ESTG - DM - Engenharia Informática
Permanent URI for this collection
Browse
Recent Submissions
- Real-Time Data Analysis Tool for Decision Support in Streaming EnvironmentsPublication . Torres, Diogo Pinto; Carneiro, Davide Rua
- Web3 Ecosystem for Device as a ServicePublication . Tavares, Maria Da Conceição Pereira; Pinto, António Alberto dos SantosThe management of distributed physical devices, especially in large-scale scenarios such as governmental digital inclusion initiatives, poses challenges related to security, control, and transparency. The management of the device lifecycle in unsupervised environments is exacerbated by concerns such as theft, loss, mistreatment, and lack of trustworthy processes for transparency and remote monitoring. This dissertation presents a novel blockchain-based Device as a Service (DaaS) system that combines smart contracts, a web portal for legitimate users, and an embedded verification module at the United Extensible Firmware Interface (UEFI). The architecture ensures decentralised, transparent, and tamper-resistant device management, allowing remote locking and unlocking, real-time state verification, and immutable activity recording without the need for intermediaries. Implemented on the Ethereum blockchain (Sepolia testnet), the system uses smart contracts written in Solidity for device registration and state management, Keyed-Hash Message Authentication Code (HMAC)-based cryptographic tokens to ensure data integrity, and Role Based Access Control (RBAC) to restrict actions to legitimate entities. A hardware-level security layer is supplied by a customised UEFI module, that validates device status updates before the Operating System (OS) boots. The functionality, security, and scalability of the system are demonstrated by validation in realistic scenarios, such as device registration, access control processes, and audit logs. This evaluation proves the solution’s ability to securely automate remote device management while providing transparency and resistance against attacks such as replay attacks. The present study addresses the gaps presented in current remote management strategies by offering a comprehensive and novel system for decentralised device management that can be applied in business, government, and educational environments. The work received recognition with a Best Paper Award, highlighting its scientific value and practical application.
- Análise e Monitorização da Qualidade de Dados no contexto da Indústria 4.0Publication . Peixoto, Teresa Maria Oliveira; Oliveira, Bruno Moisés Teixeira de; Oliveira, Óscar António Maia deA industria 4.0 representa uma mudança de paradigma nos sistemas de produção, caracterizada pela integração de tecnologias digitais avançadas, como a Internet das Coisas, os Sistemas Ciberfísicos e a Inteligência Artificial. Estas tecnologias possibilitam a recolha massiva de dados em tempo real, permitindo monitorizar, automatizar e otimizar processos industriais com um nível de precisão e eficiência sem precedentes. No entanto, o valor dos dados gerados depende intrinsecamente da sua qualidade. Dados incompletos, imprecisos, inconsistentes ou desatualizados podem comprometer não apenas a fiabilidade das análises, mas também a segurança e eficácia das decisões tomadas com base nesses dados.
- Desenvolvimento de um Sistema para o Moodle com Integração de Modelos de Linguagem de Grande Escala para Apoio à Atividade LetivaPublication . Silva, Ricardo Rodrigues da; Oliveira, Bruno Moisés Teixeira de; Oliveira, Óscar António Maia deEsta dissertação apresenta o desenvolvimento de um sistema para o Moodle com integração de Modelos de Linguagem de Grande Escala (LLMs) destinado a apoiar a atividade letiva através de automação inteligente de tarefas pedagógicas. A solução proposta combina o protocolo Model Context Protocol (MCP) para orquestração cliente-servidor, um pipeline RAG (Retrieval-Augmented Generation) para recuperação semântica de conteúdos, a base de dados vetorial Qdrant para indexação eficiente, e modelos Gemini (comutáveis dinamicamente consoante custo e desempenho) para geração de linguagem natural. O sistema integra-se com o Moodle via web services para descoberta de cursos e obtenção automática de PDFs, processa documentos com extração de texto e segmentação, cria embeddings com sentence-transformers e disponibiliza funcionalidades de alto impacto educativo: geração de resumos contextuais, criação de questionários com validação de dificuldade, flashcards, recomendações de leitura e, opcionalmente, geração de vídeos com IA (Gemini Veo). A interface web, desenvolvida em Flask, oferece autenticação, gestão de conversas, seleção de modelos e invocação de ferramentas MCP em tempo real. Metodologicamente, adotou-se uma abordagem iterativa com validação funcional contínua e logging estruturado. A avaliação incide em três dimensões: (i) desempenho do RAG (tempo de resposta e relevância de recuperação), (ii) qualidade pedagógica do conteúdo gerado (clareza, alinhamento com fontes, adequação de dificuldade), e (iii) robustez da integração MCP–Moodle (confiabilidade de chamadas e sincronização). Os resultados demonstram ganhos de eficiência na preparação de materiais, apoio à diferenciação pedagógica e redução de esforço docente em tarefas repetitivas. As principais contribuições incluem: uma arquitetura modular replicável para LMS, um conjunto de ferramentas MCP focadas em educação e um fluxo RAG otimizado para PDFs académicos. São discutidas limitações (dependência de qualidade dos PDFs, custos de inferência, privacidade) e perspetivas futuras, nomeadamente exportação direta para formatos Moodle, algoritmos de repetição espaçada para flashcards e estudos de eficácia com turmas reais. Este trabalho evidencia a viabilidade de integrar LLMs no ecossistema Moodle para potenciar práticas pedagógicas baseadas em conteúdo e dados.
- Authentication API - A SSO Authentication and Authorisation Infrastructure for WebPublication . Fernandes, José Pedro; Silva, Fábio André Souto daModern web applications leverage various login techniques, such as Single Sign-On (SSO), passkeys, and password-less authentication, to enhance user experience. Many SSO solutions exist, that enable users to log in once and be authenticated across multiple applications. In this project a custom web authentication system, tailored to the specific needs of a corporate team, was developed. In this team, the lack of web-based authentication infrastructure inhibited the transition from desktop to web applications. The primary objective was to develop a SSO authentication system that not only supports human users but also provides authentication for processes running without a browser, such as automated scripts which will not use SSO but Windows authentication instead. By utilising JSON Web Tokens (JWTs) and refresh tokens, the solution ensures authentication and fast re-authentication, while a distributed cache enables scalability allowing multiple instances to run concurrently. As a result, an Application Programming Interface (API) called AuthenticationApi was developed alongside three internal connection libraries to simplify integration for both web applications and services. A management console was also created to manage the whitelisting of clients, being them web applications or technical processes. The API was rigorously tested, achieving 96.1% code coverage through unit and integration tests, and successfully deployed in two geographical locations, New York and Paris. Structured logs were implemented, offering insights into API performance and usage patterns. Currently, the API is being used in production and serves as a key infrastructure component for the team.
- Microservices Orchestration vs.Choreography: A Comparison and AnalysisPublication . Marques, David Miguel Sousa; Santos, Ricardo Jorge da SilvaThis dissertation explores a critical evaluation of orchestration and choreography in microservices architecture, with particular attention to how these elements affect implementation complexity, latency, and resilience. Given the growing importance of microservices in modern software development, it is critical for developers and architects to comprehend these architectural principles. The study uses a mixed-methods approach to collect data on the efficacy of each approach in practical applications, including qualitative interviews with industry practitioners and the implementation of a solution based on a real-world scenario. The results indicate that while orchestration enables more control over error management and process integrity, choreography provides improved scalability and service independence. The centralisation of orchestration can lead to weaknesses such the possibility of a single point of failure and, in some cases, a rise in latency. This present paper highlights how crucial it is to align architectural decisions with system specifications and provides an overview for visual decision-making that shows the considerations associated when deciding between orchestration and choreography.
- Aplicação de Modelos de Machine Learning para Previsão de Eventos de Stress FinanceiroPublication . Fernandes, Ana Beatriz Esteves; Carvalho, Mariana Valério; Borges, Ana Isabel CoelhoO stress financeiro nas organizações pode manifestar-se através de eventos críticos, como falência, e a capacidade de prever esses eventos é crucial para a gestão de riscos e a tomada de decisões estratégicas. O presente estudo envolveu a aplicação e comparação de cinco modelos distintos de sobrevivência para prever eventos de stress financeiro: Regressão de Cox, Random Survival Forest (RSF), Kernel SVM, Multi-Task Logistic Regression (MTLR) e DeepSurv. Cada modelo foi selecionado com base nas suas características específicas e o seu potencial para lidar com dados de sobrevivência, oferecendo uma abordagem abrangente para a análise preditiva. Este trabalho detalha também o processo de seleção e preparação dos dados, abordando todo o processo seguido desde a recolha dos dados até à análise de correlações entre variáveis. A identificação e remoção de variáveis altamente correlacionadas ajudaram a otimizar o desempenho dos modelos e a simplificar a interpretação dos resultados. Os resultados obtidos indicam que todos os modelos aplicados foram eficazes na previsão de eventos de stress financeiro, com o RSF destacando-se pela sua performance superior. O estudo demonstra a aplicabilidade e a eficácia dos modelos de sobrevivência baseados em Machine Learning (ML) na identificação de riscos financeiros, oferecendo informações valiosas para a gestão financeira e a tomada de decisões estratégicas. Em conclusão, este trabalho contribui para a literatura existente ao aplicar e comparar uma vasta gama de técnicas de ML de sobrevivência na previsão de eventos de stress financeiro. As descobertas oferecem uma base sólida para futuras pesquisas e práticas na área, enfatizando a importância da escolha adequada do modelo para a previsão e a gestão eficaz dos riscos financeiros.
- Sistema de Rastreabilidade de Madeira: Uma abordagem baseada em mecanismos semânticos para integração e validação dos dados das atividades de exploração florestalPublication . Silva, Hugo Daniel Martins; Sousa, Cristóvão DinisIn recent years, growing concerns about deforestation have driven the need to monitor the origin and history of the wood arriving at factories. This has led to the adoption of traceability systems in the forestry sector. However, many of these systems are still manual and paper-based, which makes them susceptible to errors and falsification. With the advancement of Industry 4.0 and the digitalisation of forestry operations, there is an opportunity for the digital transformation of traceability. However, the digitalisation process faces challenges at various levels, one of the main issues being that the information sources are dispersed among the various stakeholders in the forest supply chain, resulting in inaccurate and hard-to-access data, which hinders a comprehensive analysis of the wood’s journey. In this context, the present work proposes a traceability system that integrates data from the various stages encompassing forestry exploitation, from the forest to the factory, ensuring a continuous flow of information. The system is based on an ontology that, in addition to formalising the knowledge necessary for traceability, allows for the identification of errors and inconsistencies through reasoning mechanisms, thereby ensuring the transparency and reliability of the collected records. Furthermore, based on the instantiated ontology, Graph Machine Learning techniques are used to train a model capable of predicting missing data and identifying implicit semantic relations. The approach was evaluated in the context of the Floresta 4.0 project and showed promising results in terms of its effectiveness. In addition to addressing the needs of traceability, it detected inconsistencies that had not previously been identified by domain experts.
- Multimedia data extraction and analysis tool: focus on video and image processingPublication . Bragança, João Miguel Teixeira; Silva, Fábio André Souto daIn today’s digital landscape, the rapid growth of multimedia content, particularly from influencers, has created a critical need for advanced monitoring tools. Building on previous research in multimedia data analysis, this dissertation proposes the development of a tool for extracting and analysing multimedia data to detect violations in influencer-produced content. The tool leverages pre-trained models such as Whisper.AI for speech recognition, YOLOv8 for object detection, and EasyOCR for Optical Character Recognition (OCR). Additionally, sentiment analysis models are employed and tested, with YOLOv8 further trained for specific tasks such as logo detection, ensuring adaptability to various use cases. The objective of this dissertation is to design a versatile and customisable tool capable of performing precise content analysis, including object detection, speech transcription, OCR, sentiment analysis, image classification and logo detection. The solu
- LIME: Optimising the creation of explanationsPublication . Pereira, João Tiago Moreira; Carneiro, Davide RuaExplainable Artificial Intelligence (XAI) techniques are increasingly necessary for ensuring trust and acceptance of complex machine learning models across various fields. One widely used XAI method, Local Interpretable Model-agnostic Explanations (LIME), is particularly popular for image-based explanations but faces challenges in terms of speed, accuracy, and applicability in different contexts. An improvement to LIME is proposed to optimize its performance, including faster training times and better prediction accuracy, with a focus on finding an alternative machine learning algorithm that can outperform the current one used by LIME. Additionally, this project defines and explores metrics derived from LIME explanations that can help evaluate the quality of image classification models, even in concept drift scenarios where labeled data may be scarce. These metrics are validated against human feedback, identifying four key metrics that could prove useful for automated systems to assess model outputs. Furthermore, in domains like manufacturing, LIME explanations must be adapted to context-specific challenges. In the case of defect detection in the textile industry, the permutation generation process used by LIME can mislead the underlying model, generating poor explanations. A methodology is proposed to mitigate this issue, supporting more accurate and contextually relevant explanations that can enhance decision-making and human-centric approaches in industrial scenarios.