FCT: DI - Dissertações de Mestrado
URI permanente para esta coleção:
Navegar
Entradas recentes
- A DECENTRALIZED MEMBERSHIP SERVICE FOR ENVIRONMENTS WITH DYNAMIC LINKSPublication . Pereira, João Gonçalo Freitas; Leitão, JoãoSwarm computing is becoming increasingly relevant in the context of distributed systems by supporting the coordination of a large number of autonomous devices in highly dynamic environments. This model is of particular interest for satellite swarms, where intermittent connectivity and rapidly changing topologies present significant challenges. One of the fundamental components required for the efficient operation of such systems is a reliable membership service that ensures nodes are aware of the active participants in the system. However, current membership services are primarily designed for static or minimally dynamic environments and often fail to adapt to the unpredictable conditions found in large-scale satellite constellations. Moreover, existing tools for emulating these conditions are insufficient for accurately modeling selective communication patterns and intermittent connectivity. This dissertation presents a decentralized membership service combined with a visi- bility model designed to capture the unique dynamics of satellite swarms. The solution was implemented and evaluated through large-scale experiments on a cloud-based infras- tructure, systematically scaling the network up to 170 nodes to assess convergence speed, message dissemination efficiency, and communication reliability. The results demon- strate that the proposed solution operates effectively in moderately sized satellite swarms, achieving stable convergence and reliable dissemination under dynamic conditions, while also revealing scalability limitations as the network size grows. These findings confirm the feasibility of a visibility-aware decentralized membership service for satellite networks and highlight opportunities for future improvements to support larger and more demanding swarm deployments.
- Prediction of debris state Based on Neural Ordinary Differential EquationsPublication . Dyreby, Katarina Elisabeth Tomé; Soares, CláudiaThe space environment has become increasingly crowded, with space debris in Low Earth Orbit expected to grow exponentially in the coming years. Alongside the surge in satellite launches, this congestion raises collision risks and highlights the need for scalable avoidance measures to prevent further debris generation. Reliable orbit prediction is therefore essential for ensuring operational safety and prolonging satellite lifespans. Satellite trajectories are shaped by external forces, some stochastic in nature and thus impossible to model perfectly. This introduces cumulative errors in orbital propagation, further compounded by limited knowledge of satellite properties and simplified force models. This Master’s thesis introduces an approach to Orbit Prediction using Neural Ordinary Differential Equations, with Physics Informed Neural Networks as a physics informed baseline and a Multilayer Perceptron as a non physics informed baseline for comparison. Relying on information about the space environment and the characteristics of satellites, the proposed approach also uses historical trajectory data to propagate a satellite’s state vectors. In case of the Neural Ordinary Differential Equations this is done by modeling the underlying differential equation of motion governing the satellite’s orbit and using an ordinary differential equation solver to propagate its trajectory forward in time. The results show that this approach effectively combines numerical integration techniques, which provide high accuracy, withNeuralNetworks, which capture complex patterns and generalize to unseen scenarios. As a result, it outperforms both baseline models by over 100 km, achieving a position mean RMSE of 0.25 km. By successfully modeling orbit dynamics, this approach offers a high-fidelity solution for Orbit Prediction, ultimately contributing to safer and more sustainable satellite operations. This thesis has contributed two publications: one presented at the International Astronautical Congress 2025, corresponding to the preliminary material from Chapters 5 to 6, and another at NeurIPS 2025, based on Chapter 4. In addition, a journal paper is in preparation covering the developed material from Chapters 5 to 6.
- Gesture classification for assembly tasks with event camerasPublication . Estróia, Pedro Afonso Fulgêncio; Soares, Cláudia; Valdeira, Filipa; Mendes, NunoIn recent years, the use of collaborative robots in manufacturing environments has in- creased significantly, enabling humans and robots to work together in shared environments. To ensure efficient cooperation, robots must be capable of detecting and predicting the actions performed by human operators. Traditionally, this detection is performed us- ing frame-based cameras (e.g., RGB cameras). These cameras capture a vast number of frames, requiring substantial bandwidth and processing time, which is a limitation in real-time detection of human gestures. Event cameras have emerged as a promising alternative, as they only capture changes in luminosity, significantly reducing storage and processing requirements. However, they remain an understudied field in the context of manufacturing. We propose a gesture classification model for assembly tasks using data captured with an event camera. This event data is from the Event-based Dataset of Assembly Tasks (EDAT24), which contains recordings of primitive tasks (such as picking, placing, and screwing) in a simulated industrial environment. First, with the goal of comparing between frames and events, we built a frame-based model (CNN–LSTM) trained on conventional frames and event frames created by aggregating events over short intervals—we use identical architectures and training to isolate the effect of the input representation. Then, we implemented an event-based PointNet++ pipeline that treats each video as a sequence of 3D point clouds of events, extracts the features hierarchically from each cloud, and then aggregates the classifications. We evaluated each of these models using several metrics. Frame-based and aggregated- event CNN–LSTM variants achieved very similar performance, with both reaching an accuracy of 84%, indicating that aggregation of events does not bring improvements over standard frames. In contrast, the PointNet++ pipeline substantially outperformed both, reaching an accuracy of 97%, while avoiding event-to-frame conversion and better exploiting the temporal structure of event data. Therefore, this research suggests the use of event data for classification of assembly tasks in manufacturing environments.
- Desenvolvimento de um módulo de visualização para simulador da florestaPublication . Amorim, João Miguel Lobão Santos Rodrigues; Birra, Fernando; Pires, João; Tomé, MargaridaAs florestas desempenham um papel central na biodiversidade, na regulação do clima e no desenvolvimento económico e social. Contudo, enfrentam crescente pressão das alterações climáticas, da desflorestação e da intensificação das atividades humanas. Em Portugal, são um recurso económico relevante, com impacto em setores como a produção de madeira e cortiça, cujo potencial depende de uma gestão florestal eficiente e sustentável. Neste contexto, as ferramentas de apoio à gestão florestal assumem uma importância cada vez maior, sobretudo quando disponibilizadas através de interfaces com elevados níveis de usabilidade e acessibilidade. Em Portugal, existem ferramentas dedicadas à gestão florestal, como o simulador de povoamentos standsSIM 1, que permite a simulação da evolução dos povoamentos sob diferentes regimes de gestão florestal. No entanto, estas ferramentas carecem, em geral, de componentes de visualização que facilitem a interpretação dos dados de simulação. A ausência de representações gráficas adequadas limita a compreensão dos resultados por parte dos utilizadores, comprometendo a sua eficácia na tomada de decisão. Este trabalho apresenta uma solução para este problema, através do desenvolvimento de um módulo de visualização para o simulador standsSIM 1. Este módulo permite a visualização tridimensional da evolução dos povoamentos ao longo do tempo, utilizando modelos 3D de árvores, bem como a apresentação analítica dos principais indicadores de gestão florestal. Com a abordagem efetuada, pretende-se melhorar a usabilidade e acessibilidade das ferramentas de apoio à gestão florestal, enriquecendo o processo de análise das opções de gestão. Espera-se, assim, contribuir para uma melhor compreensão dos dados de simulação e, consequentemente, para uma tomada de decisão mais informada e eficaz no domínio da gestão florestal.
- A Flexible Tool for Monitoring System CallsPublication . Fidalgo, Tiago Jardim; Resende, João; Martins, InêsIn recent years, cybersecurity threats have been evolving, making real-time threat detection and response critical for securing modern computing environments. Host-based security monitoring is vital in capturing detailed telemetry data, such as system calls and process activities, providing insight into potential security incidents. However, existing telemetry collection solutions often lack the depth and integration required for valuable real-world security analysis. This research proposes a comprehensive security telemetry collection framework that integrates host-based data collection tools with a monitoring system designed to track system calls and processes as key structural and functional units of a system. Additionally, our approach incorporates a security environmental test, simulating real-world threats and vulnerabilities to generate a complete, practical, and real-time host- based security dataset. To enhance the usability of collected data, we also integrate a data visualization dashboard, facilitating more effective analysis and interpre- tation of security metrics. To evaluate the effectiveness of our proposed solution, we compare our security telemetry collection approach, using Miningbeat, with Metricbeat. This comparison is conducted through simulated attacks based on the Red Team Atomic Testing framework, replicating various adversarial tactics. By an- alyzing the distinct sets of metrics collected by each tool and their behavior during different attack scenarios, we assess the capabilities of our approach in providing enhanced security telemetry for host-based threat detection. This thesis aims to contribute to security telemetry collection by demonstrating the advantages of a more integrated approach. Our findings offer insights into how host-based mon- itoring can be optimized for real-time threat detection and response, ultimately improving the security posture of modern computing systems.
- Dynamic Substitutional Reality In Adaptive Virtual ExperiencesPublication . Figueira, Guilherme Ribeiro; Nóbrega, RuiThis dissertation aims to enable the creation of Substitutional Reality (SR) experiences that dynamically adapt to unknown spaces. A key limitation of Substitutional Reality (SR) today is its static deployment: most experiences are designed for a specific physical environment, preventing portability and restricting broader adoption. Researching the practical applications of SR and its advantages in certain fields is also important to encourage further research. To address these challenges, this work explores techniques for developing interactive SR experiences that adapt to arbitrary rooms while supporting the integration of narrative and user goals. A dynamic SR system was developed, which extends existing scene-understanding capabilities of modern Head-Mounted Displays (HMDs) with a runtime substitution pipeline. This pipeline maps physical objects to predefined virtual counterparts, allowing experiences to run across heterogeneous environments without manual calibration. Ad- ditionally, the system introduces versatile screen-space portals that serve two purposes: enabling seamless transitions between multiple substitutional environments and inte- grating video see-through Augmented Reality (AR) for a stronger connection to the real world. Designers can therefore construct SR experiences that combine portability, tangible interaction, and hybrid AR/SR configurations. To evaluate the system, a conceptual museum experience was implemented as a case study. This application demonstrated how Dynamic SR can support cultural heritage use cases, combining interpretive content and interactive artifacts with adaptability to different physical spaces. Evaluations included user studies on how substitution and passthrough integration affect locomotion and engagement, an investigation of portal interactions within the museum scenario, and adaptability tests assessing performance across diverse environments. The results validate the system’s effectiveness in creating adaptive SR experiences and demonstrate its applicability to cultural heritage contexts.
- Causal Inference Explanations for Graph Neural NetworksPublication . Kumar, Sahil Satish; Soares, Cláudia; Caldas, FranciscoThe increasing popularity of Artificial Intelligence, notably Deep Learning, has prompted its application across diverse real-world scenarios. Graph Neural Networks, a type of Deep Learning model developed for graph-structured data, have gained significant popu- larity owing to their versatility and efficacy. Despite their high accuracy and performance, contemporary Deep Learning models often function as black box systems, lacking trans- parency and trustworthiness crucial for deployment in critical decision-making domains. In response, Explainable Artificial Intelligence has emerged, aiming to enhance the trustworthiness of black box models by devising explanation methods that clarify their inner workings. However, prevalent explanation techniques predominantly leverage correlation and association rather than employing causality, a significant aspect of human comprehension. In this thesis, we propose a novel explanation method using causal inference developed specifically for Graph Neural Networks. Our approach seeks to illuminate the decision- making process of Graph Neural Networks, thereby augmenting their transparency and trustworthiness. To show the generality of our approach we do a benchmark evaluation of our method against state-of-the-art explanation methods on a variety of datasets, tasks and Graph Neural Networks architectures. Furthermore, we apply and evaluate our explanation method in two real-world use cases: medical referral recommendation and space object conjunction prediction. In both cases, we demonstrate the applicability of our method and discuss its limitations.
- Large Language Model for Querying Databases in PortuguesePublication . Figueiredo, Lourenço Maria Morais da Silva Pinto de; Marques, Nuno; Cavique, LuísThis dissertation presents a system designed to generate SQL queries from the Portuguese natural language, using Cedis’ sports facility management application ESPORT.IA as a case study. The system’s performance was validated with real-world client queries, focusing on its ability to interpret synonyms and handle complex query requirements. Experiments highlighted challenges such as model hallucination and time interval pro- cessing, emphasizing the need for precise context in improving query accuracy. A custom benchmark, ESPORTNL2SQL, based on the Spider SQL Hardness Criteria, was developed to evaluate the system across varying query difficulties. For the benchmark’s first version, the GPT-4 Turbo and GPT-4o scored 88.79% and 80.17%. For the second version, they scored 75.00% and 80.17% respectively. Results gave evidence that adding contextual information improves accuracy. While the system shows potential in simplifying database interactions through natural language, limitations exist, particularly in managing larger contexts without degrading performance. This research offers insights into the practical application of NL2SQL systems in enterprise environments and their potential for broader deployment.
- Automated Data Extraction from Documents Using AI: Response Automation and Integration in Document Management SystemsPublication . Silva, Afonso Pimenta de Oliveira e; Silva, Joaquim; Campos, Adriano; Roque, MiguelOrganizations are increasingly challenged by the need to process large volumes of un- structured documents such as invoices, receipts, and contracts. Traditional rule-based approaches often fail to capture the variability and complexity of such data, leading to inefficiencies and high error rates. This thesis investigates two AI-driven strategies for automated information extraction from unstructured documents. The first combines Optical Character Recognition (OCR) through Tesseract with a Large Language Model (LLaMA) to identify and extract key fields from recognized text. The second employs Donut, a transformer-based vision-language model designed for end-to-end document understanding, which directly maps document images to structured outputs without an intermediate OCR step. A comparative evaluation is conducted across multiple mea- sures like character and word error rates, extraction accuracy, latency, and scalability to assess the trade-offs between modular and end-to-end pipelines. The results highlight the strengths and limitations of each method, providing insights into the practical deployment of intelligent document processing systems. This work contributes to advancing the field by offering a systematic comparison of OCR-LLM pipelines versus transformer-based end-to-end models for real-world unstructured document processing.
- Ferramenta para conversão de linguagem natural em açõesPublication . Lopes, Rita Gomes; Carvalho, Alberto; Medeiros, PedroA rápida digitalização de serviços aumentou a necessidade de ferramentas capazes de compreender instruções em linguagem natural e de realizar ações em conformidade. Con- tudo, soluções tradicionais, baseadas em regras rígidas, revelam-se insuficientes perante interações complexas, comprometendo a eficácia do apoio ao cliente e a experiência do utilizador. Neste contexto, esta dissertação apresenta um sistema conversacional avançado, focado na gestão de marcações em serviços consulares digitais e integrável em aplicações web, nomeadamente nas soluções da Opensoft. A solução é composta por três componentes principais: um agente conversacional, um servidor e uma interface de utilizador. O agente foi desenvolvido com recurso a um Large Language Model (LLM) e à framework LangChain. O servidor, utilizando FastAPI, assegura a comunicação com a interface de utilizador, sendo também responsável pela invocação do LLM. A interface, desenvolvida em React, disponibiliza funcionalidades como internacionalização, gestão dinâmica de tema e mecanismos de feedback, proporcionando uma experiência de utilização moderna. A solução foi ainda alvo de avaliações da sua robustez e fiabilidade, através de testes manuais, unitários, de integração e end-to-end, realizados com recurso a ferramentas como Pytest e RAGAs. Estes testes asseguraram a utilização correta das ferramentas pelo agente, a consistência das respostas produzidas e a conformidade com os requisitos de segurança definidos. Os resultados demonstram que a integração de LLMs em agentes conversacionais com acesso a ferramentas ultrapassa as limitações de chatbots tradicionais, permitindo melhores interpretações de linguagem natural, execuções de ações fiáveis e respostas ajustadas ao contexto. Além disso, a arquitetura modular e escalável desenvolvida constitui uma base sólida para adaptação a diferentes domínios, contribuindo para modernizar o atendimento digital e melhorar a experiência do utilizador.