Programa de Pós-Graduação em Engenharia Elétrica e de Computação
Permanent URI for this communityhttps://repositorio.ufrn.br/handle/123456789/11949
Browse
968 results
Search Results
Doctoral Thesis Treinamento consciente de quantização vetorial evolutiva: um método adaptativo para compactar modelos de aprendizado de máquina(Universidade Federal do Rio Grande do Norte, 2026-02-06) Flores, Thommas Kevin Sales; Silva, Ivanovitch Medeiros Dantas da; Costa, Daniel Gouveia; https://orcid.org/0000-0003-3988-8476; http://lattes.cnpq.br/0200709054904002; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; https://orcid.org/0000-0003-2808-8529; http://lattes.cnpq.br/0630479458408181; Vieira, Tiago Figueiredo; http://lattes.cnpq.br/8601011832053651; Martins, Allan de Medeiros; https://orcid.org/0000-0002-9486-4509; http://lattes.cnpq.br/4402694969508077; Villanueva, Juan Moisés Mauricio; https://orcid.org/0000-0002-8760-9390; http://lattes.cnpq.br/1715030763709959; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941The consolidation of Large Language Models poses significant challenges to their deployment on edge devices, as the reliance on cloud-based infrastructures compromises data privacy, increases latency, and limits critical applications in scenarios with unstable connectivity. Traditional compression techniques, predominantly applied after training, prove inadequate in this context, as they may degrade model performance or rely on opaque execution pipelines, hindering the analysis and verification of system behavior. In response to these challenges, this thesis proposes Evolving Vector Quantization-Aware Training (EVQAT), whose central principle consists of integrating compression directly into the learning process. In this approach, quantization is no longer treated as a post-training step but rather as an adaptive mechanism during training, allowing the transformations imposed on model parameters to be explicitly controlled and analyzed. mThe methodology is grounded in evolutionary vector quantization theory by treating the weights of the language model as continuous data streams, enabling incremental clustering algorithms such as AutoCloud, Mean-Shift, and Affinity Propagation to dynamically adjust the codebooks in synchrony with the optimization of high-dimensional parameters, thereby enabling model compression with explicit control over quantization error. The integration of codebook evolution into the training process ensures that the partitioning of the vector space follows the learning dynamics, preserving relevant statistical properties and reducing the impact of quantization noise. This characteristic makes EVQAT suitable for quantizationsensitive architectures such as Transformers, extending its applicability from multilayer neural networks to generative language models under strict memory and computational constraints. The proposed approach is evaluated across multiple experimental scenarios, including training without quantization, quantization-aware training using evolutionary vector quantization and 8-bit integer (INT8) representations, as well as post-training quantization under the same variants, with deployments on microcontroller-based platforms. In this context, auditability is achieved through transparent access to model parameters, including the vocabulary, the weights of each hidden layer, and the attention mechanisms. As a result, the methodology enables the automatic generation of standard, self-contained C++ code, with real-time inference capabilities and direct execution on diverse hardware architectures, without reliance on third-party libraries or operating systems, ensuring explainability at the levels of training, model, and execution, as well as interoperability across platforms.Master Thesis Traçamento de raios sísmicos com múltiplas condições iniciais usando Redes de Kolmogorov-Arnold com Informações Físicas(Universidade Federal do Rio Grande do Norte, 2025-12-05) Marques, Lucas Torres; Barros, Tiago Tavares Leite; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941; https://orcid.org/0000-0001-9665-2238; http://lattes.cnpq.br/1321568048490353; http://lattes.cnpq.br/0281845604882859; Silva, Ivanovitch Medeiros Dantas da; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; Nose Filho, Kenji; http://lattes.cnpq.br/7933130617235231Seismic wave propagation modeling is fundamental for seismic imaging and reservoir characterization in the oil and gas industry. While ray tracing is an efficient technique, traditional methods face challenges in complex geological models due to instability and high computational cost, as they require solving for a single initial condition at a time. This work proposes the use of Physics-Informed Kolmogorov-Arnold Networks (PIKANs) to overcome these limitations by allowing the simultaneous tracing of multiple rays with different initial conditions. The proposed methodology includes velocity model smoothing with B-spline filters, the use of a compound loss function, and the application of adaptive weights in the physics-informed loss function to enhance model training. Two experiments were conducted using synthetic data: the first with a parameterized velocity model, and the second with a non-parameterized model widely used as a benchmark in seismic processing, known as the Marmousi model. The results demonstrate the consistency of the approach, with R2 values close to 1 for the first experiment and exceeding 0.90 for the second. Qualitatively, the method accurately traces seismic rays, even in regions not used for training and in areas with rapid spatial velocity variation, proving to be more precise at predicting the ray’s spatial coordinates than the slowness vector.Master Thesis Substituições de aminoácidos como representação de mutações em transformers para análise e classificação de variantes do SARS-CoV-2(Universidade Federal do Rio Grande do Norte, 2025-12-19) Silva, Alessandro Soares da; Fernandes, Marcelo Augusto Costa; https://orcid.org/0000-0001-7536-2506; http://lattes.cnpq.br/3475337353676349; http://lattes.cnpq.br/6391394214446756; Câmara, Gabriel Bezerra Motta; https://orcid.org/0000-0002-5736-0782; http://lattes.cnpq.br/3849103334728892; Terrematte, Patrick César Alves; https://orcid.org/0000-0002-0385-0030; http://lattes.cnpq.br/4283045850342312; Barbosa, Raquel de Melo; https://orcid.org/0000-0003-3798-5512; http://lattes.cnpq.br/8174411728088087The COVID-19 pandemic, caused by SARS-CoV-2, highlighted the need for scalable and automatable computational methods for genomic surveillance in the face of the continuous emergence of variants of concern. In this context, this work proposes an innovative approach based on vector representations of amino acid substitutions, leveraging Transformer models for the analysis and classification of SARS-CoV-2 viral variants, with a focus on the Alpha, Beta, Gamma, Delta, and Omicron variants. Sequences of amino acid substitutions were processed into high-dimensional embeddings, which supported two complementary experiments. In the first, unsupervised techniques, such as Fuzzy C-Means clustering and t-SNE projection, were applied to reveal groupings consistent with known variants, as well as transitional regions and ambiguous samples. In the second, supervised learning models, including SVM, Random Forest, k-NN, and XGBoost, were evaluated for variant classification, with XGBoost achieving accuracy above 99% on an external dataset. The results demonstrate that representations derived exclusively from amino acid substitutions are capable of discriminating viral variants with high performance and capturing mutational signatures of biological relevance, without requiring genomic alignment. Thus, the proposed approach represents a scalable, flexible, adaptable, and purely computational alternative for automated pathogen genomic surveillance, with potential applications in public health scenarios.Master Thesis Musicalization with educational robotics(Universidade Federal do Rio Grande do Norte, 2025-12-19) Andrade, José Lindenberg de; Gonçalves, Luiz Marcos Garcia; Alsina, Pablo Javier; http://lattes.cnpq.br/3653597363789712; https://orcid.org/0000-0002-7735-5630; http://lattes.cnpq.br/1562357566810393; http://lattes.cnpq.br/8823342043920921; Curvelo, Carla da Costa Fernandes; http://lattes.cnpq.br/2329425562585943; Souza, Anderson Abner de Santana; https://orcid.org/0000-0001-6353-8674; http://lattes.cnpq.br/2563070123322776The integration of music and educational robotics remains underexplored, even though both areas offer strong potential to promote active, interdisciplinary, and meaningful learning. This work investigates how musicalization can be incorporated into robotics to enhance student engagement and support the practical understanding of musical, elec tronic, and computational concepts. To this end, the study proposes EcoMusic, an ecolog ical musical device built from electronic components and recycled materials, programmed to play melodies through interactive buttons. The research is grounded in the competen cies of Brazil’s National Common Curriculum (BNCC) and in theoretical foundations from music education, robotics, sustainability, and creative learning. The study was con ducted with prior authorization from the Research Ethics Committee of the Federal Uni versity of Rio Grande do Norte (UFRN), ensuring compliance with ethical standards for research involving human participants. EcoMusic was implemented in a public school with 25 students aged 11 to 17, over two instructional sessions that introduced concepts of sound and frequency, programming, device assembly, and musical experimentation. Data collection involved questionnaires administered to both students and the classroom teacher. Results indicate that integrating musicalization into robotics increases motiva tion, stimulates curiosity, facilitates understanding of musical concepts, and strengthens skills such as collaboration, creativity, and logical thinking. The use of recycled ma terials also encouraged reflections on sustainability and conscious consumption. The findings demonstrate that EcoMusic is an accessible, interdisciplinary, and curriculum aligned pedagogical tool that expands opportunities for active experiences in music and technology within basic education.Master Thesis Uma metodologia baseada em agentes com avaliação autônoma para extração de conhecimento em textos técnicos(Universidade Federal do Rio Grande do Norte, 2025-12-19) Andrade, Matheus Gomes Diniz; Silva, Ivanovitch Medeiros Dantas da; https://orcid.org/0000-0002-0116-6489; http://lattes.cnpq.br/3608440944832201; https://orcid.org/0009-0002-0268-2247; http://lattes.cnpq.br/7308297435660284; Brandão, Dennis; https://orcid.org/0000-0003-1558-0581; http://lattes.cnpq.br/6838931677289559; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941; Silva, Marianne Batista Diniz da; https://orcid.org/0000-0002-8277-7571; http://lattes.cnpq.br/6470261020797104In the current scenario of Operational Technology (OT) and Information Technology (IT) convergence and Industry 5.0, the maintenance and resilience of OT systems depend on knowledge continuity. The difficulty and inefficiency in accessing the dispersed technical knowledge within unstructured documentation of legacy systems, such as the PROFIBUS (Process Field Bus) protocol, critically impact maintenance and decision-making. Therefore, this work investigates the use of Generative Artificial Intelligence architectures to retrieve and synthesize this legacy industrial knowledge. The study compares the performance of three distinct configurations: (i) a standalone LLM (Baseline), (ii) a Single-Agent RAG (Retrieval-Augmented Generation) model, and (iii) a Multi-Agent RAG architecture. Through an experiment based on PROFIBUS technical texts, performance was evaluated using quantitative metrics (ROUGE, BERTScore) and qualitative assessment (LLM-as-a-Judge). The results demonstrate that Multi-Agent orchestration establishes a superior performance hierarchy. The collaboration and specialization among agents significantly increase the contextual precision and factual accuracy of the responses, while simultaneously reducing hallucination to 80% faithfulness to the context. The study validates the Multi-Agent RAG as a necessary architectural pattern to guarantee reliability in the transfer of engineering knowledge, providing a pathway for human-centric automation in Industry 5.0.Doctoral Thesis Control of an automated insulin delivery system using artificial intelligence(Universidade Federal do Rio Grande do Norte, 2025-10-02) Farias, João Lucas Correia Barbosa de; Bessa, Wallace Moreira; https://orcid.org/0000-0002-0935-7730; http://lattes.cnpq.br/3256782908311485; https://orcid.org/0000-0002-3324-2239; http://lattes.cnpq.br/0575163993612915; Bertachi, Arthur Hirata; http://lattes.cnpq.br/2283634850801378; Cunha Júnior, Américo Barbosa da; https://orcid.org/0000-0002-8342-0363; http://lattes.cnpq.br/5659403706694491; Dórea, Carlos Eduardo Trabuco; https://orcid.org/0000-0002-3999-2874; http://lattes.cnpq.br/0143490577842914; Medeiros, Philippe Eduardo de; https://orcid.org/0000-0002-1803-241X; http://lattes.cnpq.br/5594617038786753Type 1 Diabetes Mellitus is a disease that affects millions of people worldwide. Recently, devices that automatically regulate blood glucose concentration in diabetic patients were developed. The Automated Insulin Delivery (AID) system can provide a life with higher quality, autonomy, and comfort for patients. This work aims to design an intelligent nonlinear controller with an estimator developed with Artificial Intelligence (AI) algorithms for an AID system. An extensive review on diabetes is conducted introducing its definition, physiological context, diagnosis, global impact and available state of the art therapies. Furthermore, a literature review is carried out presenting the types of AID systems, current challenges, virtual patient models, control strategies and AI algorithms that shape modern designs. The OHSU virtual patient model, used to simulate the virtual patient population in this work, is introduced along with its mathematical formulation, which incorporates physical activity to the insulin-glucose dynamics. The proposed controller combines Feedback Linearization with an intelligent neural network-based estimator to compensate for model uncertainties, external disturbances and time delays inherent of the biological system in question. It incorporates a Reinforcement Learning algorithm, Q-learning, to optimize neural network learning rate across a diverse virtual patient population while minimizing hypoglycemic events. Results show the controller was able to safely and effectively regulate glycemia with average Time in Range (TIR) of 81% and Coefficient of Variation (CV) of 35% on a diverse training population of 10 subjects, and average TIR of 76% and CV of 34% for a different unseen validation population of 5 subjects. In order to avoid poor control during the network’s initial learning period, offline network pre-training is proposed and tested on the validation population. Severe hyperglycemic events in the early days of simulation were completely avoided with the added benefit of a slight improvement in TIR (76% vs. 78%), indicating a clear path on how to introduce the controller to real-life subjects. Overall, the proposed controller was able to provide safe and effective glycemic control under exercising across a diverse virtual patient population without the need for user input on meal intakes or physical activities.Doctoral Thesis A Knowledge Graph-Based approach for modeling legislative texts: representation and document similarity analysis(Universidade Federal do Rio Grande do Norte, 2025-08-29) Oliveira, Gisliany Lillian Alves de; Silva, Ivanovitch Medeiros Dantas da; https://orcid.org/0000-0002-6037-9660; http://lattes.cnpq.br/8193679068029541; Viegas, Carlos Manuel Dias; Villanueva, Juan Moisés Mauricio; Oliveira, Luiz Affonso Henderson Guedes de; Barros, Thiago MedeirosThe most prominent task of the Legislative Branch — lawmaking — depends on a complex and demanding process in which new proposals must be examined, debated, and revised in light of existing legislation. These activities are often labor-intensive for humans due to the technical language, substantial length, and interdependence of legal texts. At the same time, these characteristics present a tangible opportunity for Artificial Intelligence (AI), particularly through the integration of Natural Language Processing (NLP) and structured data representations. Aiming to model legislative documents in a way that preserves their rich structural semantics, this work proposes an approach for transforming legislative texts into domain-specialized Knowledge Graphs (KGs) that capture their inherent hierarchical organization. The proposed method, based on LexML standards — a Brazilian XML schema for legal documents — extracts explicit structural relationships (e.g., articles, paragraphs, items) and augments them with contextual entities and relationships extracted by a Large Language Model (LLM). The resulting KGs, stored in a Neo4j database, capture both the internal topology of legal texts and their semantic nuances, enabling structured representations that support more meaningful analysis than unstructured raw text. To assess the effectiveness of this approach, comparative experiments were conducted on document similarity tasks, a core component of legislative workflows. Three scenarios were evaluated: (i) a text-only baseline using BERT-based sentence embeddings averaged across document sections; (ii) structure-aware KGs encoded via FastRP and GraphSAGE embeddings; and (iii) contextually enriched KGs also encoded via FastRP and GraphSAGE embeddings. Results using legislative proposals from the Legislative Assembly of Rio Grande do Norte (ALRN) show that while the text-based model achieved the highest precision, recall, and F1-scores, the KG-based representations provided interpretable, structure-driven insights. Contextual enrichment improved FastRP’s performance over structure-only graphs, while GraphSAGE performed best with structure-only representations, suggesting that LLM-derived relations may have introduced semantic noise for this graph model. Although the KGs were inherently heterogeneous, homogeneous graph algorithms were applied for simplicity, which may have limited performance. Nevertheless, the results demonstrate the feasibility of converting legislative documents to KGs, and the inclusion of structural information in embeddings was achieved, demonstrating potential for future improvements via heterogeneous models, advanced pooling strategies, or self-supervised pre-training. By bridging NLP and graph-based AI, this work advances approaches for legal document modeling, offering a pipeline for document similarity analysis and improving the legislative process efficiency.Doctoral Thesis Características de projeto ótimo de um sistema de geração híbrido baseado em máquina de indução com acionamento por meio de um Regulador Eletromagnético de Frequência(Universidade Federal do Rio Grande do Norte, 2025-10-17) Crisóstomo, Daniel Carlos de Carvalho; Medeiros Júnior, Manoel Firmino de; https://orcid.org/0000-0002-9408-3948; http://lattes.cnpq.br/4960078797028638; https://orcid.org/0000-0002-9752-688X; http://lattes.cnpq.br/5456416955432542; Medeiros, Arthur Salgado de; Pinheiro, Ricardo Ferreira; Ramos, Thales Augusto de Oliveira; Bezerra, Ubiratan HolandaThe need for the implementation of policies aimed at energy transition drives the renewable energy sector, prompting increased investments in this field. Globally and nationally, data reveal significant growth in the incorporation of wind and solar energy into power grids, fostering further technological advancements to ensure efficiency and quality in the generation process. This study optimizes a hybrid wind generation system using an Electromagnetic Frequency Regulator (EFR), aiming to minimize electrical losses, reduce component weight, and decrease reliance on secondary sources. The optimization explores the relationship between the number of poles in the generator and EFR, as well as the gearbox transmission ratio. The methodology combines numerical simulations, metaheuristic optimization (Ant Colony Optimization), electric machine design tools, and financial analysis. Results demonstrate that the optimal configuration heavily depends on the pole numbers and gearbox ratio, which directly affect the electrical parameters of the EFR and induction generator. Five topologies were analyzed, with the last proving the most efficient under the proposed criteria. The study provides a scalable methodology to assess the technical and economic feasibility of hybrid systems in diverse locations, contingent on wind data and generator specifications.Master Thesis Trilhas para auditoria aplicada ao sistema único de saúde no contexto das órteses, próteses e materiais especiais (OPMES): estruturação de uma base de dados integrada a partir do projeto fiscalizasus(Universidade Federal do Rio Grande do Norte, 2025-06-26) Alves, Luca Pareja Credidio Freire; Valentim, Ricardo Alexsandro de Medeiros; https://orcid.org/0000-0002-9216-8593; http://lattes.cnpq.br/3181772060208133; http://lattes.cnpq.br/8782837208577002; Valentim, Janaína Luana Rodrigues da Silva; https://orcid.org/0000-0001-8525-8155; http://lattes.cnpq.br/8236259645905686; Santos, João Paulo Queiroz dos; https://orcid.org/0000-0002-9130-7723; http://lattes.cnpq.br/2413250851590746; Coutinho, Karilany Dantas; https://orcid.org/0000-0002-2051-8611; http://lattes.cnpq.br/8409211766785367; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941; Rodrigues Júnior, Methanias Colaço; Costa, Theo Duarte da; https://orcid.org/0000-0002-9355-8382; http://lattes.cnpq.br/8305343735444335Background: The Unified Health System (SUS) was a historic milestone established by the 1988 federal constitution. However, over the years the SUS has undergone several transformations, especially concerning the amount of data generated about its internal processes. The National Audit Department of SUS (DenaSUS) was created to manage the SUS's internal control and audit processes. As a result, new demands regarding auditing in the SUS have emerged. One of these activities was organized in partnership between DenaSUS, the Federal University of Rio Grande do Norte (UFRN), the Federal Institute of Rio Grande do Norte (IFRN), the Federal Prosecution Office of Rio Grande do Norte (MPF/RN), the Federal Court of Auditors (TCU) and other competent authorities to build the FiscalizaSUS project, which aims to provide intelligent methods for analyzing large volumes of data in the field of health, with the ultimate goal of delivering audit trails focused on Regulation and Special Orthotics, Prosthetics and Special Materials (OPSM) in the SUS domain, as these are inputs that have already been identified as objects of interest in corruption scandals involving health professionals, as was the case with the “máfia das próteses”.Goal : Thus, the objective of this work is to demonstrate how the application of statistical methods and Big Data can enhance the feasibility for developing OPSM audit trails through the integration of public and private databases. Methods: On the first hand, a systematic mapping of the literature was conducted in order to evaluate which methods are being applied for Big Data-based health audit research. From this point of view, it was possible to analyse and adopt approaches to develop a relational model for OPSM audit trials, and also develop metrics based on statistical analysis of the collected data samples. Results : The Systematic Literature Mapping (SLM) initially included a total of 344 articles, obtained from the search string used. This left 40 articles for evaluation and data collection, after the inclusion and exclusion steps. Heterogeneous solutions were identified to address the issue of health auditing, presenting techniques also related to computational methods and Big Data approaches.From an operational perspective, the tables were organized oriented towards a star schema, containing fact and dimension tables, which enable the integration of different data from various sources, also concerning the goals of each audit trail and assessing the presence of common values between the databases. Nevertheless, the integration between the databases also provided a visualization layer by using Apache Superset’s virtual interface, allowing a comprehensive view of the anomalies through geographical location, links between the data sources, and also through the proposed metrics. Conclusion: The results allow us to conclude that the integration of different data sources related to OPSM can contribute to the identification of irregularities in transactional processes involving this type of medical supply. In addition, the Big Data architecture provided by the FiscalizaSUS project also made it possible to aggregate these different data sets and brought new perspectives to the health audit trails within the scope of the project.Doctoral Thesis Deep Learning Applications in Geophysics: Data Matching and Inversion(Universidade Federal do Rio Grande do Norte, 2025-08-26) Perez, Katerine de Jesus Rincon; Souza, Samuel Xavier de; https://orcid.org/0000-0001-8747-4580; http://lattes.cnpq.br/9892239670106361; http://lattes.cnpq.br/0756224243207855; Carmo, Bruno Souza; https://orcid.org/0000-0002-2486-7026; http://lattes.cnpq.br/8129545750278083; Corso, Gilberto; https://orcid.org/0000-0003-1748-4040; http://lattes.cnpq.br/0274040885278760; Senger, Hermes; http://lattes.cnpq.br/3691742159298316; Araújo, João Medeiros de; https://orcid.org/0000-0001-8462-4280; http://lattes.cnpq.br/3061734732654188; Oliveira, Luiz Affonso Henderson Guedes de; https://orcid.org/0000-0003-2690-1563; http://lattes.cnpq.br/7987212907837941This work applies Deep Learning (DL) methodologies to solve geophysical problems, with a focus on Time lapse seismic, an essential technique for monitoring changes in reservoirs and supporting strategic management, using synthetic pre-salt seismic data acquired with Ocean Bottom Nodes (OBN). The first approach focused on mitigating one of the main limitations of TL seismic: data non-repeatability between surveys, caused by various environmental factors and equipment positioning variations, which can mask real reservoir changes. To address this challenge, we proposed Conditional Generative Adversarial Networks (cGANs). This methodology was developed to correct non-repeatability effects, improving data quality before applying Double Difference Full Waveform Inversion (DDFWI). As a second application, designed to estimate velocity variations and serve as a technique parallel to inversion, we constructed a synthetic dataset representing scenarios with different velocity anomalies. A Convolutional Neural Network (CNN) was trained using synthetic baseline and monitor seismograms generated from a pre-salt velocity model. The network received the time-lapse difference between baseline and monitor data as input, and produced the corresponding velocity model as output. To benchmark the approach, conventional Double-Difference Full Waveform Inversion (DDFWI) was also performed, providing a reference for computational cost, efficiency, and inversion quality. Results demonstrate that the CNN achieved accurate inversion performance while delivering significantly higher computational efficiency compared to the traditional DDFWI workflow. Both scenarios were trained with 80% of the data and tested with the remaining 20%. The predictions were evaluated using quantitative metrics such as the mean squared error (MSE), the normalized root mean squared error (NRMS), and the structural similarity index (SSIM), indicating a good inference capacity of the model.