42
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Archived Theses
Design of complex-geometry parts for multi-axis robotic additive manufacturing technology and its simulation
This thesis presents the design, development, and simulation of complex-geometry parts for multi-axis Robotic Additive Manufacturing (RAM). A Rhino–Grasshopper–KUKA | prc software environment is integrated to generate customized multi-axis toolpaths for geometries that conventional CAD software cannot easily produce. A KUKA KR210 R2700-2 anthropomorphic robot with a positioner is used for simulation, while Altınay Robotics provided an identical robot (without a positioner) for the physical implementation of the generated trajectories. The RAM system is intended for producing intricate geometries using both polymer extrusion and metal deposition. To identify the most suitable method, the study evaluates Wire Arc Additive Manufacturing (WAAM), Wire Laser Additive Manufacturing (WLAM), and Wire-Laser Hybrid Additive Manufacturing (WLHAM) in terms of structural efficiency, deposition quality, and integration feasibility. A multi-color polymer pellet extruder was co-developed with Altınay Robotics and MEF University interns under the supervision of the author and Altınay's mechanical manager, demonstrating multi-material printing. Additionally, an industrial Nolega polymer extruder was procured with MEF BAP research funding for future use. Beyond its technical contributions, this study emphasizes the integration of robotics and additive manufacturing within a unified production workflow. The research adopts a multidisciplinary and collaborative framework, combining engineering, design, and computational modeling to advance manufacturing education and practice. Experimental studies were conducted in collaboration with students from diverse disciplines, fostering hands-on learning and cross-domain knowledge exchange. This integrative approach highlights the thesis as a convergent study bridging theory, practice, and education through robotics-driven digital fabrication. The outcome is a globally competitive RAM technology offering scalable, precise, and sustainable solutions for advanced manufacturing in both engineering and creative fields.
Anayasa Mahkemesi Kararlarının Simülasyonu: Türk Bireysel Başvuruları İçin Çok Etmenli Bir Büyük Dil Modeli (LLM) Çerçevesi
This research examines the ability of large language models (LLMs) to emulate judicial decision-making in constitutional court cases. We used three cutting-edge models—GPT-5, Gemini, and Claude—to look at 343 decisions made by the Turkish Constitutional Court between 2014 and 2024. We did this using a two-stage evaluation framework that mirrored how courts really work. The study evaluated model efficacy in both admissibility determinations and substantive rights infringement judgments. During the admissibility stage (Stage 1), the accuracy rates varied from 68.80% for Claude to 81.34% for GPT-5, with majority voting achieving 79.59%. GPT-5 had the fewest total mistakes and a balanced approach, while Gemini and Claude were more likely to think that something was not admissible. On the other hand, courts had a more moderate acceptance rate. In Stage 2, when rights were violated, all three models had the same accuracy of 81.50%. However, majority voting did better, with an accuracy of 83.24%. In this case, GPT-5 tended to have partial matches, Gemini had the most exact matches, and Claude was in the middle, showing that each had different strengths in legal reasoning. Patterns of inter-model agreement showed that there was a lot of convergence, but it wasn't always the same. In Stage 1, unanimous agreements, though less common, had the highest accuracy (87.32%), showing that consensus decisions are reliable. In Stage 2, GPT-5 and Claude were the most in line with each other (88.52%). These results indicate that ensemble methods and hybrid human–AI approaches could improve the consistency and robustness of judicial decision-making. The results show that even general-purpose LLMs can understand complicated constitutional principles and come up with structured, court-like reasoning that is serious enough to be used in legal situations. Although existing models struggle to replicate the comprehensive intricacies of judicial reasoning, the persistent superiority of ensemble methodologies suggests that specialized legal AI systems may exceed general models, potentially revolutionizing constitutional jurisprudence by improving efficiency, consistency, and accessibility to justice
Performance evaluation of llm based chatbots with E2e method:LLama-8b,LLama-7b,Gemma-7b and mistral-7b
This study investigates the performance of large language models (LLMs) within the context of customer support chatbots by employing an end-to-end (E2E) evaluation framework. Specifically, it compares three prominent open-source models (Gemma-7B, Mistral-7B, Llama-7B and Llama-8B) based on their ability to comprehend and respond to user queries in a meaningful and accurate manner. The chatbot application under review was designed to provide assistance on an educational content platform and was tested using over 3000 curated question-answer pairs. The evaluation combines both semantic and lexical metrics, using cosine similarity to measure the alignment of model responses with expert-written answers, and ROUGE metrics to assess word-level accuracy. Additionally, the study incorporates prompt engineering techniques and analyses how models handle random or off-topic inputs, providing a comprehensive view of their reliability and contextual sensitivity. Results indicate that Gemma-7B and Llama-8B performs most consistently across all metrics, while Mistral-7B offers balanced outputs with occasional variance. Llama-7B, although structurally robust, struggled to deliver semantically aligned and contextually appropriate responses. Overall, the findings highlight the practical implications of model selection for real-world chatbot deployments and demonstrate the importance of multi-dimensional evaluation methods when assessing LLM performance in customer interaction settings.
Image classification of customs procedure documents using machine learning and deep learning models
The images of documents used in customs procedures were classified using machine learning and deep learning models. As image classification methods, Gaussian Naive Bayes, Random Forest, and Support Vector Machine were used in the machine learning field, and Convolutional Neural Network, Vision Transformer, ConvNext and EfficientNetV2 were used in the deep learning field. Datasets were organized specifically for each model type, and models were created with parameter values determined to ensure the models worked at the optimum level. Training and testing were conducted with the relevant datasets. The models were compared in terms of training times and accuracy rates. The best results were achieved with Convolutional Neural Network, which had a training time of approximately 11 minutes and an accuracy rate of 98,32%; second-best accuracy result achieved with Random Forest, which had an accuracy rate of 97,94% and second-best overall results achieved with EfficientNetV2, which had a training time of approximately 22 minutes and an accuracy rate of 95,89%.
Pid-kontrollü yüz takip iha'sinin performans analizi
This thesis examines closed-loop performance of a vision-based tracking controller on a DJI Tello micro-UAV using classical Proportional–Integral–Derivative (PID) control. We perform step-response system identification for four axes—yaw, lateral (left–right), longitudinal (forward–backward), and vertical (up–down)—by logging camera-derived error, controller outputs (RC commands), and vehicle states. Axis-wise SISO models indicate that yaw behaves as a pure integrator, while the translational axes are well captured by type-1 (integrator-with-lag) dynamics. We implement P, PI, PD, and PID controllers under identical test scripts and evaluate percent overshoot and settling time within a ±5% band from flight logs sampled at 30 Hz. Results show PD provides the best trade-off on translational motion—reduced overshoot with shorter settling—whereas yaw is adequately regulated by proportional action alone. Integral action without anti-windup causes saturation and oscillation; adding back-calculation anti-windup and a first-order derivative filter improves robustness. Contributions are: (i) reproducible axis-wise models and tests, (ii) a unified logging/plotting pipeline for fair controller comparisons, and (iii) practical guidance on derivative filtering and anti-windup for small UAVs with vision feedback. These findings establish a quantitative baseline for transitioning from classical PID to more advanced control on resource-constrained aerial platforms.
Beyoğlu'nda bir rota üzerinde binaların yaşamını ve ölümünü yeniden düşünmek
This study has emerged on the neglected and abandoned structures left to decay, which have become a familiar sight in almost every city, turning into a common image. Focusing on the tendency in architecture to attribute human characteristics, the study delves into the question of how a building can be subject to death when confronted with the endings of buildings metaphorically likened to living beings. The idea that the acceptance of death can be a driving force for positive actions is emphasized in the study. This study aims to cover and document the life and death of buildings by addressing global issues and emphasizing the impact of the concept of death on human life. This is realized through a personal observation trip on a specific trail in Beyoğlu, conducted in January 2023. The findings are reflected through mapping, recorded with photos, and highlighted with drawings through photos. The first section includes the conceptual definition of death and discusses the perspective of individuals and coping methods in the face of inevitable ends, whether it be death, uselessness, or failure. This section, stemming from the notion of the impact of death on human life, also provides a perspective for observing buildings encountered along the designated route for the study. The second section examines the current state of buildings located on the route determined in Istanbul's Beyoğlu district. During the semi-structured observation trip conducted in January 2023, 130 locations were marked and classified on the map based on the diagnoses obtained through analogies. This section provides explanations of the background of diagnoses and analogies, emphasizing the distinct traces on buildings toward death through drawings based on some of the marked structures.
Changing ideals continuing visions: prefabricated micro living unit experiments from 1960s-1970s and 2000s-2010s
In this research, prefabricated micro living unit experiments are examined over two historical periods, 1960s-1970s and 2000s-2010s. The aim of the study is to reveal the different motivations that resulted with a common production on the same housing typology in the context of socio economic, technological developments and popular culture. In addition to questioning whether the productions in the 1960s-1970s did turn into a common housing typology in their period or not, questioning the successful and unsuccessful aspects of these projects as well as identifying continuities and changes in the 2000s-2010s form the basis of this research. One of the aims is to reveal what role 1960s-1970s productions played in the 2000s-2010s. In order to achieve this aim, first of all, the historical development of prefabricated and micro living units was examined separately and the emergence of prefabricated micro living units as a housing typology was examined. Later, in chapter two the dynamics of the two historical periods are addressed separately through socio-economic, technological developments and popular culture and case studies of each period are introduced. Continuing visions and changing ideals between two historical periods spanning 1960s-1970s and 2000s-2010s are examined in chapters three and four. And finally, Volu-te, a prefabricated micro living unit developed under the Alternative Architectural Practices master's program, is located within the context of prefabricated micro living units in the 2000s-2010s
The story of trailer and its effective actors; as a mobile home case
"House" is example of a place, while "person" exemplifies a user. The relationship between the two entities culminates in the creation of a "home," which persists as long as the relationship endures. This study seeks to explore the nature of this relationship in the context of trailers, which serve as examples of mass-produced mobile homes. The trailer's features, such as mobility, minimalism, and mass-production, set it apart from traditional houses with respect to the user's relationship. A comprehensive perspective has been obtained by delving into the trailer's lifelong history and examining it within the sociological context of different periods.The study is confined to the United States, where trailers were born and evolved. The trailer has undergone several transformations over time, responding to the critical impacts of the era and reaching its present form. The study is divided into four sections, each covering a distinct period in history: the wagon era, the post-Ford T era, the post-World War II mobile home era, and the post-technology RV era. The trailer has undergone both minor and significant transformations, with individual actors, cultural influences, and crises identified as the primary drivers of these changes. In conclusion, the relationship between place, people, and crisis is questioned. In this study, the trailer is regarded as a case of place, and the necessary attributes for creating and transforming a place is discussed. Finally, speculation on the future is made based on the findings.
Customer churn prediction for the Pay-TV sector
Understanding the reasons for customer churn provides added value in terms of retaining existing customers, as customer attrition leads to revenue loss for companies and incurs marketing costs for acquiring new customers. In this study, the 6-month historical data of a Pay-TV company operating in Turkey was used, and due to the imbalanced nature of the dataset on a label basis, the oversampling method was applied. During the model development phase, various artificial learning algorithms (Random Forest, Logistic Regression, K-Nearest Neighbors, Decision Tree, AdaBoost, XGBoost, Extra Tree Classifier) were utilized, and their performances were compared. Based on the evaluation of success criteria for each model, it was observed that the tree-based Random Forest, Extra Tree Classifier and XGBoost achieved the highest performance for this dataset.
Fraud detection and prediction with machine learning applications
The main purpose of this study is to determine the fraudulent activities on transactions of the customers of a company that is active in the factoring sector, and accordingly, to capture measurable parameters with exploratory data analysis based on the historical transaction and connection data of the customers, and then to perform predictive models for the target. A hit rate of around 79% was achieved in XGBoost and CATBoost models, which are classification model algorithms. In this way, it is aimed to directly detect fraudulent activities on a trasnaction basis by acting in a more effective, efficient and correct approach after detecting the customer that shows high potential to make fraud.