APRÈN. Portal de la producció docent del professorat de la UPC

Aquest projecte explora la implementació i la validació de funcionalitats de l'IEEE 802.11ax (Wi-Fi 6) en un punt d'accés (AP) basat en OpenWrt, utilitzant el GL-MT3000 (Beryl AX) i un client ESP32-C6. A més de la validació de Wi-Fi 6, el projecte també© se centra en el desenvolupament d'aplicacions OpenWrt implementant una aplicació per a la gestió d'estacions. Les funcionalitats estudiades del WifFi 6 inclouen el Target Wake Time (TWT) per millorar l'eficiència energètica i el Basic Service Set (BSS) Coloring per reduir les interferències en canals compartits. S'analitzen el consum energètic i el rendiment de la xarxa mitjançant mesuraments reals i captures de trànsit amb Wireshark. Els resultats mostren les potencials millores d'aquestes funcionalitats en entorns IoT, ajudant a establir configuracions eficients per a aquests dispositius.

This project focuses on the design of an improved analog circuit for quartz crystal microbalance (QCM)-based sensors used to monitor biofilm formation. Limitations of the existing circuit have been addressed, particularly concerning the accuracy of impedance and resonance frequency measurements. A new design based on a sensing resistor is proposed, which has shown moderately improved performance through simulations and experimental tests.

Kubernetes is an open-source platform developed for containerized workload management. This technology is an essential part for modern applications development and deployment, being professionals skilled in its use highly valued in the job market. The deployment of Kubernetes involves setting-up a cluster composed by several machines, a process that may be expensive for those with limited budgets and challenging for non-experienced users. This bachelor's thesis aims to develop a cost-efficient environment, in which Kubernetes beginners can apply their knowledge. The following report provides a detailed explanation on how this objective has been achieved, while also ensuring a comprehensive understanding of the project by offering the necessary context of the used technologies.

This thesis explores the application of AI in odour recognition using the Bosch BME688 sensor. The goal is to create an open, automatic, flexible and cost-effective solution that overcomes the limitations of the Bosch solution in transparency, adaptability and accuracy. The project develops compressed neural networks optimized for the ESP32 microcontroller and transforms the information from the BME688 sensor into matrices, which are visualized as images, for processing with the Edge Impulse. The results improve odour classification compared to the Bosch solution, overcoming challenges with data capture and preprocessing. An open solution, in the future, allows for refining the detection capabilities, expanding the data set and exploring additional applications. The project contributes to advancing AI-driven sensor solutions and applications such as food spoilage detection, environmental monitoring and fire detection highlight the relevance of this work, promoting sustainability

This project focuses on the design and development of an automated solution to provide realtime visibility and efficient management of deviations in security controls within cloud environments, aiming to improve operational efficiency. The activities include analyzing the current situation and identifying limitations, selecting tools, designing integrations and data flows, developing automated processes, and creating interfaces that enable clear visibility of the environment's security posture. The conclusions highlight a significant optimization of time spent on repetitive tasks and greater accuracy in identifying deviations in security controls, effectively meeting the established requirements.

Family networks play a pivotal role in shaping our decisions, influencing our interactions with online services, and impacting the utilization of communication network resources. This research investigates the identification of influential individuals within family networks that maximize the spread of influence. Family networks were generated using an agent-based simulator in OMNeT++. Subsequently, a Graph Neural Network (GNN) model was employed to predict the most influential nodes. Results demonstrate that the GNN effectively approximates the influence spread for the Independent Cascade Model (ICM) achieving 94% maximum influence in significantly less time than a traditional algorithm. However, its performance decreased for the more complex Linear Threshold Model (LTM). This suggests that more complex GNN architectures, incorporating additional layers and neurons, may be required to accurately predict influence spread in scenarios with intricate interdependencies, as exemplified by the LTM model.

Image sensor design is a sector whose products are gaining more and more uses. Selfdriving vehicles, medical procedures, factory processes monitoring, and many others are examples of applications where image sensors can revolutionize the industry. For this growth in applications, new types of image sensors are being designed, including new photosensive elements. With this importance in the industry, parties signing a product contract define a set of specifications and requirements. In addition, to ascertain compliance with the specifications, the image sensing devices have to be tested and characterized. In this testing stage, it is essential that the test system should not pollute the test results. This is why a test system should also have its own test procedures, which validate it for testing the targeted device under test (DUT). This thesis covers the design of a test system whose final purpose is to test an image sensor, but putting emphasis on including tools that allow the validation of the test system without dependence on this sensor. From printed circuit board (PCB) design to user application software code, going through embedded control software and programmable logic data processing. This project was carried out at IMASENIC Advanced Imaging [1], a company that is a reference in the imaging sector worldwide.

This thesis presents the development carried out to improve the Machine Protection System (MPS) of a particle accelerator. The document describes the general proposal for the improvement of the system and focuses on the design and validation of the RTL corresponding to two of the main components of the MPS. The RTL design has been programmed using VHDL as hardware description language. The designed system receives and analyses data from the Local Control Systems (LCS), acting accordingly by generating alarm signals and providing information needed to monitor the system status in real time. As this is a critical system, the development process included extensive validations through simulations, hardware testing and a final demonstration at the accelerator facility.

This thesis proposes an end-to-end framework for malware detection that models endpoint memory as a dynamically evolving mathematical object. The system acquires memory snapshots via a PowerShell-based VirtualBox control script, compresses them into complex-valued representations capturing magnitude and orientation changes (implemented in Rust), and analyzes these representations as two-dimensional time series in Python. This approach simulates a “MEMORY ECG,” where the delta of page content across consecutive snapshots reveals anomalous behavior without requiring process specific knowledge. At the core of this framework is an autoencoder (AE) that captures the latent structure of normal memory blocks across multiple snapshots. By modeling how blocks evolve over time, the AE identifies deviations from these learned distributions, signaling potential malware. Feature extraction, essential for AE input, is performed using Wavelet Scattering Transform (WST) or alternative techniques such as Fourier Transform, Cepstrum and wavelets analysis. Unlike WST, which generates invariant features directly, these other methods involve an additional layer of preprocessing to extract features from raw memory deltas, enabling the AE to detect anomalies effectively. Developed in Python, the solution offers extensive configurability and automation for both signal processing and machine learning through a Keras-based pipeline. Users can finetune parameters such as block size, transform type, scaling, and PCA, as well as define operational modes (training, incremental training, or testing). By integrating memory acquisition, compression, and feature extraction into a continuous workflow, this scalable framework lays the foundation for real-time behavioral malware detection, enabling live anomaly monitoring through endpoint memory analysis as a time-varying signal.

This thesis explores the implementation of a Self-Sovereign Identity (SSI) system using Ethereum and Decentralized Identifiers (DIDs). The project focuses on leveraging blockchain technology to create a secure and decentralized framework for digital identity management, incorporating Verifiable Credentials (VCs) and Verifiable Presentations (VPs). Key components include DID document management, secure user authentication, and user-friendly interface. What makes this system different is the integration with existing wallets, privacy and user control through selective disclosure, allowing users to share only necessary information, and key rotation. It also uses EIP-712 signatures for secure and structured data signing, which allows users to clearly see and understand what they are signing while the cryptography is securely handled by the wallet. Future work will focus on adding more wallet support, improving data storage, and enhancing system scalability and security.

This thesis presents the design and implementation of an automotive network system where multiple ECUs transmit information to the TCU using two network protocols: CAN and Automotive Ethernet. Beyond the system's design, a comprehensive evaluation is done, focusing on the reliability of the different network technologies under various test configurations. Particular emphasis is placed on comparing two media access control methods for Automotive Ethernet: CSMA/CD and PLCA. Regarding this comparison, the PLCA method demonstrates a significant improvement over CSMA/CD, achieving 100% reliability due to its deterministic arbitration mechanism. However, further in-depth evaluation under worst-case scenarios and with additional devices is recommended to fully assess the system's robustness.

The present project, “Firmware Design for Centralized and Distributed Battery Management Systems”, aims to design, implement and test various components of the firmware for two types of Battery Management Systems (BMS): centralized and distributed. These BMSs will manage the batteries, balancing their cells and ensuring safe operation. Although this works is focused on the firmware design of each BMS, the document will also explain the hardware to provide the needed context information. Furthermore, it also describes the hardware and functional tests carried out to verify the correct implementation of the system. Finally, several improvements are proposed, which could be implemented in the future to enhance safety or add additional functionalities.

Gastric capsules emerged as an alternative to conventional approaches for diagnosing gas- trointestinal (GI) diseases. Proper localization of these capsules is crucial for identifying GI regions with abnormalities, delivering biomarkers at specific locations, and adjusting video quality based on their position. However, this is not an easy task due to the com- plexity of biological tissue, energy constraints, and the need for a high level of accuracy. Nevertheless, ReMix, a novel wireless system based on backscatter communication, is ca- pable of localizing a deep-tissue device with zero power consumption. The device backscat- ters a radiofrequency (RF) signal, which can be isolated from surface interference and signal deflection due to its non-linear behavior. This signal can then be utilized for lo- calization purposes, achieving an accuracy of 1.4 cm in animal tissue. Considering these advantages, this thesis explores the development and validation of a gastric capsule local- ization system based on the ReMix methodology. The primary focus is on replicating and validating the performance of a backscatter device and its non-linear behavior. The device was constructed by welding a flat patch antenna to a diode, enabling the generation of harmonics required for backscattering the signal and facilitating localization. Hardware components, including Universal Software Radio Peripherals (USRPs), were used to transmit and receive RF signals at approximately 1 GHz. Signal visualization was performed using GNU Radio, while inspectrum software was used to store the received signals. Several modifications to the original ReMix setup were implemented, including the use of two types of antennas, conducting experiments outside shielded chambers, and adjusting parameters to accommodate laboratory constraints. As a result, the findings differ from those of the original ReMix study. Nevertheless, the backscatter device's ability to generate harmonics was demonstrated across various scenarios, including open air, inside a chicken body, and embedded in pork tissue. However, the device's performance was influenced by the arrangement of the antennas, the type of antennas used, and parameters such as the distance and gain settings of the receiving antennas. The harmonics generated during in-animal experiments were used to validate the local- ization algorithm. When the receiving antennas were aligned along the same y-axis, the algorithm failed to optimize the y-axis coordinate, resulting in errors ranging from 15.71 cm to 37.5 cm. However, when three receiving antennas were distributed at different po- sitions within a 3D plane, and certain assumptions were made, the algorithm achieved a localization accuracy of 8.45 cm. This thesis lays the groundwork for further research aimed at enhancing the precision and real-time tracking capabilities of gastric capsule localization systems.

This Master's Thesis addresses the challenge of Catalan–Castilian code-switching in speech recognition. It develops synthetic corpora by combining machine-translated text chunks, large language model prompts, and semi-structured dialogues, then uses text-to-speech to produce audio samples. After fine-tuning ASR models, results indicate better recognition of artificially generated code-switching data but highlight limitations in real-world scenarios, such as emergency calls. The research underscores the need to incorporate more authentic recordings for stronger generalization, offering a practical framework to improve code-switching handling in multilingual ASR systems.

Continuous-time pipeline (CT-Pipeline) ADCs are an emerging analog-to-digital medium-resolution converter architecture. Recently demonstrated examples prove that they are capable of achieving larger bandwidths than those of state-of-the-art continuous-time ΣΔ ADCs, which until now were the preferred wideband mediumresolution ADC choice. In this work, this type of converter is demonstrated through a MATLAB Simulink model. Throughout these pages, a detailed analysis and explanation of this topology is provided, while proposing solutions to some of the recurring design challenges. With regard to the implementation, the use of power-efficient VCO quantizers has been evaluated. A comparison is also conducted between two alternatives for the digital reconstruction filter. In addition, the most common non-idealities and their effects are also introduced and validated by simulation. The developed model of a 4-stage continuous-time pipeline ADC achieves a SNDR of 80 dB over a 400-MHz bandwidth. It is intended to serve as an initial step towards the implementation of a CT-Pipeline that achieves an SNDR of 70 dB.

This thesis explores the optimization of a semi-supervised model for object detection combining the Soft Teacher framework with the DINO object detector. A student model is trained with a combination of labeled and unlabeled images, while a teacher model generates pseudo-labels for the unlabeled data. To enhance the quality of the generated pseudo-labels, different strategies are explored to filter them and discard the bad ones. This study shows the importance of this crucial step for the final performance of the model.

This thesis explores the question: can vehicle tracking be improved using Automated License Plate Recognition (ALPR)? The project builds upon an existing modular approach to vehicle tracking by adding two new modules: one to correct license plate text predictions and another to use these corrections to improve vehicle tracking. As an alternative to the traditional approach of reading license plate text, which can be time-consuming, a deep learning model has been trained from scratch to determine whether two license plates are the same. While the main question has been answered positively, this project lays the groundwork for further improvements in re-identification. The system's performance, when integrated into the modular framework, leaves room for refinement in future studies.

Developing effective deep learning methods for digital pathology image analysis remains a challenging task due to the scarcity of large annotated datasets in the medical domain. This is particularly notable for cell classification tasks, which require very specific labels. This thesis explores a potential solution to this problem by using a vision-language model specialized in histopathology to generate cell classifications in a zero-shot setting, only with textual descriptions of the cell classes. We evaluate the model on this task and develop a finetuning strategy using a dataset of image-caption pairs to improve its cell specific understanding. The results demonstrate the potential of the model for zero-shot cell classification, particularly after the fine-tuning process. However, its ability to generalize to more granular tasks remains limited. These findings highlight the potential of histopathology based vision-language models for providing specific cell type classifications, without requiring specific training, only from natural language descriptions.

This thesis investigates the relationship between ionospheric ionization levels and GNSS positioning errors, focusing on data from the ALEX2 station between 2013 and 2019. Us ing single- and dual-frequency Precise Point Positioning (PPP), ENU positioning errors were analyzed in correlation with ionospheric Total Electron Content (VTEC) derived from Global Ionospheric Maps (GIMs) and geomagnetic indices (Kp, ap). The analysis revealed that dual-frequency PPP mitigates ionospheric delays more effectively, resulting in improved accuracy compared to single-frequency PPP. Statistical correlations between positioning errors and ionospheric/geomagnetic indices highlighted significant temporal variations across daily, weekly, and monthly resolutions. Custom Python workflows, ad vanced GNSS processing with RTKLib, and space weather integration provided robust insights into ionospheric impacts on GNSS accuracy. The study aimed to develop a posi tioning error index that could be utilized in space weather applications, bridging the gap between GNSS error characterization and ionospheric monitoring.

3D reconstruction using multiple cameras is a growing field with applications in augmented reality, virtual reality, and telepresence. Accurate calibration of cameras and color consistency in point clouds are critical to achieving high-quality reconstructions. This project aims to improve the calibration process by automating the computation of extrinsic parameters and ensuring consistent color correction across multiple cameras. By leveraging tools such as OpenCV and Open3D, we propose methodologies to reduce human intervention, enhance efficiency, and improve the overall quality of 3D reconstruction. The project employs a setup with several Azure Kinect cameras, focused on a central point. For extrinsic calibration, we utilize ArUco markers and iterative registration techniques to compute precise camera alignments. For color correction, overlapping camera views are analyzed to establish a consistent reference color space using lookup tables. These advancements streamline the calibration process and harmonize color across point clouds, enabling robust and scalable volumetric capture systems suitable for real-time applications.