Doctoral School

Reading date: 02/04/2025

• PÉREZ I GONZALO, RAÜL: End-to-end learning for wind turbine blades: from imagery data to defect repair recommendations
  
  Author: PÉREZ I GONZALO, RAÜL
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN AUTOMATIC CONTROL, ROBOTICS AND VISION
  Department: Institute of Robotics and Industrial Informatics (IRI)
  Mode: Normal
  Deposit date: 26/02/2025
  Reading date: 02/04/2025
  Reading time: 11:00
  Reading place: Sala d'Actes de la Facultat de Matemàtiques i Estadística (FME) de la Universitat Politècnica de Catalunya, C/de Pau Gargallo, 14, 08028 Barcelona
  Thesis director: AGUDO MARTÍNEZ, ANTONIO
  Thesis abstract: The European Union's (EU) reliance on external energy sources underscores the urgent need for energy security and affordability, driving the transition to renewable energy with wind power as a key renewable solution. However, wind turbine operation and maintenance still account for 30% of energy production costs, due to their prolonged exposure to harsh environmental conditions. Timely defect detection and repair are critical, as turbines must often be halted during visual inspections and repairs. Streamlining the process from inspection to decision-making is essential to reduce downtime and operational costs.This thesis presents a comprehensive end-to-end blade assessment system designed to determine defect severity, quantify their impact on energy production, and deliver actionable repair recommendations. By enabling wind turbine owners to act proactively, this system helps minimize operational costs. The framework emphasizes efficient image transmission that preserves quality, followed by the generation of detailed blade assessments to establish a consistent and effective repair strategy.To this end, this project proposes first segmenting images to isolate blade regions, simplifying subsequent tasks through algorithms tailored for imagery acquired under diverse conditions. These include a Blade U-Net model, which introduces dense conditional-random-field regularization to enhance segmentation accuracy, and advanced post-processing involving iterative refinement through hole-filling and noise reduction via an unsupervised random forest. Two deep discriminant analysis frameworks integrate class separability and probabilistic modeling into robust non-linear architectures to derive precise defect boundaries, handle complex textures, and improve generalization across varied inspection data. Additional contributions include a modular region-growing classifier for efficient segmentation in data-scarce conditions and diffusion-based models with dual-space augmentation to improve generalization and robustness, leading to substantial superior performance than competing techniques. Together, these segmentation methods form the foundation for automated defect detection and diagnostics.In the second part, to address the challenge of handling large volumes of high-resolution inspection data, this work also presents a novel region-of-interest (ROI) image compression framework. Traditional methods often compromise critical defect information. The proposed framework leverages segmentation outputs to ensure high-fidelity compression in blade regions, employing lossless or high-quality lossy techniques while aggressively compressing non-relevant areas. Key innovations include multi-layer nested latent variable models for lossy coding and parallelized bits-back coding optimized for industrial-scale applications. These advancements achieve state-of-the-art performance while significantly reducing computational costs. By coupling compression with our proposed multi-task defect detection model, this approach supports timely and accurate diagnostics, ensuring minimal disruption to turbine operations.In summary, this thesis contributes a hierarchy of low-level to high-level algorithms designed to streamline wind turbine maintenance processes. The combination of advanced segmentation and compression enables a fully automated pipeline for blade defect assessment, encompassing defect localization, classification, and repair prioritization, directly improving energy efficiency by reducing downtime, optimizing maintenance schedules, and minimizing repair costs.
  

Reading date: 04/04/2025

• UGRINOVIC KEHDY, NICOLAS: Modeling and Reconstruction of 3D Humans under Context
  
  Author: UGRINOVIC KEHDY, NICOLAS
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN ARTIFICIAL INTELLIGENCE
  Department: Department of Computer Science (CS)
  Mode: Normal
  Deposit date: 05/02/2025
  Reading date: 04/04/2025
  Reading time: 15:00
  Reading place: Sala de Juntes de la FME
  Thesis director: SANFELIU CORTES, ALBERTO | MORENO NOGUER, FRANCESC D'ASSIS
  Thesis abstract: The study of human's and their behavior through the analysis of images and videos has long been a central topic in Computer Vision. The reconstruction and modeling of human behavior have garnered increasing attention, due to their potential applications in virtual environments, including AR/VR, sports, fashion, and the film industry. Despite this growing interest, accurately capturing and generating the 3D pose and motion of humans remains an important challenge, primarily due to the vast diversity of human movements and the inherent complexity of the human body. Furthermore, the ability to capture and replicate subtle human interactions---such as a hug---that are intuitively understood by humans continues to be a significant obstacle for machines. This complexity arises from the need for a deep understanding of the physical world, its constraints, and the nuanced ways in which humans interact with it.This thesis presents the development of several methodologies for reconstructing and modeling various aspects of humans in 3D, including detailed shape, pose, and motion, mainly from RGB images. A key emphasis is placed on capturing or incorporating contextual information as additional information. First, we introduce a method for modeling the detailed body shape of individuals, which includes elements such as clothing across a wide range of poses. Subsequently, the focus shifts to the simultaneous pose estimation of multiple individuals, wherein scene constraints are employed to enhance the accuracy of these estimations. This approach addresses the fundamental challenges of depth and scale ambiguity inherent in 3D reconstruction. The work is then extended into the temporal domain, to reconstruct interacting individuals, particularly in scenarios involving close interactions. A significant challenge under such situations is the lack of compliance with physical laws, such as body collisions. To address this, we integrate a fully-featured physics simulator within a motion estimation pipeline to account for these physical inconsistencies. Lastly, we propose a method capable of generating human motion that interacts with a virtual environment. All proposed methods have undergone extensive evaluation.In summary, this thesis introduces a suite of tools for the modeling and reconstruction of 3D humans, advancing the field towards more accurate capture and recreation of realistic behavior for virtual humans, with a particular emphasis on their interactions with its surrounding environment.
  

Reading date: 11/04/2025

• LIZÁRRAGA SÁNCHEZ, SALVADOR: Bacardí Tultitlán, México. Mies van der Rohe
  
  Author: LIZÁRRAGA SÁNCHEZ, SALVADOR
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN THEORY AND HISTORY OF ARCHITECTURE
  Department: Department of History and Theory of Architecture and Communication Techniques (THATC)
  Mode: Normal
  Deposit date: 03/03/2025
  Reading date: 11/04/2025
  Reading time: 11:00
  Reading place: ETSAB (Escola Tècnica Superior d'Arquitectura de Barcelona) - Planta Baixa - Sala de GrausAv. Diagonal, 649-651 - 08028 - Barcelona
  Thesis director: GARCIA ESTEVEZ, CAROLINA BEATRIZ | ROVIRA GIMENO, JOSE MARIA | ROVIRA GIMENO, JOSE MARIA
  Thesis abstract: This thesis focuses on the office building for Bacardí y Cía S. A. in Tultitlán, Mexico, which Mies van der Rohe and his team designed and built from 1958 to 1961. Several Mexican companies – Knoll Internacional de México S. A., Constructora Maya, Campos hermanos and SACMAG de México– were involved in the process. For its construction, Mies’ architects –Gene Summers, Jan Lippert and Friedrich Wagner– made dozens of trips from Chicago to Tultitlán, while Mies visited Mexico only once. The thesis has two main objects of study. The first is the archive of the building, which contains about a thousand documents related to the Mexican building stored in the Mies van der Rohe archive at MoMA –hundreds of letters, telegrams, photographs, sketches and plans. The second is the architecture itself, whose peculiar materiality is contrasted with the information in the archive.The Tultitlán building is placed on the margins of the history of Mexican architecture, of Mies' history and, therefore, of Western architectural history. However, by extracting the object from that marginal position and forcing it to take a central position, it drags with it an entire architectural culture and forces the hegemonic discourses of those histories to reconstruct themselves, or at least to be questioned. The unprejudiced dissection of the archive and its building puts to the test historian Manfredo Tafuri's dictum that positioning oneself at “a particular angle of observation allows facts mute in themselves to be forced to become eloquent.” Among others, the archive forces us to place ourselves in the particular angle of vision of its secondary characters in order to understand them as principal and eloquent; from the foreshortening of a marginal city for the history of Western architecture that shows us that it became actually an international center; in the standpoint of a technological and constructive reality that allowed the materialization of a Mies building, but with methods different from those of a rich country; among many others. The research does not hide an inevitable conflict between the “historical word” of our present and that of the documents of another era -because the letters, plans, publications and films used in this research were created in a reality that no longer exists-. In other words, on the one hand, the documents were forced to speak in a language unknown to them –ours– and, at the same time, they were allowed to speak freely, without trying to hide their contradictions for the sake of a supposed historical or scientific congruence acceptable for the present. The collision of times forced to seek support in other languages, disciplines and characters –from Florence Schust Knoll and Lina Bo, to popular office cinema– to make intelligible the transnational context that allowed the existence of the objects of study of this thesis: the archive of the Bacardi offices in Tultitlán and its architecture.
  

Reading date: 09/05/2025

• DAVIS ORTIZ, ALBERTO: Development of a Fuzzy Logic-Based Algorithm for Stroke Detection in Non-Contrast Computed Tomography Images
  
  Author: DAVIS ORTIZ, ALBERTO
  Thesis file: (contact the Doctoral School to confirm you have a valid doctoral degree and to get the link to the thesis)
  Programme: DOCTORAL DEGREE IN BIOMEDICAL ENGINEERING
  Department: Department of Automatic Control (ESAII)
  Mode: Normal
  Deposit date: 18/02/2025
  Reading date: 09/05/2025
  Reading time: 16:30
  Reading place: Aula 28.8, 1a planta, Edifici I, Escola Tècnica Superior d'Enginyeria Industrial de Barcelona, Campus Diagonal Sud, Av. Diagonal, 64708028 Barcelona
  Thesis director: AYMERICH MARTINEZ, FRANCISCO JAVIER | GORDILLO CASTILLO, NELLY
  Thesis abstract: The present work addresses the problem of early stroke detection, not only from the perspective of detection accuracy, but also focusing on computational efficiency, considering the limited availability of cases for training. To this end, several algorithms have been developed to optimize different processes, such as a brain extraction algorithm, an affine transform algorithm, and a specific adaptive filter for noise in computed tomography images. This research has generated valuable resources, such as a brain atlas of healthy Mexican patients and a template of vascular territories. The use of atlases allowed the extraction of features from specific areas. The features used were relatively simple, such as histograms and Haralick textures, which were combined with linear discriminant analysis and an adaptive neuro-fuzzy inference system as a second layer of feature extraction, and finally with a support vector machine as a classifier. Together, these methods achieved a performance of 98.25%. The results show that using the adaptive neuro-fuzzy inference system as a feature extractor improves the performance of other classifiers due to its ability to handle uncertainty and identify nonlinear relationships between variables. This study contributes to the development of low computational cost algorithms and provides new perspectives and tools that could be applied in a real environment in the future