrical signals, the ECG is the most attractive and low-cost diagnostic technique used in healthcare institutions. Arrhythmia is a medical term for aberrant cardiac impulses. Cardiac arrhythmia is hazardous and, in most cases, fatal. Arrhythmias come in a variety of forms, and an ECG test detect it. Electrocardiogram analysis has received great scientific value in the medical field for the classification of different heart arrhythmias. Patients benefit from automated systems that may be tailored to test for arrhythmia categorization, and doctors benefit from them as well. Medical professionals' diagnostic errors are one of the primary causes of death. Humans can now automate patient prognosis and disease deduction from report analysis thanks to technological advancements. Data mining and artificial intelligence approaches are extremely valuable in the creation of expert systems for disease identification. CNN and LSTM are a very helpful algorithm used for various disease identification systems so we will use CNN and LSTM for the detection of ECG signal. Our proposed CNN model achieving the maximum accuracy consists of 13 layers and the LSTM model consists of just five layers but with 512 filters through which will be trained on data than we tested it in different samples. Every ECG signal is preprocessed to eliminate the baseline before being segmented using a straightforward approach. The suggested approach obtained improved precision for ECG classification in simulation experiments using the MIT-BIH benchmark database. When it results are compared to previously state-of-the-art approaches, our model give better results. Keywords: Machine Learning (ML), Deep Learning, Convolutional Neural Networks (CNN), Long-Short Term Memory (LSTM) Deep Neural Networks, ECG Signals, Computer-Aided Diagnosis Systems.

Prediction of students' success employing data mining algorithms

Most higher education institutions now priorities student achievement as a key strategic goal. Academic institutions are focusing more on keeping students enrolled in their classes without sacrificing educational level in response to budget constraints and rising operational costs. The use of machine learning to forecast students' success using academic and behavioural data has been the subject of numerous initiatives and studies. We applied the popular machine learning algorithms in order to predict student achievement including Random Forest, Support Vector Machine, Logistic Regression etc. The employed Support Vector Machine has achieved to provide the best accuracy (93.8%) and sensitivity (98.75 %) scores. Additionaly, the implementation of stack-based ensemble model is very effective in the predicition of student success. Therefore, a stacked-based ensemble learning model is cerated in order to compare the efficiency of it in the prediction of students' succes with the other machine learning algorithms. At the end, the stacked-based ensemble model has provided the best accuracy score (93.9%) in the prediction of students' success.

Classification of skin cancer with deep transfer learning method

Skin cancer is a significant threat to human health. The key solution to completely treating skin cancer patients is early detection of lesions, which are the basis of skin cancer. With the advancement of artificial intelligence technology, significant progress has been made in developing automated approaches to help dermatologists detect skin cancer. In this study, eight different transfer learning networks are examined on the International Skin Imaging Collaboration (ISIC) dataset (64000 images) for the classification of skin cancer. The models used in the study are DenseNet121, Xception, InceptionResNetV2, ResNet50, Inception, EfficentNetB6, VGG16, and MobileNetV2 which were successful in various studies recently. In the preprocessing part of the study, the dataset is split into three parts training, validation, and testing. The results of the study showed that some transfer learning models are appropriate for classification success, with high classification accuracies. With an accuracy rate of 99.6%, the finetuned DenseNet121 model outperformed all other pre-trained models used in the study. In addition, when the ensemble learning method was applied with the best three models (DenseNet121, MobileNetV2, and Xception), 99.96% accuracy was obtained.

Deep ensemble learning-based classification of stroke

Stroke is one sort of brain disease that profoundly impairs people's quality of life and health. The quantitative analysis of brain Magnetic Resonance (MR) images is crucial for both the diagnosis and treatment of strokes. The method of early diagnosis is crucial for preventing stroke instances. Deep neural networks, which have the capacity for vast data learning, enable stroke prediction. Therefore, in this study, several deep neural network models are proposed for transfer learning to classify MRI images into two categories (stroke and non-stroke), in order to study the characteristics of the stroke lesions and achieve full intelligent automatic detection. These models include MobileNet, EfficientNetB2, ResNet50, DenseNet121, and EfficientNetB2. 1901 training images, 475 validation images and 250 testing images make up the study dataset. Data augmentation was employed to increase the number of images on the training and validation sets, which helped the models to learn more effectively. Results from the experiment outperform those from all state of the art methods that used the same dataset. The top models of the study, which use the DenseNet121 and Xception models for transfer learning, obtained an overall accuracy of 98.4% with the same values for precision, recall, and F1-score. Additionally, ensemble learning method is used with the top three models of the study, EfficientNet, DenseNet, and Xception, and a 100% overall score is obtained.

Computer network traffic classification using data mining

In this work, we explore different data sets for Network Intrusion detection. Classify the attack signals from normal ones. With the increase in network connections, computer security and intrusion detection became a challenging problem. Our research used different feature selection techniques: Information gain, Gain ratio, Chi2, and Relief. We train different models based on ten thresholds with a quantization range of 10% to 90% for each feature selection algorithm we used. We train Decision trees, Random Forest, Extra trees, GradiantBoosing, XGBoost, AdaBoostm, Linear, SVM, and Logistic Regression. We use cross-validation and feature selection with different thresholds for each of these models. Then, we compare all of these models based on the validation accuracy obtained from cross-validation. The stacking model achieves 97.79% test accuracy on the UNSW-NB15 data set using only 15 features with Random Forest, Extra Trees, and XGBoost as base models (level-0 models) and Logistic Regression as a meta-classifier model. By changing the top model to MLP with 5 hidden layers we achieve 97.82% test accuracy, but the complexity of the stacking model is increased too much to improve the accuracy with only 0.03% this note should be considered when choosing the model for deployment which will affect the inference time of the stacking model.

Design of a new biometric system based on hand geometry images using deep learning methods

Dorsal hand vein patterns are unique to each individual, making them a promising tool for accurate identity recognition. This thesis focuses on biometric analysis of human identification. The tecnocampus hand image dataset is used to validate feature extraction methods on dorsal hand images. The study compares and analyzes three Convolutional Neural Network CNN architectures: ResNet-50, GoogleNet, and SqueezeNet, which offer distinct approaches to feature extraction. It evaluates the performance of each method by applying them to a dataset. Based on the analysis, the most effective features are selected from the respective architectures. By leveraging the robustness of assessed CNN architectures and harnessing the power of Euclidean distance, the proposed approach achieves remarkable outcomes in the realm of image recognition. It uses feature vectors to measure similarity between dorsal images, and uses Euclidean distance to measure the similarity between feature vectors. This allows the algorithm to accurately identify and match hand vein patterns with high accuracy. With an overall accuracy of 97.32%, this research has significant implications for the field of biometric identification and could serve as a valuable foundation for future advancements in this area.

Ağ yapılı veriler için öznitelik tabanlı özellik öğrenimi

Ağlar, birçok gerçek hayat senaryosunda veriler (düğümler) arasındaki etkileşimleri (bağlantıları) yakalama yeteneklerinden dolayı birçok sistemde yaygın olarak kullanılmaktadır. Ağlar, makine öğreniminde farklı tahmin tabanlı analizler yapmak için kullanılabilir. Ancak mevcut makine öğrenimi algoritmaları, veri örneklerinin birbirinden bağımsız olduğu ve ağ yapılı verilerin düzensiz ve değişken boyutlu doğası nedeniyle ağ bilgilerinin makine öğrenimine dahil edilmesini engellediği yönünde temel bir varsayıma sahiptir. Son zamanlarda, ağdaki yapısal bilgiyi koruyarak her düğüm için düşük boyutlu bir vektör temsilini öğrenmeyi amaçlayan özellik öğrenme yöntemleri önerilmiştir. Bu teknikler yalnızca ağın topoloji bilgisinden faydalanır ve düğüm özniteliklerini dikkate almaz. Ancak gerçek dünya ağları genellikle çok sayıda düğüme ve yüksek boyutlu düğüm özniteliklerine sahiptir. Bu çalışmada, topoloji bilgisinin yanı sıra düğüm özniteliklerinden faydalanmak için rasgele yürüyüş prensibini benimseyen node2vec yönteminin nasıl genişletilebileceği gösterilmiştir. Node2vec'teki bir rasgele yürüyüş, düğüm kimlikleri dizisi olarak temsil edilir, öznitelikleri kullanmak için düğüm kimlikleri dizisi yerine özellik vektörleri dizisi kullanılır. Tek kategorik özniteliğe sahip ağlar için node2vec-öznitelik, çok nitelikli ağlar için ise node2vec-öznitelik+ modelleri önerilmiştir. Önerilen yaklaşımların yapısal rollerin benzerliği ve düğüm sınıflandırması üzerindeki etkinliğini değerlendirmek için gerçek dünya veri setleri üzerinde çeşitli testler gerçekleştirilmiştir. Gerçekleştirilen tüm deneysel çalışmalar, her bir algoritmanın etkinliğini kanıtlamaktadır. Genel olarak sunulan sonuçlar, önerilen modellerin ağ ile ilişkili düğüm içerik bilgilerini yakalama ve kodlama konusundaki etkinliğini ve sağlamlığını ortaya koymaktadır.

Detection of faulty solar panels using artificial intellegence methods

Solar energy is gaining popularity as a renewable energy source, but maintaining the efficiency of solar panels poses challenges due to potential defects that can lead to significant energy production losses. The objective of this thesis is to develop an accurate and robust classification model using deep learning techniques to detect faulty solar panels. The proposed method relies on Convolutional Neural Networks (CNNs) and transfer learning techniques to achieve high accuracy, reaching 93% in accurately classifying faulty solar panels using the proposed CNN model. The transfer learning architectures are DenseNet121, ResNet50, MobileNetV2, and Xception models, were implemented with accuracies of 81.66%, 51.21%, 77.47%, and 90.31% respectively. This study demonstrates the effectiveness of employing deep learning techniques alongside suitable data preprocessing methods to develop the classification of defective solar panels.

Lise giriş sınavında öğrenci başarılarının makine öğrenmesi teknikleri ile analizi

Son yıllarda Eğitsel Veri Madenciliği çalışmalarında Makine Öğrenmesi algoritmaları oldukça yaygın biçimde kullanılmaya başlanmıştır. Bu çalışmada öğrencilerin ders başarı notları ile bireysel ve çevresel özelliklerini içeren veri seti kullanarak LGS sınavındaki başarı tahminlemesi yapılmıştır. Çalışma aynı zamanda tek tabanlı ve topluluk (Ensemble) algoritmalarının performanslarını karşılaştırılmak suretiyle en iyi modeli de önermektedir. Oluşturulan modelleri kıyaslamada R-Squared değeri referans kabul edilirken MSE, RMSE, MAE ve MAPE metriklerine de bakılmıştır. Çalışmada aynı zamanda Feature Selection metodlarının da model performansına etkileri araştırılmıştır. Elde edilen sonuçlarda Feature Selection uygulanmadan tüm veri seti kullanılarak Stacking mimaride Super Learner model ile en iyi performans elde edilmiştir. Modelin R-squared değeri 0,79 olurken MSE 161,32 olarak hesaplanmıştır. Çalışmanın sonuçları çalışılan modelin gelişime açık bir model olduğunu ortaya koymaktadır. Öte yandan öğrencilerin öğrenim ve kişisel özelliklerini içeren veriler kullanılarak merkezi sınavlardaki başarı düzeylerinin modellenebileceğini de göstermektedir. Kullanılan bu teknik ilerleyen dönemlerde öğrencilerin güçlü ve zayıf yönlerinin belirlenmesinde ve öğrencilerin başarısının arttırılmasında önemli bir yere sahip olabilir.

Investigation of financial applications with blockchain technology

This thesis focusing on utilizing machine learning (ML) and deep learning (DL) methodologies for forecasting cryptocurrency prices. The research concentrates on interpreting historical price data of prime cryptocurrencies, including Bitcoin, Ethereum, Dogecoin, USD Coin, Binance Coin, and Cardano. These virtual currencies demonstrate unique and unpredictable behavior; therefore, the need to understand their price dynamics is vital for making informed financial decisions. The essential objective of this study is to utilize various ML and DL techniques to construct tailored prediction models for each of these cryptocurrencies, aiding investors, traders, and financial institutions in making more precise and lucrative decisions in the rapidly evolving cryptocurrency market. The central tenet of this study is the use of diverse ML and DL algorithms to develop specialized prediction models for each cryptocurrency. Several techniques are deployed, such as LASSO (Least Absolute Shrinkage and Selection Operator), Linear Regression, Ridge Regression, Decision Tree, AstroML, Convolutional Neural Networks, Support Vector Machines, K-Nearest Neighbors, LSM (Least Squares Method), XGBoost Regression, and Gaussian Process Regression. These techniques were chosen based on their ability to handle complex non-linear relationships, missing values, and large data volumes. The approach was to model each cryptocurrency individually, recognizing that each might display unique characteristics and behaviors influenced by different external factors. The efficacy of these prediction models was evaluated using robust performance metrics like root mean square error (RMSE) and R-square. RMSE measures the average magnitude of the prediction error, providing an understanding of how accurately the model forecasts the prices, while R-square represents the proportion of the variance for the dependent variable that's explained by the independent variables in the model. These metrics allowed for a comprehensive and rigorous evaluation of the prediction models. The findings of the research provide a wealth of insights for various stakeholders in the cryptocurrency market. By demonstrating the effectiveness of ML and DL algorithms in predicting cryptocurrency prices, the research offers investors, traders, and financial institutions valuable tools to aid their decision-making processes. Furthermore, by comparing the performance of various models, it provides insights into which ML techniques are most suitable for cryptocurrency price prediction. Of particular interest is the outstanding performance of the Gaussian Process Regression in predicting cryptocurrency prices. This technique, known for its capability to handle a large number of predictors and complex relationships among variables, outperformed the other models tested. This highlights the potential of Gaussian Process Regression in predicting cryptocurrency prices, providing a promising direction for further research in the field.

Detection of phishing urls with deep learning based on the GAN-CNN-LSTM network and swarm intelligence algorithms

Phishing attacks are one of the challenges of the Internet and its users. Phishing attacks are an example of social engineering attacks based on deceiving users. In phishing attacks, fake pages that are very similar to legitimate pages are created on the Internet. In phishing attacks, the victim is directed to fake pages, and their valuable information is stolen. Most of the targets of phishing attacks include online payment services, banking, and online sales, so the losses of these attacks are significant. One way to detect phishing attacks is to use machine learning and deep learning methods. The challenge of machine learning and deep learning methods is intelligent feature selection. The lack of feature extraction and intelligent feature selection reduces the accuracy of learning methods in detecting phishing attacks. This paper presents a combined method with deep learning, machine learning, and swarm intelligence algorithms to detect phishing attacks. In the first phase, the dataset is balanced by deep learning based on the GAN network. In the second step, the convolutional neural network extracts the primary features from the links and code of web pages. In the third step, the White Shark Optimizer (WSO) algorithm selects the essential features. In the last step, the LSTM neural network classifies the samples. The proposed method has been evaluated on ISCX-URL-2016 and Phishtank datasets for feature extraction and selection. The proposed method's accuracy, precision, and sensitivity in the ISCX-URL-2016 data set are 97.94%, 97.82%, and 97.76%, respectively. In the Phishtank dataset, the proposed method has accuracy, precision, and sensitivity of 96.78%, 95.67%, and 95.71%. The proposed method is more accurate than LSTM, CNN, CNN-LSTM, CNN+GA, DNN, VAE-DNN, and AE-DNN methods in detecting phishing.

Makine öğrenme yöntemleriyle öğrencilerin akademik performanslarının farklı öznitelik seçim teknikleri uygulanarak sınıflandırılması

Öğrenci ve eğitimci açısından akademik başarının artırılması her zaman istenilen durumdur. Akademik başarıyı etkileyen etmenlerin önceden bilinmesi çok önemlidir. Çalışmada, farklı öznitelik seçim teknikleri ve makine öğrenmesi algoritmaları kullanarak ortaokul öğrencilerinin akademik performanslarını sınıflandırmak amaçlanmıştır. Kullanılan veri seti, MEB'e bağlı bir ortaokuldaki 728 öğrenciden elde edilmiştir ve 24 öznitelik içermektedir. Elde edilen veri seti üzerinde ön işleme ve öznitelik seçim teknikleri uygulanmıştır. Veri ön işleme aşamasında, veri seti içerisinde yer alan kategorik değerler sayısal değerlere dönüştürülmüştür. Kullanılan veri setinin dengesiz sınıf dağılımı olduğu için SMOTE algoritması kullanılarak dengeli hale getirilmiştir. Temel veri seti ve 7 farklı öznitelik seçim tekniği kullanılarak toplam 8 veri seti elde edilmiştir. Bu veri setleri Hold out ve 5 kat çapraz doğrulama yöntemleri ile bölünmüştür. Öğrencilerin sene sonundaki başarıları 3-seviyeli sınıflandırma tekniği ile 9 farklı algoritma kullanılarak sınıflandırılmıştır. Çalışmada kullanılan sınıflandırma algoritmaları için en uygun parametre setleri deneme yanılma yöntemi ile belirlenmiştir. Öznitelik seçim yöntemleri kullanmadan, veri setindeki 23 öznitelikle kurulan Gradyan Artırma algoritmasıyla doğruluk değeri %95.17'ye ulaşılmıştır. Gömülü yöntemler ile öznitelik seçim yöntemleri kullanarak 11 özniteliğe indirgenen veri seti ile kurulun Gradyan Artırma algoritmasında doğruluk değeri %93.31'e düşmüş fakat Rastgele Orman algoritmasında doğruluk değeri %95.54'e çıkmıştır. Sonuç olarak, Rastgele Orman algoritması gömülü yöntemler öznitelik seçim teknikleri ile birlikte kullanıldığında, öğrenci performansını tahmin etmek için en etkili algoritma olduğu görülmüştür.

Cloud based veins recognition and authentication using CNN

In the current digital landscape, numerous systems require a reliable recognition method to verify the identity of individuals accessing their services. As technology continues to evolve, there's an escalating demand for robust and secure biometric identification. With increasing advancements in the interface between humans and computers, new biometric modalities have emerged and gained traction. Among these, the finger vein has emerged as a particularly compelling area of research. The underlying rationale for this heightened interest is the unique characteristic of the human finger vein, its intricate structure, and the inherent difficulty in counterfeiting or stealing this information. This thesis delves deep into proposing a finger vein-based recognition system tailored for individual identification. The finger vein stands out primarily because of its distinctiveness. Each individual's finger vein pattern is intricate and different, almost akin to the uniqueness of a fingerprint but concealed within the body, making it a challenging target for malicious intents. Unlike external features like facial structures, fingerprints, or iris patterns, which can be potentially copied, mimicked, or stolen, the finger vein remains protected inside the finger, making it an optimal choice for biometric identification. The methodology of the proposed recognition system is built on a series of sequential processes. First, the acquired finger vein image undergoes a preprocessing stage. This initial step is essential to refine the image by removing any superfluous details and enhancing its core features. The color image is transitioned into a grayscale format, streamlining the data and eliminating any color-based anomalies. To accentuate the details and improve clarity, histogram equalization is employed. This technique amplifies the contrast, ensuring that the vital vein patterns stand out prominently. Once the image is preprocessed and enhanced, the system shifts focus towards feature extraction. The choice of feature extraction method is pivotal, as it directly impacts the subsequent identification accuracy. This research leverages the acclaimed Linear Discriminant Analysis (LDA) for this purpose. LDA, being one of the most prominent feature extraction techniques, ensures that the extracted features are not only distinct but also optimal for classification. The heart of the recognition system lies in its classification model. This research proposes a deep Convolutional Neural Network (CNN), renowned for its capacity to discern patterns and classify with astounding precision. The advantage of employing a deep CNN in this context is its capability to deliver high accuracy without necessitating extensive datasets. A notable outcome of this research is the exemplary recognition performance yielded by the deep CNN model. When tested on the SDUMLA-HMT finger vein dataset, the model achieved an accuracy of 99.65%. Meanwhile, on the UTFVP finger vein dataset, it further exceeded expectations with an accuracy rate of 99.72%. These results are a testament to the efficacy of the proposed model, underscoring the potency of using a 1-dimensional convolutional network combined with dense layers. Another significant contribution of this thesis is the amalgamation of LDA with the deep model. This fusion has demonstrated superior prowess, not just in terms of recognition accuracy, but also in training speed, ensuring a swift and efficient system. Lastly, to bridge the recognition system with the cloud database, the TCP/IP protocol is employed. This ensures that once an individual is recognized, their pertinent information can be swiftly retrieved from the cloud database. This integration not only augments the recognition system's functionality but also elevates its application potential in real-world scenarios.

Automatic based framework by using PCA and deep learning for website phishing classification

An example of online identity theft is phishing. Phishers utilize social engineering to obtain victims' financial account information and personal digital identity information. Social engineering scams utilize phony emails to deceive unsuspecting victims into visiting bogus websites that ask for financial information and is called as deceptive phishing attack (PA). Deep learning (DL) algorithms are constantly improving, for making it simpler to identify phishing websites and reducing human errors and minimizing detection times. The overall objective of this research was to achieve high accuracy (Acc) that would aid in spotting phishing websites. DL algorithms, branches of artificial intelligence (AI) are applied in this work to capture the inherent characteristics of the website and classify websites as phishing or non-phishing. The results show that the proposed CNN model achieved the highest Acc of 97.28%.

Document classification using improved word embeddings

In this study, one of our primary objectives is to develop a trained model capable of classification using deep learning techniques for prediction. Neural networks, especially in the realm of natural language processing, have demonstrated impressive results, notably in document classification. Researchers have focused extensively on classification prediction. Convolutional network models, recurrent networks, and other embedding mechanisms are employed where texts are extracted (embedded) from documents either at the sentence or word level. Historically, the Word2Vec model was utilized in natural language processing to extract words based on context. This was later augmented with Long Short-Term Memory (LSTM) networks. The use of N-gram properties, in context with the text and associations between words, has proven to enhance prediction accuracy in classification tasks. Previous studies have primarily based document classification on visual methods or formats, perhaps concentrating on titles and abstracts. However, this article posits that classification should be anchored in word inclusion. Utilizing a dataset comprised of 47,000 texts and topics, we employ word embeddings to determine document themes. These embedded words — vast textual content — are sorted into seven primary categories, serving as foundational classes in our dataset. This data then trains deep learning models designed for document classification (both for training and testing). Once trained, this model can autonomously classify documents based on embedded words and texts. Our approach begins by extracting words from the dataset's texts. Subsequently, two models are constructed using Word2Vec. The words undergo lemmatization, reverting to their original form. Superfluous elements, such as symbols and punctuation, are purged to ensure the text remains pure, concentrating solely on semantically significant words. These cleansed word series are then used to train the two models, aiming to establish correlations between words. Both models strive to construct associations based on word sequences within the text. The first model assigns vectors to words based on context and endeavors to predict context via these words. In contrast, the second model hinges on the interrelationships between words and predicts specific words based on classification, yielding a relational concept termed "Neighbor word". Finally, we employ a deep learning model rooted in Long Short-Term Memory (LSTM). This is buttressed by the relationships deduced from the two Word2Vec models. Evaluations between them are conducted to ascertain which offers superior performance in predictive classification

Solar power harvesting design for wireless communication based wireless sensor networks application

This study conducted a thorough exploration of the efficiency of different maximum power point tracking algorithms within photovoltaic energy harvesting-based wireless sensor networks, with a specific focus on how these algorithms respond to environmental factors such as temperature and solar irradiance. Among the various algorithms evaluated, the Particle Swarm Optimization (PSO) algorithm stood out due to its exceptional performance. It exhibited a notably low overshoot of 14.8 volts and 25 mA at 0.1 seconds. Comparatively, the Perturb and Observe (PO) algorithm displayed a slightly higher overshoot of 15 volts and 25.5 mA. The Incremental Conductance (IC) and Fuzzy Logic (FL) algorithms both showed even higher overshoots, reaching 15.3 volts and 27 mA within the same timeframe. In the critical area of output power, the PSO algorithm again emerged as the most efficient, with the IC and FL algorithms following behind, and the PO method demonstrating the weakest performance. Furthermore, the study uncovered that environmental conditions, specifically temperature and solar irradiance, significantly influence the charging process of wireless sensor batteries. These factors have a profound effect on the charging efficiency and rate, which in turn impacts the overall performance and lifespan of the batteries in wireless sensor networks. The study recommends the use of the PSO algorithm for optimizing WSN battery charging processes due to its efficiency, adaptability, and responsiveness to changing conditions. However, for limited computational resources, alternative algorithms like IC or FL can be considered. The study also emphasizes the importance of incorporating environmental variables like temperature and solar irradiance into the battery charging optimization strategy to enhance charging efficiency, improve battery longevity, and enhance overall performance in WSN systems. Despite its computational complexity, PSO offers significant improvements over the basic P&O method. The selection of batteries that can endure a range of weather conditions is essential, especially considering the impact that extreme temperatures can have on battery performance. This research provides valuable insights that can contribute to enhancing the robustness, sustainability, and reliability of wireless sensor network systems, enabling them to operate effectively even in challenging environmental conditions. It offers a comprehensive evaluation of various MPPT algorithms used in photovoltaic energy harvesting-based wireless sensor networks, assessing their performance across different environmental scenarios using MATLAB 2021b.

Noise removal from the image using convolutional neural networks-based denoising auto encoder

The increasing use of digital cameras and imaging devices has led to a surge in daily images, increasing the demand for high-quality images in fields like medical imaging and surveillance. However, noise poses a significant challenge in image processing and analysis, as it degrades image quality and disrupts the preservation of essential features like edges, corners, and textures. Traditional denoising techniques struggle to balance these challenges. In response to this challenge, this thesis introduces a novel approach to image denoising, utilizing a denoising autoencoder based on convolutional neural networks (CNNs). The core objective of this research is to develop a method that effectively reduces noise in digital images while preserving key features. This is achieved through a two-step process involving an autoencoder and a CNN. The initial phase of this process involves categorizing input images into training and testing datasets. This categorization is crucial for the subsequent training and evaluation phases of the model. The training phase employs a denoising autoencoder, a variant of the traditional autoencoder specifically designed for noise reduction. This autoencoder learns to reconstruct noise-free images from their noisy counterparts. The success of this phase is contingent upon the autoencoder's ability to capture the essential features of the images while discarding the noise. Following the autoencoder's training, the denoised images are then fed into a convolutional neural network. The CNN, renowned for its effectiveness in image recognition and processing tasks, further refines the denoised images. It does this by learning hierarchical representations of the data, enabling the extraction of more sophisticated features. The training of the CNN is a critical step in enhancing the overall quality of the denoised images. The testing phase involves evaluating the performance of the proposed system using the test dataset. This phase is pivotal in determining the efficacy of the model in real-world scenarios. The evaluation metrics employed are the Root Mean Square Error (RMSE) and the Peak Signal-to-Noise Ratio (PSNR), both standard in assessing image quality. To facilitate a comprehensive evaluation, the MATLAB programming language was used due to its extensive support for image processing and neural network functionalities. The experiments were conducted on two distinct datasets: the COVID-19 Radiography Database and the SIIM Medical Images Dataset. These datasets were chosen for their relevance in medical imaging, a field where the quality of images is critical. The results of the evaluation demonstrate the superiority of the proposed method over the baseline techniques. On the COVID-19 Radiography (CXR) dataset, the proposed method achieved an 8% improvement in PSNR and a 53% reduction in RMSE compared to the baseline method. Similarly, on the CT Medical dataset, the proposed method outperformed the baseline by 5% in terms of PSNR. These results are indicative of the method's efficacy in not only reducing noise but also in enhancing the overall image quality. One of the key strengths of the proposed approach is its versatility. While the research focused on medical imaging, the method is applicable to a wide range of domains where image quality is critical. Furthermore, the two-step process of using an autoencoder followed by a CNN provides a robust framework for denoising. The autoencoder effectively reduces noise while the CNN enhances the image features, resulting in a synergistic effect that significantly improves the quality of the images. Finally, this thesis presents a novel and effective approach to image denoising, leveraging the strengths of denoising autoencoders and convolutional neural networks. This research contributes significantly to the field of image processing, offering a promising solution to the ever-present challenge of noise in digital images.

Analysis of cancer dataset with statistical learning

Cancer continues to pose a significant global health challenge, underscoring the criticality of early and accurate diagnosis for enhancing treatment outcomes and patient well-being. The classification of cancer types assumes a pivotal role in tailoring treatment plans, minimizing unnecessary procedures, and optimizing therapeutic success. This thesis presents an extensive analysis of statistical learning algorithms and machine learning (ML) algorithms on breast cancer, lung cancer, and prostate cancer datasets. The primary objective was to evaluate the algorithms' performance in distinguishing between benign and malignant samples across diverse cancer types. To ensure robust and reliable results, a comprehensive steps of preprocessing techniques was implemented, encompassing data cleaning to address null values and duplicate records, data scaling for feature normalization, random over-sampling to tackle class imbalance, and an 80:20 data splitting ratio for training and testing. Additionally, cross-validation was employed to assess model generalization and robustness. The paramount importance of accurately diagnosing cancer types lies in its potential to significantly impact patient outcomes and guide treatment strategies. The results showcased impressive accuracies ranging from 95.8% using ridge logistic regression to 97.2% using lasso logistic regression for breast cancer. Similarly, ML algorithms, such as Decision Tree, SVM, Random Forest, and XGBoost, achieved accuracies between 93% using random forest to 98.6% using XGBoost for breast cancer. Additionally, lung cancer statistical learning algorithms demonstrated accuracies between 93.75% using Ridge regression to 96.87% using Lasso regression, while ML algorithms achieved accuracies from 95.83% using Decision tree to 98.95% using Random forest. For prostate cancer, statistical learning algorithms achieved accuracies between 74.11% using ElasticNet regression to 77.64% using Lasso regression, and ML algorithms achieved accuracies ranging from 63.53% using Decision tree to 75.29% using SVM. These findings underscore the effectiveness of both statistical learning and ML algorithms in cancer classification, affirming their potential applicability in real-world scenarios to advance cancer detection and diagnosis.

Cryptocurrency analysis using machine learning approaches

Cryptocurrency is (CRYPTO) popularity and commercial acceptance are pivotal in reshaping the financial system. The allure of potentially high returns has especially piqued the interest of investors in CRYPTO trading. To maximize returns on Bitcoin (BTC) investments, precise price prediction (PP) becomes indispensable. Since CRYPTO PP is inherently a time series task, utilising Deep Learning (DL) models is highly advisable. In this context, our study delves into a comprehensive assessment of different DL models, focusing primarily on Artificial and Convolutional Neural Networks, to predict daily BTC prices (BP). Our initial dataset was sourced from Kaggle, a premier hub for data science endeavours. To ensure consistency during preprocessing, we employed the min-max scaler technique. Alongside the DL models, our analysis also encompassed a spectrum of Machine Learning (ML) models such as, Decision Tree Regression (DT), Linear Regression (LR), Random Forest Regression (RF), and others. This was to offer a holistic comparison across various predictive methodologies. To gauge the predictive prowess of these models, we employed key regression metrics like mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (R). It's essential to note that a lower RMSE signifies better model performance, indicating reduced prediction errors. In our results, CNN emerged as the top contender among DL models with an RMSE of 0.0543, MAE of 0.0324, and an R-value of 0.960, underscoring its exemplary capability in forecasting BP. In our evaluation of traditional ML models, the RF model showed remarkable performance with an RMSE of 0.0561 and an MAE of 0.0246. Per previous discussions, a lower RMSE value indicates better predictive accuracy, making the RF model highly effective in this context. Despite this, it's noteworthy that the RF model's performance was slightly inferior to that of the CNN model, especially regarding RMSE. Additionally, while the R2 value for the RF model was 0.958287, indicating a solid fit to the data; the correlation coefficient (R-value) was not specified for this or other ML models. Including the R value could provide a more nuanced understanding of each model's performance and should be considered in future evaluations.

Derin öğrenme yöntemlerini kullanarak X-ray görüntüleri üzerinden farklı implant türlerinin sınıflandırılması

Bu çalışma, X-ray görüntüleri üzerinde farklı implant türlerini sınıflandırmak için derin öğrenme modellerinin kullanılması üzerine odaklanmaktadır. Açık veri kaynaklarından elde edilen 597 görüntüden oluşan 4 sınıflı omuz implant veri seti ile 2376 adet görüntüden oluşan diş implant veri seti kullanılarak gerçekleştirilmiştir. Bu sınıflandırma işlemi, InceptionResNetV2, ResNet152V2, Xception ve DenseNet201 gibi derin öğrenme modelleri kullanılarak gerçekleştirilmiştir. Model eğitimi için Adam ve RMSprop algoritmaları ve farklı öğrenme oranları seçilmiştir. Ayrıca, veri setlerinde veri artırma ve CLAHE filtresi uygulanarak, modelin performansı artırılmaya çalışılmıştır. Bu tekniklerin uygulanmasıyla elde edilen sonuçlara bakıldığında her iki veri setinde de Xception modeli en iyi başarımı göstermiştir. Omuz implant veri setinde modelin test doğruluk değeri 0,85, F1-skoru 0,85 ve AUC değeri 0,95 iken diş implant veri setinde test doğruluk değeri 0,90, F1-skoru 0,91 ve AUC değeri 0,97 ulaşmıştır. Bu araştırma, x-ray görüntüleri üzerinde farklı implant türlerini sınıflandırmak için derin öğrenme modellerinin kullanılabilirliğini ve etkinliğini vurgulamaktadır. Elde edilen sonuçlar, bu teknolojilerin klinik uygulamalarda potansiyel olarak değerli bir rehber olabileceğini göstermektedir. Veri artırma ve CLAHE filtresi uygulanması, modelin performansını artırmaktadır.

A deep learning approach for load demand forecasting of power system

This study presents a deep learning methodology tailored for the precise forecasting of power system load demands, a critical component in the management and planning of electrical grids. The focus of our thesis lies in comparing the performance of several regression models, including both traditional machine learning algorithms and a deep learning model, specifically a Long Short-Term Memory (LSTM) network. The performances of models were evaluated based on various metrics, namely Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and for all except the LSTM model, the Root Mean Absolute Error (RMAE) and losses. Our findings reveal that the LSTM model significantly outperforms traditional regression techniques, achieving an MAE of 2.1351 and an MSE of 3.8831, metrics that are substantially lower than those recorded for the other models. This highlights the LSTM's superior ability in handling the complexity and variability inherent in power load demand forecasting. This study underscores the potential of deep learning, particularly LSTM networks, in enhancing the accuracy and reliability of power load demand forecasting. Such advancements are vital for optimizing the operation and efficiency of power systems, facilitating better energy management, and contributing to the overall stability of electrical grids.

Makine öğrenmesi ile BİLSEM'de öğrenci başarısını tahmin etme

Bu tez, Türkiye'deki üstün yetenekli öğrencilerin belirlenmesi süreçlerinde makine öğrenmesi algoritmalarının kullanımını ve etkilerini incelemektedir. Çankırı ve Ankara'daki Bilim ve Sanat Merkezleri (BİLSEM) tablet uygulamasının birinci aşamasını geçen öğrencilerin verileri kullanılarak öznitelik seçimi ve model eğitimi yapılmıştır. Bu sayede uygulamada başarılı olacak öğrencilerin tahmin edilmesi sağlanmıştır. Kullanılan 19 öznitelik seçimi ve 12 tane model eğitim algoritması arasında en çok Variance Thereshold öznitelik seçimi algoritması ile birlikte RadialSVM model geliştirme algoritması 0.67 başarı oranı ile uygulamadaki başarılı öğrencilerinin tahminini yapmıştır. Bulgular, bu algoritmaların öğrenci başarılarını tahmin etmede yüksek doğruluk oranlarına sahip olduğunu göstermektedir. Öğrencilerin bireysel başarıları, ailelerin sosyo-ekonomik durumu, ebeveynlerin eğitim düzeyi ve okulun fiziksel koşulları gibi faktörlerin öğrenci başarıları üzerinde önemli bir etkisi olduğu tespit edilmiştir. Böylece, BİLSEM seçim sürecinde kullanılabilecek yeni bir karar destek sistemi önerilmiştir. Önerilen sistem, öğrenci verilerinin analizi ve modelleme süreçlerinde doğruluğu artırarak seçim sürecinin daha objektif ve adil yürütülmesini sağlamaktadır. Bu sayede, üstün yetenekli öğrencilerin erken yaşta tespit edilmesi ve uygun eğitim programlarına yönlendirilmesi mümkün olacaktır.

Derin öğrenme ve boyut indirgeme destekli makine öğrenmesi yöntemleriyle böbrek tümörü tespiti

Bu tez, böbrek tümörlerinin BT görüntülerinden hızlı ve doğru biçimde saptanması için derin öğrenme ve boyut indirgeme destekli makine öğrenmesi tabanlı bir yaklaşım önermektedir. Çalışmada iki büyük ve kapsamlı BT veri seti (V1, V2) ön işlenmiş, Gri Düzey Birlikte Oluşum Matrisi (GLCM) ile doku özellikleri çıkarılmış ve T-SNE ile iki boyuta indirgenmiştir. Elde edilen düşük boyutlu öznitelikler üzerinde KNN, Bagging Ağaçlar, Karar Ağacı, İnce Gauss SVM ve üç katmanlı Yapay Sinir Ağı eğitilmiştir. Karşılaştırma amacıyla AlexNet, EfficientNet-B0, Darknet-53, Xception ve DenseNet-201 ile derin öğrenme deneyleri yürütülmüştür. Sonuçlar, DenseNet-201'ün 0,0001 öğrenme oranında %99,75 doğruluğa ulaştığını; önerilen kademeli GLCM+T-SNE yaklaşımının ise çok daha küçük model boyutu ve yüksek tahmin hızıyla KNN'de V1 için %99,65, V2 için %99,98 doğruluk sağladığını göstermektedir. T-SNE kullanımı, yalnızca GLCM'e göre benzer doğrulukla anlamlı hız ve bellek kazanımları sunmuştur. Önerilen yöntem, düşük sistem gereksinimleri, gerçek zamanlı kullanım potansiyeli ve veri kümesi bağımsız sağlam performansıyla klinik karar destek süreçlerini güçlendirmektedir. Gelecek çalışmalar, farklı modaliteler, transfer öğrenmesi ve açıklanabilir yapay zekâ teknikleriyle yöntemin genellenebilirliğini ve yorumlanabilirliğini artırmayı hedeflemektedir. Değerlendirme doğruluk, kesinlik, duyarlılık, F1 ve özgüllük metrikleriyle, karışıklık matrisleri üzerinden yapılmış; hiperparametreler ızgara aramasıyla seçilmiştir. Model boyutu ve tahmin hızındaki iyileşmeler, sınırlı belleğe sahip uç cihazlarda dağıtımı kolaylaştırmakta ve güvenilirliğini artırmaktadır.

Zararlı yazılım saldırılarının öznitelik seçimi kullanılarak yapay zeka tabanlı tespiti ve analizi

Gününüzde siber güvenlik saldır yöntem ve tehditleri giderek karmaşıklaşan yapıları ve gelişen saldırı teknikleriyle küresel ölçekte büyük bir sorun haline gelmiştir. Güncel tehditlerin en önemlilerinden olan zararlı yazılımların sürekli olarak kendilerini yenilemesi ve saldırı tekniklerinin gelişmesi, geleneksel güvenlik yöntemlerinin yetersizliğini ortaya koymaktadır. Bu tez çalışması, yapay zekâ ve makine öğrenimi tekniklerinin zararlı yazılımların tespit ve analiz süreçlerindeki etkinliğini artırarak, siber tehditlere daha hızlı ve etkili bir müdahale sağlama kapasitesine odaklanmaktadır. Tez kapsamında, zararlı yazılım türleri ve saldırı vektörlerini içeren kapsamlı bir veri seti oluşturulmuş ve bu veri seti, verilerin ön işleme adımlarıyla temizlenerek analiz için uygun hale getirilmiştir. Makine öğrenmesi sürecinde karar ağaçları, random forest, k-en yakın komşu (k-NN) ve yapay sinir ağları (Neural Network) gibi farklı algoritmalar kullanılmış, öznitelik seçimi ve optimizasyon teknikleri uygulanarak model performansları karşılaştırılmıştır. Veri seti, test ve analiz olmak üzere iki gruba ayrılarak modelin değerlendirilmesi sağlanmıştır. En iyi performansı gösteren model, derinlemesine incelenmiş ve elde edilen analiz sonuçları üzerinde tartışılmıştır.

Road segmentation in satellite images using deep learning

We demonstrate several approaches for dealing with road segmentation problems. Some rely on the encoder-decoder method, some on the Generative Adversarial Network (GAN) technique, and others on a fully convolutional network. The approaches utilizing Encoder- Decoder and GAN seem to have promise. Due to the great performance of Encoder-Decoder Deep Convolutional Neural Networks in many segmentation problems. Our aim is to apply all recent model architectures that use the DCEP network as a primary base model on two open-source data sets DeepGlope, and Massachusetts. We choose the most common encoder-decoder models that proved great performance for different data sets of image segmentation. We choose Unet, FPN, PSPNet, Unet++, PAN, LinkNet, DeepLab- v3, DeepLab-v3+, and MA-Net for our experiments and we give a brief comparison based on the result. We show the results for each model we use, both with and without the bilateral filter, and we show how the IOU (Intersection Over Union) and Dice loss of the results for all models on the Massachusetts data set are very similar. In an effort to improve model performance, we also use different data augmentation parameters, however, the results are the same for this data set. The Unet model has an excellent IOU for the DeepGlobe data set, scoring 95.46% accuracy.

Human identification using palm print images based on deep learning methods and gray wolf optimization algorithm

In this thesis, we used palm print images for human identification. This thesis contains four steps. First, the features of the palm print images are extracted by pretrained network that uses the deep learned network. Thre pretrained networks such as Googlenet squesse net and Alexnet are used to extract the features from the images. Then the best features were selected by using the Gray-Wolf optimization method. In the third step, these features are used in the nearest neighborhood method for recognition of the dataset to be used in the test data. Finally, we evaluated the results with recognition rate that calculated the percentage of the recognitions of the mistake index and correct index from the dataset. The aim of this thesis is to use the Gray-Wolf Optimization method and deep learning to get high recognition rate. In this study we used two famous dataset such as Polytechnic Hong Kong university dataset and Tongji Contactless datasets, the recognition rate for the different methods is evaliated and tested. We are showing that the proposed method has a very high performance in the recognition rate than other different methods such as principle compomnent analysis, Local binary pattern and laplacian of gaussian gabor transform. The recognition rate from the proposed method we obtain 96.72.

Deep learning neural network based on the PSO for leukemia cell disease diagnosis from microscope images

Leukemia is a cancer of blood tissues, including the bone marrow and the lymphatic system. White blood cells (WBCs) are affected by leukemia, causing the bone marrow to produce an excessive amount of abnormal, and immature WBCs, which do not function properly and circulate through the body. Leukemia is one of the most difficult diagnoses to make that uses image features extracted from a microscope. In this thesis, deep learning neural networks based on optimization methods have used for leukemia cell disease diagnosis from microscope images. First, two powerful Convolutional Neural Networks (CNNs) architectures with pre-trained on ImageNet have used to extract features from leukemia images, which are: Deep Residual Networks (ResNet-50) and GoogLeNet. Then, Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO) have used to select the effective features to achieve the most accurate and fastest prediction that will be obtained by using different of machine learning algorithms which are: Support Vector Machine (SVM), K-Nearest Neighbor (K-NN), and Decision Tree (DT). The obtained results were as follows: 100%, 100%, and 89.7% for SVM, K-NN, and DT, respectively.

Üç performans analizi fotovoltaiksistemlerin faz paralel invertör

Unified model for fingerprinting code applications

The enormous popularity of digital contents contains unlicensed and illegal activities which are involved in different perspectives. Despite several efforts are made toward an illegal copy and redistribution (legal and technical), still the problems exist with growing use of digital application such as storing and distribution of contents, the problem is becoming more worrisome. Attaching a unique invisible marking to every unit of object, is viable trading with problems of unauthorized copying, that is if an illegal copy is released elsewhere, it is going to be practical to find the original owner of the object (traitor). This phenomenon of enforcement, in a copyright protection, is what is referred to as fingerprint code. Dealing with code, for fingerprinting in digital contents, in the context of several users combining (colluding) to create a new, unauthorized copy (pirate). Theoretical bounds on the performance in the code are realized, with general methods of theoretically observe whether certain users are involved in the illegal activity (copy and redistribution) is presented, through checkmate items. In this paper, we have introduces collusion–secure fingerprinting code have been similar to construction by aggeloss and serdar; but attain a shorter code in length, faster accusation by splitting the group of users in to partitions and allow for code over arbitrary alphabets, and asymmetric fingerprinting scheme out of symmetric scheme from the early idea work of tardos in his fingerprinting code.

A fast and accurate skew detection algorithm

Skew detection of digital documents is one of the important preprocessing steps of document layout analysis and optical character recognition, because it affects the quality and speed of these algorithms. In literature, there is a variety of algorithms and studies. Three of these algorithms are popular, namely i) projection profile, ii) Hough transform and iii) nearest neighbor clustering. In this thesis, we propose a novel fast and accurate skew detection method. The proposed method forms rectangular blocks and uses the centroids of blocks to estimate the skew angle. Unlike other methods, the output of the proposed algorithm is a deskewed document. In other words, no additional skew correction step is required. Moreover, it is independent of the script used in the document. Having a wide range of scan interval is another advantage of the proposed method over other methods. Tests made on 180 open access images give us promising results in terms of speed and accuracy. In comparison with the most popular methods, beyond being the fastest, the proposed algorithm gives the second most accurate results

Traffic signaling control at highway intersections using morphological image processing technique

Nowadays, the result of growing population in intercity areas vehicle traffic congestions become an important issue to be solved. For the solution of this problem various methods and techniques have been suggested in many recent studies. Any preventive measure to ease traffic flow along the roads is valuable to save time and fuel wastage, and also reducing CO and CO2 emissions. In this thesis, we attempt to implement the morphological image processing technique on the images obtained from a simulated junction model. The comparisons of different images are made from the matching score results to decide green light signaling duration for each road of the junction. During implementation of the morphological image processing technique we focus on the individual effect of each operator to improve the accuracy of matching scores. Since we are basically interested in the area occupied by the vehicles, it is observed that ignoring edge detection operator does not produce any loss in the image quality, but rather increases the accuracy of image matching scores resulted from long vehicles, daylight and seasonal variations. The results indicate that data of matching scores for each road provides a successful decision for the green signaling duration at different daytime periods. The method suffers from vehicles having dark colors, since they are accounted as a part of the road. Suggested future work involves the implementation of the procedure for live traffic at a junction and eliminating the problem resulted from colored vehicles under different weather conditions.

Face feature selection using genetic algorithm under different biometric variations

In the current study face recognition under different biometric variations is investigated applying Principal Components Analysis (PCA). In order to improve the recognition performance Genetic Algorithm (GA) is selected. The algorithm follows optimized selection of PCA features based on GA operations on the datasets ORL, FERET and BANCA. The maximum recognition rate (MRR) results obtained with ORL and FERET databases are found to be close to the results of computed with WAVELET-PCA-GA-SVM method. Further the MRR results obtained for BANCA database is 100% as that of the computed with WAVELET-PCA-GA-SVM method for YALE and YALE-B databases. Generally PCA on GA is found to be effective in removing irrelevant data groups and therefore it improves the performance.

A survey on image illumination enhancement techniques and a new approach to assess the image illumination

Presently, digital images are used in many parts of the daily life and various fields of academic and scientific researches. These applications depend on the quality of the image obtained. Illumination enhancement is required for all the applications which are using digital images in some part, to make them work properly, also assessment of the illumination quality is, as important as the illumination enhancement for comparing images and having an idea about the proposed output of the academic research results in the field of digital image processing. Illumination problems has become a demanding matter in various applications of digital image processing. A relevant argument in the field of illumination enhancement is lack of a quantitative measurement method to assess the illumination of an image. A quantitative measurement that is demonstrating the illumination case (contrast level, brightness etc.) of an image is proposed. Proposed new approach in this thesis, appropriates the input image's estimated Gaussian distribution and the Kullback-Leibler Divergence of the estimated Gaussian distribution and the expected Gaussian distribution for measurement and the calculations. In the results part, the experimental works show the influence and the accuracy of the proposed approach for the assessment of image illumination quality.

Electronic public distribution system e- government food department

This thesis entitles "Electronic Public Distribution System" is a web based an online application intended for citizen. The main problem of this thesis is to automating "Food Department Office (FDO)" and digitalizes deliver services mainly to the citizens and public sectors. (FDO) has been automated specially for Governorate of Sulaimani Region. The main objectives of this system are to make it interactive and facilities to the users for solving daily problems. Citizens can use this system to search for information and send necessary documents within the electronic government for the Ministry of Commerce in client side for solving their problems. Also it is used to complete their information in the database system by using full name and ration card number of family to enter the system. Citizen can follow this case and can find the result for this request from the system. To tackle this problem the government's types such as C2G and G2C have been used. "Citizen to Government and Government to Citizen" users can send their request to Electronic Public Distribution System directly and E-center of the system can take action for these cases by staff member who has a special user name and password, staff can do update and increase in the database files such as family, person, and agent. In order to evaluate the system " Usability test Model" have been used to evaluate the satisfactions of the users during using the system and for this reason four administrators of the system and 300 three hundreds of online users have been involved to test the measurement of the reliability, usability and availability the system. The results shows that the users satisfaction feedbacks obtained by using Usability test Model.

Automated multi-functional smart home system using arduino

Monitoring the condition of domestic and its appliances has been always ranked among the main concerns of the daily lives of people. This concern is far more prominent for those who are outdoors during the day and have to keep watch on their children or elderly people. Maximizing home security is also proposed in many other. The smart home refers to a system that uses information technology and computers or smart phones to monitor domestic appliances. This technology can successfully resolve the concerns described above. In this thesis, a method is proposed that implements a smart home in practice in several phases called Providing security (detecting theft and notifying mobile alerts), providing safety and comfort (setting the desired home temperature, adjusting ambient lighting) and controlling the domestic appliances through mobile phones and GSM. The results obtained from the implementation of the proposed system show that the proposed system has been able to perform multiple tasks efficiently and can to be used by the public in many circumstances while, putting into effect factors such as security, safety and remote control of home appliances via smart devices.

A Study On Representing Cultural Heritage By Virtual and Augmented Reality

This study gives an overview of education, preservation and virtual learning of cultural heritage through virtual and augmented learning. Works that have been done on cultural heritage have a responsibility of replicating, visualizing and representing a virtual environment as if they were in a real world through the presentation of the virtual reality. The explains that cultural heritage is an important accept when it comes to learning and learning is not based on what the developer of the system wants but what the user is looking for. However, even if most virtual and augmented environment is done from what the developer thinks is appropriate still it offers learning environment and the preservation of the culture is still done (Central European University, 2017). This is because learning can take place to advance technology that the virtual and augmented reality offers to mean that individual who can do reach the cultural environment can still have the learning experience. It is also important to identify the need for the users of the system on what exactly they would like to learn and see for the creation of the virtual environment to be said it is successful to offer learning environment. Technology is what has facilitated the use of the virtual and augmented reality but this also has posted a challenge since not all users can assess the advanced technology. The users of the virtual environment would also like to have an opportunity of interacting with the objects in the virtual environment which still is a challenge to the developers of the system. The study gives evidence that the virtual and augmented environment offers educational, learning and preservation opportunity for culture.

Voltage regulation for grid connected photovoltaic systems using reactive power

Nowadays the microgrids with renewable energy sources (RESs) usage and development rapidly increase due to govermentel incentives,dropping the price of DERs and improvement of the technology. The high penetration of DERs espicialy if it is intermittent sources such as Photovoltaic (PV) sources, will effect the stability of the distribution network and causes disturbances for the voltage. To solve this issue the connected PV sources have to contribute to the stability of the distribution network. The conventional control of distrubition networks generally controlled by human which has many limitations and slow response time, so the it is important to use a real time controllers.Currently, the nine-zone diagram control (NDC) is widely used in the reactive power control (VQC) to keep voltage and reactive power within limits. This NDC plane is divided into nine regions, which determine on-load tap-changer (OLTC) and Static Var Compensators (SVCs) control strategy according the location of real-time measured voltage and reactive power. In this research the already available infrastructures (such as OLTC and SVCs) will be used in coordenation with PV sources to contribute to the voltage regulation via reactive power. The developed solution will elemenate some device failures, save budgets and improve the reialibility for both consumer and distribution network operator A case study have been carried out on a 8 bus distribution system to illustrate the effectiveness of the proposed method. This study showed how the voltage and reactive power problem can be resolved in a timely and economical way. A small scale microgrid has been simulated in this thesis and the results helped companseting distrubition grids. Matlab/Simulink have been used to verify the developed controller.

Classification of microarray gene expression cancer data by using artificial intelligence methods

Today, the development of computer technologies has affected the studies in many areas. Advances in molecular biology and computer technologies have revealed the science of bioinformatics. Rapid developments in the field of bioinformatics have contributed greatly to the solution of many problems waiting to be solved in this field. The classification of DNA microarray gene expressions is one of these problems. DNA microarray studies are a technology used in the field of bioinformatics. DNA microarray data analysis plays a very effective role in the diagnosis of diseases related to genes such as cancer. By determining gene expressions depending on the type of disease, it can be determined with great success rate whether any individual possesses the diseased gene. The use of high-performance classification techniques on microarray gene expressions is of great importance to determine whether an individual is healthy. There are many methods for classifying DNA microarrays. Support Vector Machines, Naive Bayes, k-Nearest Neighbour, Decision Trees, such as many statistical methods are widely used. However, when these methods are used alone, they do not always give high success rates in classifying microarray data. Therefore, the use of artificial intelligence-based methods to achieve high success rates in the classification of microarray data is seen in the studies. In this study, in addition to these statistical methods, it is aimed to obtain higher success rates by using a method such as ANFIS based on artificial intelligence. K-Nearest Neighbourhood, Naive Bayes and Support Vector Machines were used as statistical classification methods. Here, studies on two different cancer data, namely breast and central nervous system cancer, have been conducted. According to the information obtained from the results, it was found that artificial intelligence based ANFIS technique was more successful than statistical methods.

Stability considerations of connecting a large-scale photovoltaic plant to the public grid

In recent years, due to global warming, energy security and environmental problems, interest in renewable energy sources, which are environmentally friendly energy production, has increased worldwide. Accordingly, it attracts great interest in modern energy systems in the installation of large-scale photovoltaic power plants (LSPVPP) connected to transmission lines. In addition to the rising trend, the control and operation problems encountered during the integration of large-scale photovoltaic energy systems into the grid are increasing. A detailed analysis and regulations with limitations are needed to deal with these problems. In this article, a detailed examination of the problems encountered during the connection of LSPVPP to the network will be made. Network regulations created by the countries will be examined. The restrictions to be examined in these regulations will be observed limitations of the grid code used in Turkey. Regarding the problems encountered during the connection of these facilities to the transmission lines; fault ride through requirement (FRT), voltage and frequency, active power and frequency control, voltage control and reactive power limitations specified in Turkey in the first network with the correct design and simulations including regulation will be observed. The evaluations to be made; Suggestions will be made for the establishment of stable facilities with correct design and design and meeting the requirements of the grid regulation. Matlab / Simulink will be used to validate the assessments to be made.

Fake news detection with deep learning and machine learning methods

In our age, especially since the rise of technology and social media, fake news has become a problem in society. News has had an impact on society throughout history. Today, when online platforms are one of the biggest communication tools, we access the news over the internet. When we cannot access the right source of information, serious problems arise for the life of society. Anymore we see fake news spreading more and faster in some situations than real information. The availability of technology has brought us many of its problems too alongside the advantages of it. Fake news has been used for directing society and for social engineering rather than an innocent information mistake. Today, we are faced with a false fraud that guides social movements, makes wrong decisions, and affects political and economic events. The events experienced show us that work such as big data processing and social media data analysis needs to be done. In this study, fake news data is used which is a more general topic than fake fabricated or fictional content that can change this economy, sociology, politics. I evaluate this different artificial intelligence method, which aims at using machine learning and deep learning methods of fake news, with fake news detection. I compare NLP, LSTMs, and traditional machine method learning. LSTMs has had higher classification success in most machine method learning methods. The fact that a lot of work has to be done related to big data processing, internet news analyzing, social media data analysis encouraged me to this field. This thesis is expected to be useful for producing more effective solutions to fake news, which is an important problem.

Türkiye'nin uzun vadeli enerji tüketiminin metasezgisel algoritmalarla tahmini

Energy is one of the most important topics for the sustainable development of countries. Due to the fact that the energy used can be depleted, it imports many energy sources, and environmental factors, it is of great importance for Turkey to forecast how much energy needs may be in the future. In this study, whale optimization algorithm (WOA) was preferred from metaheuristic algorithms in order to forecast the amount of energy consumption of Turkey until 2040. A genetic algorithm (GA) and an artificial bee colony algorithm (ABC) were used to measure the performance of the WOA. These heuristic algorithms are increasingly preferred by the high forecasting performance obtained from the models they create. All models are arranged linear and quadratic and the result is obtained. Data for independent variables such as gross domestic product (GDP), population, imports and exports affecting energy consumption were used between 1990 and 2019. These data are taken from TUIK (Turkish Statistical Institute), BP World Energy Statistics Report, World Bank and IEA (International Energy Agency). Modeling of the past 30 years has been provided to determine the equation that gives the closest result to the actual value. For the next 20 years, calculations were made according to 4 different scenarios. According to these calculations, it was observed that the ABC linear model works better than the other two models with values of 99% R^2 and 4.35% MAPE.

Multimodal classifier for disaster response

Nowadays, social media data can be used to make a huge difference in making correct decisions in time-critical situations, such as the event of natural disasters. Social media content consists of messages, images, and videos. In some cases, understanding the damage caused by natural disasters only from text is not enough in terms of analysis, the effect of disaster is better understood using visual data. Text datasets from social media platforms are widely used by researchers, and a limited number of studies have focused on the use of other content such as images. This is due to the fact that the number of tagged image datasets related to disasters is very limited. Therefore, in this thesis, we aim to address this limitation by presenting a multimodal Turkish text and images dataset. Multimodal classification studies were carried out with the late fusion technique. Also, to achieve multimodal classification; a pre-trained LSTM model is used for classifying the text while a pre-trained CNN model is used for the visual content. Overall, concatenating both inputs in a multimodal learning architecture achieved an accuracy of 91.87%.

Integration of deep learning methods in the classification of rna-seq data

Cervical cancer is a type of gynecological cancer that affects the cells in the lower section of the uterus called the cervix. Cells become abnormal as a result of gene mutations, and uncontrolled division is the primary cause of cancer. On the other hand, Alzheimer's disease is an irreversible neurological disease that causes memory loss and dementia, primarily as a result of brain cell death as people age. The parts of the brain that control thinking, learning, and memory have been injured or destroyed, causing symptoms. Cervical cancer and Alzheimer's disease are both genetic disorders. As a matter of fact, gene expression is significant in the diagnosis and classification of Cervical Cancer and Alzheimer's disease. Many genes' information is kept in RNA-Seq data. To accelerate this approach and assist clinicians in the diagnosis process, methodologies can be constructed using classification algorithms with decreasing the number of irrelevant genes. The objective of this thesis is to use statistical and deep learning approaches to examine Cervical Cancer and Alzheimer's Disease utilizing RNA-Seq datasets built with genes obtained from real samples. To lower the size of the RNA-Seq data set, gene selection is done by 5%, 10%, and 30% among all genes for both datasets, and the gene expression data were generated for each gene according to the importance level of the genes. In three scenarios, the selected genes are trained and tested in the categorization process. Deep Neural networks, convolutional neural networks, and long short-term memory approaches are implemented for classification. Following this research, it will be evaluated which approaches work best in the classifications of Alzheimer's Disease and Cervical Cancer.

Cultural heritage building segmentation and classificationusing point cloud data

The documentation and digital accessibility of historical buildings is an important issue. An integrated system that can be used in the restoration processes of cultural structures can be created by evaluating a historical building as a whole, by considering the region and period conditions in which it is located, and by realising an intelligent system that can replace existing manual systems. In this thesis, automatic segmentation of structures using point cloud data belonging to cultural structures in Gaziantep province and mesh model creation studies were carried out from segmented point cloud data. An artificial intelligence-based learning model was used to segment historical buildings and mesh model studies were completed using poisson transformation methods. Using the results obtained, the most comprehensive step of the studies carried out with the aim of developing the HBIM (Heritage Building Information Modelling) system has been completed.

Klasik makine öğrenme algoritmaları ve transformer modeli ile Türkçe tweet duygu analizi

Teknolojik çağın başlamasıyla birlikte birçok uygulama günümüzde popüler hale gelmiştir. Bunlardan biri de Twitter'dır (Yeni adıyla X). Bu platform sayesinde birçok kullanıcı kişisel düşünce ve fikirlerini yazıyla, resimle ya da video ile paylaşabilmektedir. Paylaşılan bu veriler birçok insanın ilgisini çekmektedir. Bu verilerden kişisel çıkarım yapmak isteyen bilim insanınları, akademisyen ya da başka bir meslek grubundan insanlar çalışmalarını bu yönde yapmaktadır. Örneğin suçlu profili oluşturmak isteyen bir polis twitlerden bunu yapma amacıyla çalışma yürütebilir ya da bir reklam ajansı kişinin yazdıklarına ya da paylaştığı resimlere bakarak kişiye özel reklam üretebilir ve bu sayede ürün satıcıları ürünlerini hızlı bir şekilde satabilir. Bu tezde ise yüksek oranda doğruluk oranı aldığımız bir duygu analizi çalışması yapılmıştır. Bu çalışmada hazır bir Türkçe Tweet verisi üzerinden duygu sınıflandırılması amaçlandı. Bu veri seti 5 farklı etiketle ayrılmış 4000 veriyi içermektedir. Bu ham veri seti üzerinde ön işleme uygulandı. Ön işlemden geçen veri seti eğitim ve test olarak ayrıldı. Klasik makine öğrenmesi algoritmalarının performansları ölçüldü. Bu performanslar doğruluk, kesinlik, duyarlılık ve F1 skorunun yanı sıra makro ve ağırlıklı ortalama açısından ölçülmüştür. Ayrıca, her algoritma için hesaplanan karışıklık matrisi verilmiştir. En yüksek doğruluk oranına %96,88 ile Yığın algoritması ulaşılmıştır. Bu çalışmada ayrıca derin öğrenme alanı içerisinde olan önceden eğitimiş bir Transformer modeli kullanılmıştır. Veri seti eğitim, doğrulama ve test olarak ayrılmıştır. Aynı şekilde performansı doğruluk, kesinlik, geri çağırma ve F1 skoru açısından ölçülmüştür ve her algoritma için hesaplanan karışıklık matrisi verilmiştir. Bu model ile %93 doğruluk oranına ulaşılmıştır.

Bir üretim bandında üretilen malın kalite kontrolünün PLC kullanılarak gerçekleştirilmesi

Bu çalışma da bir üretim bandından gelebilecek malzemenin kalite kontrolü yapılmaya çalışılmıştır. Bu kontrol işlemi kullanıcının seçimi ile renk, boyut ve cinse göre yapılabilmektedir. Sistemde bu işlemleri gerçekleştirmek için PLC 1er kullanılmıştır. Yine sistemde üretim kayıtlarının tutulması ve kullanıcı isteklerinin PLC lere iletilmesi için bir PC de kullanılmıştır. Malzemenin boyutu optik sensörler yardımıyla tanınmaktadır. Malzemenin kalınlığı, bir lineer potansiyometre ile ölçülmektedir. Malzemenin rengi ise bir foto direncin değişik renklerde verdiği voltaj çıkışından elde edilmektedir. Renk ve kalınlık ölçme kısımlarından elde edilen voltaj değerleri 12 bitlik ADC kullanılarak PC ye aktarılmıştır. Malzeme kullanıcının belirttiği şartlara uyuyorsa her çalışmada 4 adet değişik özelliğe göre kutulara pnömatik pistonlarla tevzi edilmektedir. Tanınmayan malzemeler ise tanımsız kutusuna hatalı üretim olarak atılmaktadır. Malzemenin cinsi ve üretim bandında malzemenin olup olmadığı temassız sensörlerle tespit edilmektedir. Bu üç değişik çalışma modunda yapılan kontrol işlemlerinin deneysel sonuçlan ve sistemdeki birimlerde kullanılan yöntemlerle ilgili teorik bilgiler tezde verilmiştir.

Omuz eklemi, kas ve tendon hasarlı hastalar için akıllı fizik tedavi mentor uygulaması (fizmenu)

Dünya üzerinde bir milyardan fazla insan hethangi bir çeşit fiziksel engel ile hayatını sürdürmektedir. Herhangi bir fiziksel engeli olan veya fiziksel yaralanma geçirmiş insanların iyileşmeleri için fizik tedavi görmesi gerekmektedir. Günümüzde, birçok alanda olduğu gibi fizik tedavi alanında da teknoloji kullanılmaktadır. Bu tezin amacı fiziksel engelli insanların hastaneye ulaşım, egzersizleri doğru yapamama ve tedavi süreçlerinin kontrol edilememesi gibi sorunlarına çözümler üreterek, hasta tedavilerinin daha doğru ve etkin yapılabilmesini sağlamaktır. Bu tez kapsamında; omuz eklemi, kas ve tendon hasarlı hastaların, fizik tedavi merkezleri ve evde yaptıkları fizik tedavi egzersizlerinin doğruluk oranını arttıran, hastaların yaptıkları çalışmaları fizik tedavi uzmanına raporlayan ve hastaların iyileşme durumlarını gösteren kol omuz ve el sorunları anketi (DASH) sonuçlarını yapay zeka teknikleri ile tahmin eden bir fizik tedavi mentör uygulaması (FizMenU) önerilmektedir. Önerilen sistem, hastaların fizik tedavi egzersizlerini doğru yapmalarını sağlayan Kinect v2 tabanlı Geliştirilmiş Omuz Fizik tedavi Uygulaması (GOFU), fizik tedavi uzmanlarının hastanın tedavi sürecini gözlemleyebildiği ve müdahale edebildiği bir web platformu, DASH sonuçlarını tahmin edebilen yapay zeka destekli bir modülden oluşmaktadır. Önerilen FizMenU sisteminin yapılan testlerle omuz fleksiyon, abdüksiyon, iç rotasyon ve dış rotasyon pozlarının ölçümlerini klinik standartta geçerli ve güvenilir bir şekilde yapabildiği gösterilmiştir. Bilecik Devlet Hastanesi'nde gönüllülerin tedavi sürecinde kullanılan FizMenU'nun hastaların tüm egzersizlerdeki limitasyonlarını istatistiksel olarak anlamlı şekilde azalttığı gözlemlenmiştir. Geleneksel tedavi yöntemleri ile kıyaslanan FizMenU'nun hastaların limitasyonlarını azaltma oranlarında abdüksiyon egzersizinde %16.55, fleksiyon egzersizinde %15.09, dış rotasyon egzersizinde %1.84, iç rotasyon egzersizinde %8.58, daha başarılı olduğu gösterilmiştir.

Çok katmanlı aşırı öğrenme makinelerinin geliştirilmesi ve sistem modellemedeki başarımlarının analizi

Aşırı Öğrenme Makinesi (AÖM), sınıflandırma ve regresyon uygulamalarında yüksek verimlilik ve kolay uygulanabilirdiğinden dolayı son on yılda önemli araştırma konusu olmuştur. Çalışmamız kapsamında regresyon ve sınıflandırma problemleri için geliştirilen klasik AÖM, Radyal Tabanlı Fonksiyon Kullanan Aşırı Öğrenme Makinesi (RTF-AÖM), Meta Aşırı Öğrenme Makinesi (Meta-AÖM) ve Çok Katmanlı Aşırı Öğrenme Makinesi (ÇK-AÖM) gibi mevcut AÖM yapıları incelenmiştir. Bu çalışmada, anılan yapıların zayıf ve üstün yönleri incelenerek yapıların iyileştirilmesi ve yeni hibrit yapılar önerilmiştir. Çalışmamızdaki motivasyon unsuru, AÖM yöntemlerinin yaygın olarak sınıflandırma problemlerinde kullanılmış olması ve sistem modelleme üzerinde çalışılmamış olmasıdır. Bu bağlamda zamana dinamik bir şekilde bağlı sistemlerin yanısıra engebeli ve/veya keskin değişimli yüzeylere denektaşı sistemlerin modellenmesi üzerinde çalışılarak iyileştirme ve yeni karma ağ yapısının geliştirilmesi hedeflenmiş olup çalışmamızda istenen hedefe ulaşılmıştır. ÇK-AÖM yapısının sistem modelleme başarımını arttırmayı ve hesaplama zamanını düşürmeyi amaçlayan iki İyileştirilmiş Çok Katmanlı Aşırı Öğrenme Makinesi (İÇK-AÖM) adını verdiğimiz AÖM ağ yapıları geliştirilmiş olup deney sonuçları ile başarım üstünlünlükleri denektaşı sistemler üzerinde gösterilmiştir. İÇK-AÖM yapıları ile yedi farklı dinamik sistem modelleme uygulaması üzerinde başarımı incelenmiştir. Ayrıca çok katmanlı aşırı öğrenme makinesi ile radyal tabanlı fonksiyon ağları birleştirerek HybRBF-ML-ELM adını verdiğimiz karma bir ağ yapısı daha geliştirilmiştir. Engebeli ve/veya keskin değişimli yüzey problemlerinde başarımı deney sonuçları ile kıyaslamalı olarak gösterilmiştir. Dinamik sistem modelleme deneylerinde geliştirilen İÇK-AÖM'nin bazı durumlarda ÇK-AÖM'ye kıyasla hem eğitim hem de test veri setleri için %70lere varan daha iyi modelleme başarım sergilediği gözlenmiştir. Örneğin, Denektaşı Dinamik Sistem (DDS) 7 için 100 düğüm kullanıldığında ÇK-AÖM, İÇK-AÖM1 ve İÇK-AÖM2 yapıları 0.627977, 0.104272 ve 0.092683 ortalama test Karesel Ortalama Hatanın Karakökü (RMSE) sonuçları verilmiştir. ÇK-AÖM ile karşılaştırıldığında İÇK-AÖM1 yapısı 83% daha iyi ortalama test RMSE sonuçları sağlarken İÇK-AÖM2 yapısı 85% daha iyi ortalama test RMSE sonuçları sağlamıştır. Yüzey modelleme deneylerinde geliştirilen HybRBF-ML-ELM yapısının ÇK-AÖM yapısına göre her problem için daha başarılı olduğu gözlemlenmiştir. Örneğin, Griewank, Dropwave ve Schaffer2 fonksiyonları için hem eğitim hem de test RMSE sonuçlarına bakıldığında ÇK-AÖM yapısına göre yaklaşık olarak %99 daha iyi başarım elde edildiği görülmüştür. Peaks fonksiyonu için de eğitim verisinde ortalama RMSE değerine bakıldığında ÇK-AÖM yapısına göre %76,87 daha iyi modelleme başarım sergilediği gözlenmiştir.

Havacılığa yönelik turboelektrik itki sistemi için matematiksel model tasarımı

Bu çalışma, havacılık uygulamaları için gelecek vaat eden bir turboelektrik itki sisteminin matematiksel modellemesi ve simülasyonunu gerçekleştirmiştir. Ana güç kaynağı olarak, GE T700 turboşaft motorunun yüksek hassasiyete sahip MATLAB / Simulink matematiksel modeli kullanılmıştır. Seçilen gaz türbininin matematiksel modeli, N-3X turboelektrik uçaklar için tasarlanan NASA elektrik mimarileri ile birleştirilmiştir. Çift motorlu uçaklarda meydana gelen motor arızalarında, tek motor ile çalışma senaryosu önemli bir tasarım ölçütüdür. Bu senaryonun analizi için MATLAB / Simulink matematiksel modeli kullanılmıştır. Bu doğrultuda yapılan çalışma doğru sistem tasarımı ve gerekli ekipman seçimi gibi konularda öngörüler içermektedir. İki önemli mimari olan Baseline Architecture ve Inner Bus Tie Architecture'ın analizleri yapılmıştır. Bu doğrultuda, seçilen mimariye göre yapılacak değişiklikler ve ekipman farklılıkları açıklanmıştır. Ayrıca, turboelektrik mimariler için en uygun güç elektroniği uygulamaları konusunda bazı ön çalışmalar yapılmıştır. Bu turboelektrik itki sisteminin her bir elektrik ünitesinin güç gereksinimleri, dağıtım yüzdeleri, güç elektroniği sistemleri için ön güç değerlendirmesi ve nominal güç kapasiteleri, ana güç kaynağı olarak günümüzde yaygın bir şekilde kullanılan GE T700 turboşaft motor modeli kullanılarak hesaplanmıştır. Turboşaft motor 3 fazlı jeneratöre bağlanarak mekanikten elektriğe güç dönüşümü sağlanmıştır. 3 fazlı jeneratörün çıkış gerilimi yükselten bir 6-pulse doğrultucu aracılığıyla doğrultulup daha sonra bir gerilim kaynaklı invertör yardımı ile motorun gereksinimlerine göre gerilim düşürülebilecektir. Fakat, gerilimi gerekenden daha yüksek bir seviyede iletmek DC iletim kayıplarının azalmasını sağlayacaktır. Birbirlerine göre farklı üstünlükleri olan Baseline Architecture ve Inner Bus Tie Architecture'ın matematiksel modelleme sonuçları, gelecekte farklı güç sistemleri için yapılacak olan çalışmalara altyapı teşkil etmektedir. Bu turboelektrik itki sisteminin her bir elektrik ünitesinin güç gereksinimleri, dağıtım yüzdeleri, güç elektroniği sistemleri için ön güç değerlendirmesi ve nominal güç kapasiteleri, ana güç kaynağı olarak günümüzde yaygın bir şekilde kullanılan GE T700 turboşaft motor modeli kullanılarak hesaplanmıştır. Turboşaft motor 3 fazlı jeneratöre bağlanarak mekanikten elektriğe güç dönüşümü sağlanmıştır.3 fazlı jeneratörün çıkış gerilimi yükselten bir 6-pulse doğrultucu aracılığıyla doğrultulup daha sonra bir gerilim kaynaklı invertör yardımı ile motorun gereksinimlerine göre gerilim düşürülecektir. Gerilimi gerekenden daha yüksek bir seviyede iletmek DC iletim kayıplarının azalmasını sağlayacaktır. Bu matematiksel modelleme ile yapılan mimari tasarımın sonuçları, gelecekte farklı güç sistemleri için yapılacak olan çalışmalara altyapı teşkil etmektedir.

Makine öğrenmesi algoritmalarıyla Türk işaret dilinde harf ve dinamik sözcük tanıma

Toplumsal hayatın bir parçası haline gelen bilgisayarlar ve akıllı arabirimler, ses ve görüntü dışında, özellikle işaretlerin ve hareketlerin tanınmasında etkin bir şekilde kullanılarak hayatı kolaylaştırmaktadır. Sunulan bu çalışmada işaret dilini kullanan işitme ve konuşma engelliler ile diğer bireylerin kolay iletişime geçmesi ve işaret dilini bilmeyen bir kişinin bu dili anlayabilmesi çerçevesinde çalışılmıştır. Günümüzde teknolojik eğilimin taşınabilir sistemlere doğru olması sebebiyle işaret tanıma çalışmalarında LMC ve Kinect cihazı kullanılmıştır. LMC cihazı bilgisayar, mobil cihazlar başta olmak üzere diğer akıllı arabirimlere gömülü yada doğrudan entegre edilebilir olması, el ve parmak hareketlerini anlık tarama hızı ve hassasiyetinin yüksek olması ve gelişime açık olması sebebiyle özellikle tercih edilmiştir. Kinect ise LMC'nin yetersiz kaldığı durumlarda destek olması amacıyla kullanılmıştır. Bu çalışmada, LMC ve Kinect cihazı kullanılarak Türk işaret dili (TİD)'ne ait çift elli statik parmak alfabesi ve dinamik sözcük tanıma sistemleri üzerine çalışılmıştır. Çalışma önişlem, izleme, görüntü karelerinin toplanması, özellik çıkarımı, özellik seçimi ve çıkarımı, boyut indirgeme, eğitim ve test olmak üzere 8 aşamadan oluşmaktadır. Çalışma dört uygulamadan oluşmaktadır. Birinci uygulama LMC kullanılarak statik parmak alfabesi tanıma üzerine gerçekleştirilmiştir. İkinci uygulamada LMC ile dinamik sözcük tanıma sistemi tasarlanmıştır. Bu uygulama için kendi arasında benzerlik ve farklılıkları dikkate alınarak hazırlanan 50 dinamik sözcük için 4 denek kullanılarak veri kümesi oluşturulmuştur. Üçüncü uygulamada Kinect cihazı kullanılarak dinamik sözcük tanıma sistemi üzerinde çalışılmıştır. Dördüncü uygulama ise LMC+Kinect cihazları kullanılarak gerçekleştirilen farklı uzunluk ve sürelere sahip dinamik sözcük tanıma sistemidir. Bu uygulamada kullanılan histogram ve zamansal özellik çıkarımı veri kümelerinin boyutlarını eşitleyerek sınıflandırıcıya girişi kolaylaştırır ve aynı zamanda verilerin boyutunun küçültülmesini sağlar. Bu uygulamalardan elde edilen veri kümelerinden özellik seçim algoritması, özelik çıkarım yöntemleri ve PCA, LDA ve PCA+LDA boyut indirgeme yöntemleri kullanılarak yeni veri kümeleri elde edilmiştir. Bu veri kümeleri kullanılarak makine öğrenme yöntemlerinden geleneksel, sinir ağı ve ELM tabanlı sınıflandırıcılar ile işaret tanıma başarımı analiz edilmiştir. ELM mimarileri ilk kez bir işaret dili tanıma sisteminde sınıflandırıcı olarak bu çalışmada kullanılmıştır. Tanıma başarımı açısında gürbüz ve kararlı genelleme yeteneği sunan ELM'in 5 farklı mimarisi ve bunların kendine özgün öğrenme yöntemleri test edilerek sonuçlar karşılaştırılmıştır. Bu çalışma ile önerilen TİD tanıma sisteminin başarım testi 10 katlamalı çapraz doğrulama yöntemi kullanılarak yapılmıştır. Elde edilen başarım metriklerine dayalı olarak, ELM tabanlı mimari ve makine öğrenmesi yöntemlerinden ML-KELM'in tüm veri kümelerinde başarım oranını koruduğunu ve en yüksek başarım oranını verdiği ve başarım açısından kararlı bir yapıya sahip olduğu gözlemlenmiştir.