Doctorado en Ingeniería - Automática

URI permanente para esta colecciónhttps://repositorio.unal.edu.co/handle/unal/81651

Examinar

Mostrando 1 - 20 de 116

Arbitrary signal constellation design for optical coherent transmission systems
(Universidad Nacional de Colombia, 2026) Márquez Viloria, David Andrés; Guerrero Gonzalez, Neil; Marquez-Viloria, David [0001592937]; Guerrero-Gonzalez, Neil [0000300500]; David Andrés Márquez Viloria [DltMmuAAAAAJ&hl]; Neil Guerrero Gonzalez [5IApxMsAAAAJ&hl]; Marquez-Viloria, David [0000000201687276]; Guerrero-Gonzalez, Neil [0000000285043741]; David Andrés Márquez Viloria [David-Marquez-Viloria]; Neil Guerrero Gonzalez [Neil-Guerrero-Gonzalez]; Environmental Energy and Education Policy E3P; Marquez-Viloria, David [57192269433]; Guerrero-Gonzalez, Neil [25929118600]
This PhD thesis proposes an unsupervised density-based approach for real-time arbitrary constellation demodulation in coherent optical communications. Traditional telecommunications receivers assume standardized, symmetrical modulation formats relying on fixed decision boundaries for symbol recovery. This work extends conventional frameworks by detecting unexpected signal variations using unsupervised machine learning. Rather than imposing rigid decision grids, information is recovered through statistical recurrence tracking, replacing static decision thresholds with dynamically updated cluster boundaries in real time. To realize this scheme in high-speed environments, DBDemod, a real-time FPGA-based demodulator for arbitrary constellations, was developed using an adapted DBSTREAM algorithm. This density-based micro-clustering method unifies equalization, synchronization, and detection into a streamlined processing chain. The heterogeneous CPU-FPGA architecture balances efficiency and flexibility: the FPGA handles massive parallelization of recurrent comparisons and statistics mapping in a single clock cycle, even with finite resources, while the CPU manages the adjacency matrix and demapping rules, enabling constellation logic updates via software without hardware redesign. Technical optimizations minimize computational and energy costs by replacing complex exponentials with bit-shift operations (powers of two) in fixed-point format. Implemented on AWS F2 VU47P via High-Level Synthesis, it achieves 210.52 Gbps throughput with <7 % resource utilization (LUT: 70,785; FF: 92,033; BRAM: 120; DSP: 512). Validated under AWGN, nonlinear phase noise, 16/32/64-QAM, and arbitrary constellations, DBDemod enables self-adaptive elastic optical networks, underwater optical channel characterization (ONRG-AFOSR N62909-24-1-2088), and next-generation format-agnostic transceivers
Explainable deep learning framework to support wind speed forecasting for sustainable energy planning
(Universidad Nacional de Colombia, 2025) León Gómez, Eder Arley; Álvarez Meza, Andrés Marino; Castellanos Domínguez, César Germán; León Gómez, Eder Arley [0001635673]; León Gómez, Eder Arley [3f_bcccAAAAJ&hl]; León Gómez, Eder Arley [Eder-Arley-Leon-Gomez]; Grupo de Control y Procesamiento Digital de Señales
Wind energy plays a pivotal role in the transition toward sustainable and decarbonized energy systems. Accurate wind speed forecasting is essential for optimizing the integration of wind power into national grids, ensuring grid stability, and supporting reliable energy dispatch. In regions such as La Guajira, Colombia, the wind energy potential is considerable—exceeding 580,000 GWh/year across more than 100 planned wind farms. However, the effective exploitation of this resource requires forecasting models capable of handling the chaotic, non-stationary, and regionally heterogeneous nature of atmospheric dynamics. Traditional physical and statistical approaches often struggle to generalize across spatial domains and temporal scales due to their rigidity and computational demands. While deep learning methods have emerged as powerful alternatives, they still face critical challenges related to data scarcity, generalization under non-stationarity, and limited interpretability, which hinder their adoption in safety-critical energy systems. This thesis addresses three core challenges in the design of reliable wind speed forecasting systems. First, data heterogeneity across wind farms—arising from geographical variability, sensor configurations, and temporal resolutions—complicates the training of transferable models. Second, non-stationary wind dynamics driven by atmospheric variability degrade the performance of conventional deep learning models, particularly for medium-term forecasting horizons. Third, the lack of interpretable and auditable model behavior limits trust and operational deployment in energy infrastructure contexts. These challenges motivate the methodological contributions presented in Chapters 4, 5, and 6, and their integration within an operational system in Chapter 7. Chapter 4 introduces a data augmentation strategy termed Neighborhood Preserving Cross-Dataset Data Augmentation using UMAP-based Neighborhood Preserving Cross-Dataset Data Augmentation (UMAP-CDDA). The proposed method constructs a low-dimensional latent space that preserves local topological structures across geographically distinct datasets, enabling the generation of synthetic time series via neighborhood sampling in the embedded space. Experimental results using SRNN, LSTM, and GRU models on datasets from Argonne, Beijing, and Chengdu demonstrate significant improvements in forecasting accuracy and cross-dataset generalization under limited data conditions. Chapter 5 proposes a kernel-enhanced deep learning architecture that integrates multi-scale Random Fourier Features with recurrent neural networks to explicitly model non-linear and non-stationary temporal dynamics. By projecting input sequences into structured spectral representations, the proposed framework leverages kernel theory without incurring the computational cost of explicit kernel matrices. A nested optimization strategy combining Bayesian optimization and gradient-based learning is employed to tune spectral and recurrent hyperparameters. Comparative evaluations against conventional RNNs and Transformer-based baselines show that the proposed RFF-RNN models achieve superior robustness and accuracy, particularly for medium-horizon forecasting under varying climatic regimes. Chapter 6 presents a kernel-scale interpretability framework for deep multi-horizon forecasting models. By combining temporal attributions with scale--time saliency analysis in the spectral feature space induced by Random Fourier Features, the proposed approach provides faithful and model-consistent explanations of forecasting behavior across horizons. Complementary dimensionality reduction techniques, including Kernel-PCA and UMAP, are employed to analyze learned representations and assess generalization. Applications to real-world wind datasets confirm that kernel-scale explanations reveal meaningful spectral–temporal patterns and support transparent model validation. Finally, Chapter 7 introduces an integrated conversational AI system that operationalizes forecasting and interpretability within an end-to-end climate intelligence framework. By combining data ingestion, forecasting models, explanation outputs, and a retrieval-augmented conversational interface, the system enables interactive and traceable access to wind forecasts across regions and horizons. Overall, this thesis contributes novel methodologies for data-efficient, interpretable, and generalizable wind speed forecasting. The proposed approaches support reliable wind energy planning and provide a scalable foundation for deploying explainable deep learning systems in renewable energy and broader environmental forecasting applications (Texto tomado de la fuente).
Planning methodologies for the operation and expansion of distribution systems considering the interactions with the transmission system
(Universidad Nacional de Colombia, 2026) González Vanegas, Wilson; Murillo Sánchez, Carlos Edmundo; González Vanegas, Wilson [0000041735]; González Vanegas, Wilson [xMG3_f4AAAAJ&hl]; González Vanegas, Wilson [0000000309699287]; González Vanegas, Wilson [Wilson-Gonzalez-Vanegas]; Gipem Grupo de Investigación en Potencia, Energía y Mercados; González Vanegas, Wilson [57193393258]
This dissertation explores the design of mathematical optimization models for planning the operation and expansion of distribution systems, paying special attention to the interactions with the transmission system. To model the three main components that form an integrated distribution-transmission system---namely, the unbalanced distribution network, the balanced transmission system, and the distribution-transmission interface---a novel low-error quadratic approximation of the AC power flow equations, termed QTPC, is devised utilizing a consistent mathematical presentation that progresses from the level of power system components to the entire network, thereby providing a comprehensive modeling approach in matrix form. This QTPC formulation is based on the concept of transformed polar coordinates, which allows deriving a quadratic model using a polar-based voltage setting that does not require additional equations to include set-points, unlike Cartesian-based models. Attained results of power flow calculations over many transmission, distribution, and joint distribution-transmission test cases demonstrate the QTPC approximation is a low-error approach. With the QTPC formulation at hand, two models for modeling the distribution-transmission interface are analyzed: a complete three-phase description of the system and a hybrid single-phase/three-phase representation, concluding that the latter is a low-error approach that requires fewer decision variables. Furthermore, the QTPC formulation serves as the central element for formulating an operations planning model based on optimal reactive power dispatch (ORPD) for transmission-informed distribution systems, which falls within the scope of quadratically-constrained quadratic programming (QCQP) problems. With respect to the expansion of distribution systems, considering the interactions with the transmission system, the distributed generation expansion planning (DGEP) problem is formulated as a mixed-integer quadratically-constrained quadratic programming (MIQCQP) problem. Both optimization-based methodologies incorporate several constraints for modeling technical conditions at the distribution-transmission interface, such as the satisfaction of a minimum exchange power factor, usually included in the power systems regulations of the different countries. A significant contribution of this dissertation is devoted to software implementation. More than a hundred classes and functions were incorporated into the widely used tool for power systems research and education, MATPOWER. New top level functions called run_orpd() and run_dgep() were designed upon a robust implementation of the QTPC formulation for solving ORPD and DGEP problems for integrated distribution-transmission systems. These classes and functions are available for academics and researcher for reproducing the results presented in this dissertation and using them for their own needs (Texto tomado de la fuente).
A deep learning approach with preserved interpretability to support multimodal remote sensing in precision agriculture with applications to the Tolima region
(Universidad Nacional de Colombia, 2025) Triana Martinez, Jenniffer Carolina; Álvarez Meza, Andrés Marino; Fernández Gallego, Jose Armando; Triana Martinez, Jenniffer Carolina [QK7GBBcAAAAJ&hl]; Triana Martinez, Jenniffer Carolina [0000000297063635]; Grupo de Control y Procesamiento Digital de Señales
Precision agriculture (PA) is an essential strategy to address rising global food demands under the constraints of climate change and resource scarcity. By integrating remote sensing (RS) technologies with multimodal data analysis, PA enables site-specific management to improve crop yields and sustainability. The analysis of heterogeneous data from sources like Unmanned Aerial Vehicles (UAVs) and in-field sensors is often performed using deep learning (DL) models. However, the practical application of these models in real-world agricultural settings is hindered by three significant challenges: the prevalence of noise and conflicts within multimodal data, the nonstationary behaviour of agroecosystems due to phenological and environmental variability, and the inherent lack of interpretability in complex "black-box" models, which impedes trust and adoption by stakeholders. This thesis delivers a methodological framework to overcome these challenges by developing a suite of deep learning techniques with preserved interpretability. First, to mitigate the effects of noisy and unreliable labels in the context of multiple annotators, the Generalized Cross-Entropy-based Chained Deep Learning (GCECDL) framework is introduced. This approach jointly models the reliability of each annotator as a nonstationary function of the input features and employs a noise-robust loss function to ensure stable convergence. The framework is subsequently extended to multimodal classification (M-GCECDL), where each modality is treated as a source with learned reliability, enabling adaptive fusion that prioritizes the most trustworthy modalities. Second, to analyze and visualize nonstationary crop dynamics, an unsupervised UMAP-based Local Biplot (UL-Biplot) method is developed. This approach integrates manifold learning with localized linear projections to uncover interpretable patterns of crop water status across different phenological stages. Finally, to reconcile predictive accuracy with transparency, a hybrid TabNet-informed UL-Biplot model is presented, which combines an attention-based deep learning architecture with the UL-Biplot to provide both global and instance-level feature relevance attribution. The proposed frameworks were validated on synthetic benchmarks and three real-world agricultural datasets, including the RiceClimaRemote multimodal dataset from the Tolima region of Colombia. The GCECDL framework demonstrated superior performance, outperforming state-of-the-art methods on 12 out of 15 benchmark datasets. The TabNet-informed UL-Biplot achieved high predictive accuracy, with [Mathematical Formula] $R^2 = 0.77 \pm 0.01$ for breeder score estimation in forage grasses and [Mathematical Formula] $R^2 = 0.79 \pm 0.01$ for Leaf Area Index (LAI) estimation in wheat, while successfully identifying key agronomically relevant predictors such as VEG, RCC, NDVI, and PSRI. The interpretability analysis confirmed that the models learned to dynamically fuse information, correctly prioritizing clean data domains and providing visually coherent, context-aware insights into crop health. By ensuring high predictive accuracy in estimating several crop water status agronomic traits, the proposed models contribute directly to precision agriculture practices. These advances support site-specific irrigation management, crop monitoring, and breeding programs, strengthening agriculture’s capacity to adapt under climate variability. This research establishes a robust groundwork for developing transparent, noise-resistant, and adaptable DL models for multimodal remote sensing in agriculture. Future work will extend these frameworks toward multimodal transformer architectures, complementing the proposed TabNet-informed Local Biplot for tabular data with the incorporation of visual transformers for image-based modalities and temporal modeling, to enhance the scalability and impact of data-driven precision agriculture.
Enhanced interpretability using regression networks for assessing domain dependences in motor imagery
(Universidad Nacional de Colombia, 2025) Gómez Morales, Óscar Wladimir; Castellanos Domínguez, César Germán; Álvarez Meza, Andrés Marino; Gómez Morales, Óscar Wladimir [0000000346547231]; Grupo de Control y Procesamiento Digital de Señales
The growing demand for efficient and accessible brain-computer interface (BCI) systems has driven research into methods that enhance the interpretation of electroencephalographic (EEG) signals. In this context, improving the accuracy and transparency of computational models represents a crucial challenge for advancing applications such as neurorehabilitation and human–machine interaction. However, three fundamental problems persist in the development of BCI systems: (a) the non-stationarity of EEG signals, which compromises model stability; (b) the high intra- and inter-subject variability in EEG recordings, which limits generalizability; and (c) the low physiological interpretability of the models used, which hinders neuroscientific validation. To address these challenges, this thesis proposes a methodological framework based on machine learning models and functional connectivity techniques, aimed at optimizing the prediction, classification, and interpretation of EEG signals in BCI systems. Three strategies were developed, aligned with the aforementioned problems: (i) a regularized regression model to predict full-channel EEG from a reduced subset of electrodes, mitigating nonstationarity and improving classification performance; (ii) a supervised feature extraction strategy to predict musical emotions in MIDI format, applying domain alignment and neural clustering techniques to address intra- and inter-subject variability; and (iii) a functional connectivity analysis approach using the Weighted Phase Lag Index (wPLI) and graph-based representations, aimed at enhancing interpretability through the detection of pre-training brain desynchronization patterns. The results show that the predictive model based on Elastic Net regression achieved an average classification accuracy of 78.16%, outperforming traditional approaches using either reduced or full-channel data. In the second objective, the combination of domain alignment and latent encoding improved emotional prediction and musical synthesis in MIDI, preserving both temporal and tonal coherence. Lastly, the functional connectivity analysis enabled the identification of key brain regions involved in desynchronization, associated with different levels of proficiency in BCI paradigms. The most relevant conclusions include: (i) regularized regression can estimate high-fidelity signals from a small number of electrodes, promoting the development of more cost-effective and portable systems; (ii) shared representations between EEG and musical data enable more accurate emotional prediction, contributing to the design of personalized neurointeraction environments; and (iii) the use of functional connectivity metrics provides validated physiological explanations, enhancing confidence in classification models. As future work, the proposed methodology will be validated using broader and more diverse datasets, and adversarial generative architectures will be incorporated to strengthen transferability across subjects and experimental conditions.
Propuesta de escenarios de digitalización para integración de demanda activa en sistemas de distribución utilizando técnicas metaheurísticas
(Universidad Nacional de Colombia, 2025) Obando Paredes, Edgar Darío; Carvajal Quintero, Sandra Ximena; Obando Paredes, Edgar Darío [0001493475]; Obando Paredes, Edgar Darío [sGk6rtQAAAAJ&hl]; Obando Paredes, Edgar Darío [0000000225157640]; Environmental Energy and Education Policy E3P; Obando Paredes, Edgar Darío [57211790612]
Esta tesis de investigación doctoral aborda el desafío de integrar la demanda activa en sistemas de distribución eléctrica mediante un enfoque de digitalización progresiva basado en escenarios. La investigación está estructurada en cinco momentos clave: revisión del estado del arte, identificación de brechas en la normatividad, establecimiento de la metodología, definición de escenarios progresivos y modelado de cada escenario. En la primera etapa, se realizó una revisión del estado del arte, identificando las principales tendencias y limitaciones en la gestión de la demanda activa y la integración de DER´s. Simultáneamente, se analizó la normatividad vigente, revelando importantes brechas que justifican la necesidad de un enfoque formal para la integración de tecnologías digitales. A partir de estas brechas, se establece una metodología por fases que guía la transición hacia sistemas inteligentes de distribución. Los escenarios definidos representan diferentes niveles de digitalización, desde la supervisión básica hasta el control avanzado basado en datos. Cada uno fue modelado matemáticamente para minimizar costos operativos y penalizaciones, maximizando la integración de fuentes renovables y optimizando el uso de infraestructuras tecnológicas habilitadoras disponibles. Los modelos diseñados incorporan técnicas y tecnologías emergentes para gestionar la demanda de manera dinámica. Finalmente, se validaron los modelos en un caso de estudio aplicado a la ciudad de Pasto, analizando el impacto de la digitalización en la sostenibilidad y la democratización energética (Texto tomado de la fuente)
Modelo matemático para la toma de decisiones con respecto a inversiones sostenibles
(Universida Nacional de Colombia, 2025) Cholo Camargo, Ingrid Milena; Olivat Tost, Gerard; Peña, Carlos; Angulo García, Fabiola; Percepción y Control Inteligente (Pci)
Lo ideal en cualquier sociedad es poder tener una buena calidad de vida, usar los recursos de manera adecuada sin dañar el medio ambiente y asegurar que estos recursos sean suficientes para las futuras generaciones. Una herramienta muy útil que ayuda a modelar este tipo de situaciones para entender su dinámica, es el modelo matemático. Por ello, se propone un modelo matemático que estudie la interacción entre el hombre y los recursos naturales, con el objetivo de optimizar la gestión sostenible de los recursos naturales y el crecimiento poblacional, ya que, una adecuada administración del recurso reduce el riesgo de su desaparición. Se considera una población aislada con parámetros constante para simplificar el análisis. Se usan técnicas de análisis de estabilidad, bifurcaciones, teoría de viabilidad y control óptimo para identificar condiciones iniciales bajo las cuales se puede lograr un equilibrio sostenible. Además, conociendo la condición inicial del sistema y un umbral mínimo de personas, se han encontrado umbrales sostenibles del bosque que pueden mantenerse. También se ha identificado el umbral máximo sostenible del bosque y la manera de mantenerlo. Se ha demostrado que, si se cuenta con una reserva mínima de bosque y/o una actividad económica adicional, la utilidad máxima será mayor. Los hallazgos sirven como base para diseñar políticas públicas que promuevan el uso sostenible de los recursos en sectores como la agricultura, la pesca y la silvicultura, promoviendo la regeneración de los recursos y el bienestar de las comunidades (Texto tomado de la fuente)
Factorización no negativa de matrices e índices acústicos para el análisis perceptual y la caracterización de paisajes sonoros
(Universidad Nacional de Colombia, 2025) Martínez Tabares, Fernando; Orozco Alzate, Mauricio; Martínez Tabares, Fernando [0000000322822016]; Grupo de Control y Procesamiento Digital de Señales
En esta tesis se presentan contribuciones innovadoras para el análisis de paisajes sonoros, enfocándose en la factorización no negativa de matrices (NMF), y en el uso del índice de complejidad acústica (ACI) como guía para mejorar la interpretabilidad y detección de eventos sonoros. El marco teórico incluye aspectos clave sobre aproximaciones numéricas, notación de orden de convergencia, aproximaciones de derivadas e integrales basadas en el polinomio de Taylor y conceptos básicos sobre funciones de variación acotada. Estos conceptos son fundamentales para las interpretaciones matemáticas del ACI presentadas en la tesis. Así mismo, se cubre el análisis tiempo-frecuencia, explicando los conceptos básicos sobre la transformada de Fourier de tiempo corto (STFT), que se utiliza como la matriz de entrada para la NMF. Además, se revisan índices acústicos comunes en ecoacústica y bioacústica, que se emplean en capítulos posteriores. El primer conjunto de resultados, correspondiente a la selección de índices acústicos, incluye enfoques supervisados y no supervisados para identificar índices relevantes en la caracterización de paisajes sonoros colombianos, destacándose el ACI como uno de los más representativos. Este análisis fundamenta la importancia del ACI, que es utilizado como guía en capítulos posteriores. El capítulo sobre interpretaciones matemáticas del ACI presenta contribuciones significativas al ofrecer interpretaciones tanto geométricas como analíticas del ACI. La relación entre el ACI y el concepto de variación total permitió proponer una metodología más eficiente para su cálculo. Finalmente, en el capítulo sobre sparse NMF basado en el ACI, se desarrolla un enfoque novedoso en el que se integra el ACI como función de regularización en la NMF, permitiendo una reducción de dimensionalidad más interpretable (Texto tomado de la fuente).
Design of a photovoltaic microinverter for active and reactive power injection
(Universidad Nacional de Colombia, 2025) Bolaños Navarrete, Mario Andrés; Osorio Londoño, Gustavo Adolfo; Bastidas Rodriguez, Juan David; Percepción y Control Inteligente (Pci)
The integration of photovoltaic (PV) systems into the power grid marks a significant stride towards renewable energy adoption, necessitating advancements in power conversion technologies. An important technology to this development are microinverters, which facilitate DC to AC conversion directly at the PV panel level. This dissertation explores the design, modeling in small and large signal, and implementation of photovoltaic microinverters with a focus on their capabilities for active and reactive power injection, a feature that enhances grid stability and efficiency. Microinverters offer a multitude of advantages over traditional centralized inverter systems. By optimizing the performance of individual solar panels, they mitigate issues such as shading and ensure a more reliable and efficient energy production. The ability of microinverters to inject reactive power into the grid is particularly beneficial, as it aids in voltage regulation and grid stability. The dissertation presents an state-of-the-art review of inverter technologies capable of reactive power control, including a detailed examination of a commercial inverter and its computational modeling. This groundwork helps for understanding the current landscape of inverter technology. The exploration of power converters for DC-DC conversion, such as the Flyback converter and a resonant converter variant, highlight the objectives of this dissertation in improving the efficiency of PV systems. A novel nonlinear averaged model for the Flyback converter, developed as part of this research, facilitates the simulations of power systems in extended time without increment the computational complexity. The detailed analysis of the resonant converter highlights its suitability for optimizing PV system performance through Maximum Power Point Tracking (MPPT) algorithms while maintain high efficiency. Further, the dissertation introduces two inverter (DC-AC) designs: the Dual Buck (DB) and the Full Bridge (FB) inverters. These designs are analyzed due to its extended use and increased efficiency when operated with unipolar commutation. Also, an innovative commutation technique for the FB to address zero-crossing distortion is presented, exemplifiying the potential of this inverter technologies to maintain the benefits of the Dual Buck design while enhancing grid stability and power quality. The analysis also shows the advantages of FB over DB in terms of grid support functionalities. Integrating the DC-DC and DC-AC stages, the dissertation culminates in a comprehensive microinverter design. This integration, detailed through control techniques and four-quadrant switching analysis for reactive power injection, confirms the microinverter’s capability to support grid stability. The theoretical models and control strategies proposed are substantiated through simulation results and experimental setups, validating the practical applicability of the research findings. In this way, this dissertation contributes to the field of renewable energy by advancing the understanding and capabilities of microinverters, not only as efficient power converters but also as key components for ensuring the stability and reliability of energy networks. The research also presents a design and control concepts that contribute to the construction of robust microinvertes based in power converters (Texto tomado de la fuente).
Diseño de una metodología para el mejoramiento de procesos odontológicos basados en técnicas de segmentación en odontología digital
(Universidad Nacional de Colombia, 2024) Tamayo-Quintero, Juan David; Gómez-Mendoza, Juan Bernardo; Guevara Perez, Sonia V.; Tamayo-Quintero, Juan David [https://scienti.minciencias.gov.co/cvlac/visualizador/generarCurriculoCv.do?cod_rh=0001523579]; Tamayo-Quintero, Juan David [https://scholar.google.com.co/citations?user=Bkfzfk4AAAAJ&hl=en]; Tamayo-Quintero, Juan David [https://orcid.org/0000-0002-0159-1573]; SHAC - Computación aplicada suave y dura
This thesis proposes a methodology for supporting orthodontic treatments through image processing techniques in digital dentistry. Led by Juan David Tamayo Quintero at the Universidad Nacional de Colombia, the research aims to improve orthodontic processes using artificial vision and image processing. The study focuses on developing tools for dental arch shape classification and analysis, employing image processing methods. The methodology integrates data structures for dental models, including various file formats like PLY, STL, OBJ, and PCD. Point cloud data representation is explored, emphasizing neighborhood and curvature estimation techniques. The document introduces the creation of a database comprising digitized dental models and outlines arch form classification methodologies, including ovoid, tapered, and square forms, along with morphometric templates. Ground truth determination methods and statistical analyses for evaluating model accuracy are also presented. The data analysis and implementation section details techniques for representing data in a 2.5D space, image projections, and segmentation methods, including watershed segmentation. Two software tools, DentalArch and DentiShape3D, were developed to aid in dental arch analysis, incorporating machine learning and image processing techniques. The conclusion reflects on achievements in image processing and database construction, highlighting future work areas such as enhancing sample diversity and reducing evaluator biases. The thesis provides a contribution to the field of digital dentistry, offering methodologies and tools to improve orthodontic processes through image processing techniques (Texto tomado de la fuente).
Regularized Gaussian functional connectivity network with Post-Hoc interpretation for improved EEG-based motor imagery-BCI classification
(Universidad Nacional de Colombia, 2024) Garcia Murillo, Daniel Guillermo; Álvarez Meza, Andrés Marino; Castellanos Domínguez, César Germán; DG García-Murillo; Grupo de Control y Procesamiento Digital de Señales
Conditions such as amyotrophic lateral sclerosis, brain strokes, and spinal injuries can cause significant disruptions in the brain's communication pathways, affecting voluntary muscle control and interactions with one's environment. Brain-computer interfaces (BCIs) are computer-based systems utilized to address these challenges; they interpret brain signals and have seen extensive applications in medical technology, rehabilitation, and entertainment industries. Particularly in BCI studies, motor imagery (MI), a widely researched paradigm, holds the potential for reviving motor functionality. Integrating MI-BCIs with skilled therapists has yielded positive outcomes in sensory and motor rehabilitation processes, proving to be beneficial for individuals with neurological disorders. Additionally, MI-BCI systems are not limited to clinical applications; they also find relevance in non-clinical settings like virtual reality, gaming, and skill acquisition. Brain signals in MI-BCI systems are often analyzed using electroencephalography (EEG), which is advantageous due to its high temporal resolution, affordability, and portability. These features make it a viable method for capturing short-evolving events and temporal neuronal activity patterns. Nonetheless, decoding EEG signals is complex due to their high sampling rate and the high number of electrodes involved, which results in an overwhelming amount of data points. Several promising methodologies, leveraging the concept of event-related synchronization (ERS) and event-related desynchronization (ERD), have been developed to ease the extraction of relevant features for decoding EEG into sensorimotor rhythm (SMR) patterns. While noteworthy, single-channel features tend to oversimplify overall phenomena since they need to consider the interactions of different brain regions. Considering the complex activations and communications within multiple brain areas during the execution or imagination of simple motor tasks, multi-channel feature extraction approaches are more accurate. For instance, the Common Spatial Patterns (CSP) method has demonstrated its significance in MI-BCI, promoting the extraction of highly discriminative features by maximizing the separability of SMR features. Brain connectivity (BC), a key determinant in this approach, describes interactions within and between brain regions and correlates to synchronization mechanisms within oscillatory modulations. BC can be divided into structural/anatomical connectivity (SC), effective connectivity (EC), and functional connectivity (FC), each with distinct characteristics. SC focuses on physical connections and is context-independent, while EC and FC, suitable for EEG-based MI-BCI systems, decode fast-evolving cognitive states where FC is particularly suited to MI-BCI applications due to its simplicity, low computational demands, and lack of rigid prior assumptions. This thesis addresses three prevailing challenges experienced with FC EEG-based MI-BCI systems. Firstly, performance deficits in MI classification tasks are often a result of high noise levels present in EEG signals. Secondly, relying on expert knowledge for feature extraction often presents a barrier. Lastly, there is a necessity for increased transparency within these models. The study delivers key contributions toward a single-trial FC in MI-BCI systems in response to these challenges. This improvement is achieved by utilizing a novel approach grounded on the kernel cross-spectral approximation afforded by extending Bochner's theorem. This new single-trial kernel cross-spectral (KCS) FC estimator tackles noise issues, spurious connections, and inter-subject variability, significantly improving accuracy. Moreover, an end-to-end approach, KCS-FCnet, has been conceptualized to enhance automatic EEG representation for MI-BCI. This approach is pivotal in reducing artifacts and accurately identifying relevant connections. In addition, it assists in the extraction of critical spectral, temporal, and spatial EEG representations. Acknowledging the necessity for model transparency, this study proposed a strategy for improving model transparency and interpretability, named interpretable regularized KCS-FCnet (IRKCS-FCnet). Capitalizing on the potential of regularization, the model's interpretability is strengthened by minimizing nonrelevant connections. This is further made possible by maximizing the cross-information potential using Renyi's entropy measured at $\alpha=2$, alongside a cross-entropy function. To supplement these enhancements, Layer-CAM and last layer weight strategies have been employed to provide both post-hoc and intrinsic interpretability, respectively. The average drop concept further supports these strategies, enhancing the system's transparency and understanding. The results show that the proposed KCS-FC achieves higher performance accuracy than classical FC estimators and has competitive accuracy against more complex FC-based strategies, reaching the second place in DBI\footnote{BCI2a~\url{http://www.bbci.de/competition/iv/index.html}} with an accuracy of $81.92\%$ and first place with $74.12\%$ in DBIII\footnote{Giga Motor Imager~\url{http://gigadb.org/dataset/100295}}, both binary motor imagery classification tasks. Moreover, the KCS-FCnet shows $76.4\%$ on DBIII while keeping the architecture with $4245$ trainable parameters, making it the highest against the compared DL models. Lastly, the IRKCS-FCnet shows better performance, increasing $2$ percentage points for low-performing individuals, and showed better localization and interpretations, demonstrating that when removing $5\%$ of the most important features, the average accuracy drop score goes to $22.46$, $5$ percentage points higher than the version without the regularization. Furthermore, the contralateral pattern can be seen in the topoplot interpretation for both the right and left hand. This study opens the path for using more FC estimators within DL frameworks, and future work can extend and improve some of the existing issues, for instance, including multivariate measuring strategies and not relying only on pairwise measures, as well as implementing a time-lag method to reduce the impact of volume conduction problems. Finally, measuring FC matrices with more appropriate distances that account for the inner links between different channels, like in the Riemann geometry, could help to better decode MI signals hindered in the intricate EEG recordings (Texto tomado de la fuente)
A machine learning framework for EEG-based BCI-MI skill prediction with preserved explainability
(Universidad Nacional de Colombia, 2024) Caicedo Acosta, Julian Camilo; Castellanos Dominguez, Cesar German; Alvarez Meza, Andres Marino; Grupo de Control y Procesamiento Digital de Señales
The understanding of brain functioning, including its structure and the dynamic networks generated under different situations, has been one of the most challenging research topics in recent years. In this sense, brain- computer interfaces (BCI) have become the main systems to acquire and process brain signals, allowing the development of increasingly specialized techniques like machine learning and artificial intelligence methods in order to replace, restore, enhance, supply, and improve brain functionality. Thus, BCI systems may be a promissory tool for several fields, including health, rehabilitation, education, marketing, and gaming, among others. Besides, BCI development is a continuously growing field of study not only in research but also in the business context, with a market that projects more than four hundred million USD by 2029. However, despite advances in building BCI systems, around 30% of BCI users are unable to effectively manage the device due to multiple factors such as intra- and inter-subject variability, overfitting, and small datasets. This issue is known as the ‘BCI illiteracy phenomenon’. Since electroencephalography (EEG) is the most used acquisition method due to its high temporal resolution, portability, and low cost compared with the other techniques, in this work we study the causes of the illiteracy phenomenon by predicting the user performance during motor imagery (MI) tasks. In this sense, we developed a machine learning framework to improve the illiterate subject’s performance under a skill prediction scheme. First, we develop a data- driven deep learning model, termed DRN, based on extracting time-frequency neurophysiological indicators that allow coding the inter-subject variability by predicting MI-BCI performance. Besides, to address the overfitting issues in the presence of small datasets and high-dimensional representations, we also developed a functional connectivity-based Monte-Carlo dropout regularized approach by modifying the proposed DRN to extract relevant patterns from channel relationships. Last, to improve illiterate subjects’ performance, we develop a general-purpose end-to-end multi-task deep learning approach founded on an autoencoder- based regularization scheme that transfers learned knowledge from performance prediction tasks to MI classification tasks. The results obtained in this work are promising and outperform the baseline methods in both BCI-performance prediction and MI classification. Furthermore, the proposed prediction methods are able to find behavioral group patterns between subjects with similar degrees of variability, while the transfer learning techniques allow to improve the performance of illiterate subjects with information from other subjects, contributing to the generalization of BCI systems (Texto tomado de la fuente)
Coordinación de protecciones : Una solución desde la calidad de la potencia
(Universidad Nacional de Colombia, 2023) Arias Guzmán, Santiago; Ustariz Farfan, Armando Jaime; Cano Plata, Eduardo Antonio; Arias Guzmán, Santiago [https://scienti.minciencias.gov.co/cvlac/visualizador/generarCurriculoCv.do?cod_rh=0001455252]; Arias Guzmán, Santiago [https://scholar.google.com/citations?user=b4hchf8AAAAJ&hl=es]; Arias Guzmán, Santiago [0000000172046401]; Arias Guzmán, Santiago [https://www.researchgate.net/profile/S-Arias-Guzman]; Grupo de Investigación en Calidad de la Energía y Electrónica de Potencia; Redes de Distribución y Potencia Gredyp
LA TESIS presenta, inicialmente, una revisión crítica del estado del arte de la protección de los sistemas eléctricos, incluyendo las funciones, esquemas y sistemas de protección. A partir de esta revisión se presentan las limitaciones actuales en los sistemas de protección, así como las tendencias para su corrección. Para superar estas limitaciones, en esta tesis se propone un nuevo sistema de protección basado en el análisis de la calidad de la potencia eléctrica; utilizando el formalismo matemático del tensor instantáneo de potencia se define un nuevo indicador de desviación global de las condiciones ideales de un sistema sinusoidal y balanceado. Mediante el nuevo indicador definido en esta tesis, se detectan y clasifican condiciones de falla a ser despejadas. El sistema de protección propuesto logra evitar el proceso de coordinación de protecciones, mejora la selectividad de fallas de alta impedancia de característica lineal y no lineal, y logra la mitigación del riesgo de arco eléctrico (Texto tomado de la fuente)
Modeling and simulation of photovoltaic systems under partial shading conditions
(Universidad Nacional de Colombia, 2023) Restrepo Cuestas, Bonie Johana; Ramos-Paja, Carlos Andres; Trejos Grisales, Luz Adriana; Restrepo Cuestas, Bonie Johana [0000491233]; Restrepo Cuestas, Bonie Johana [https://scholar.google.es/citations?user=aBg8s0QAAAAJ&hl=en]; Restrepo Cuestas, Bonie Johana [0000000152761651]; Restrepo Cuestas, Bonie Johana [https://www.researchgate.net/profile/Bonie-Restrepo-2]; Grupo de Automática de la Universidad Nacional Gaunal
This thesis introduces a methodology for modeling commercial photovoltaic panels at the cell level operating under partial shading conditions. In the first part, a review of the literature is presented, focusing on the proper representation of the current-voltage characteristics in both forward and reverse bias, the mathematical formulation, the circuit model, and the estimation of parameters for photovoltaic cells. In the second part, the single diode model (SDM), the direct-reverse model (DRM), and Bishop’s model are introduced, emphasizing their current-voltage relationship, mathematical formulation, circuit model, and parameter requirements. In the third part of the thesis, a procedure to obtain I-V curves in panel terminals without the need for any physical intervention is detailed. This procedure is necessary to compare the behavior of the three models analyzed in both quadrants. The procedure requires a panel without a bypass diode and measurement equipment capable of acquiring current, voltage, temperature, and irradiation. After considering the evaluation of some metrics such as root mean square error (RMSE) and mean absolute percentage error (MAPE), Bishop’s model is selected for use in the methodology. In the fourth part, a methodology to estimate the parameters of Bishop’s model is proposed, which formulates the estimation of the parameters as an optimization problem. The metho- dology uses a genetic algorithm, and it is validated using information from two commercial panels. The curve reconstructions for each technology are evaluated using metrics such as RMSE and MAPE to assess the accuracy of the models (Texto tomado de la fuente)
Modelado de brotes epidémicos de dengue para la toma de decisiones en salud pública : Efecto de la movilidad en el departamento de Caldas
(Universidad Nacional de Colombia, 2021) Ospina Aguirre, Carolina; Olivar-Tost, Gerard; Osorio Londoño, Gustavo Adolfo; Carolina Ospina Aguirre; https://scholar.google.com/citations?user=F3rcKnMAAAAJ&hl=en; Ospina Aguirre, Carolina [0000000339924289]; https://www.researchgate.net/profile/Carolina-Ospina; AbcDynamics
En este trabajo se presenta la formulación de un modelo compartimental en el que se utilizan ecuaciones diferenciales y redes complejas para representar la dinámica de transmisión del dengue en el departamento de Caldas. La población está dividida en cuatro compartimentos: susceptibles, infectados, hospitalizados y recuperados; y los mosquitos que transmiten la enfermedad en dos: susceptibles e infectados. Se explora el efecto de las lluvias, de aplicar medidas de control, de la hospitalización y de la movilidad sobre la cantidad de personas infectadas. En el departamento de Caldas, hay dos temporadas de lluvias al a ̃no, las cuales fueron simuladas generando un aumento en la población de mosquitos. Se encuentra que el incremento de las precipitaciones incrementa los casos de dengue en 5, 45 %. Las medidas de control vectorial analizadas son fumigación y eliminación de criaderos. Se encontró que el uso conjunto de estas medidas tiene un efecto reductor mayor en la cantidad de infectados que si se aplican de manera individual. La hospitalización temprana del 20 % los contagiados de dengue redujo en un 17,83 % la cantidad total de casos en el departamento. La red compleja implementada para modelar el transporte vehicular define la probabilidad de movilidad en- tre un parche y otro mediante una matriz de tasa de transición. Esta matriz se calcula con base en un modelo gravitacional. La estimación de los parámetros del modelo, fue realizada con datos reales de cada uno de los municipios incluidos en este estudio, esto es, los 27 de Caldas y los 7 municipios vecinos que tienen conexión terrestre directa con algún municipio del departamento. Los casos de dengue obtenidos cuando los municipios están conectados, es decir, que hay movilidad de personas, incrementaron un 83,17 % respecto a los resultados obtenidos cuando no había movilidad. Se pudo observar que cada municipio es afectado de manera diferente por el movimiento de sus residentes. En aquellos donde la incidencia de dengue es alta y una proporción de sus habitantes se desplazan a zonas de menor incidencia se presenta una disminución en la cantidad de infectados. Los habitantes de municipios sin casos de dengue contraen la enfermedad al desplazarse a zonas con presencia de la enfermedad. Es por esto que se propone restringir el acceso a municipios endémicos durante un brote para disminuir la cantidad total de casos en el departamento (Texto tomado de la fuente)
Metodología para el reconocimiento de patrones sísmico-volcánicos no estacionarios mediante técnicas de aprendizaje adaptativo
(Universidad Nacional de Colombia, 2023) Castro Cabrera, Paola Alexandra; Orozco-Alzate, Mauricio; Londoño Bonilla, John Makario; Paola Alexandra Castro; Paola Alexandra Castro Cabrera; P.A.Castro-Cabrera [0000-0002-4442-0715]; Paola Alexandra Castro Cabrera; Grupo de Control y Procesamiento Digital de Señales; Castro-Cabrera Paola Alexandra [36717473000]
El monitoreo volcánico constituye una tarea imprescindible en el contexto de prevención y gestión del riesgo; en este sentido, los observatorios vulcanológicos y sismológicos cumplen una misión trascendental en la declaración de alertas tempranas de erupción volcánica. Y dentro de esta labor, la correcta clasificación de la sismicidad representa un insumo indispensable para la interpretación del fenómeno volcánico y la caracterización de dinámicas eruptivas; por tal motivo, es necesario que la clasificación se lleve a cabo de manera ágil y confiable. A través de la sismicidad correctamente etiquetada, los expertos analistas pueden caracterizar los procesos que estarían ocurriendo al interior de un volcán, e identificar precursores de una erupción. Sin embargo, la acertada discriminación de eventos sísmicos suele verse afectada por la migración de fuentes sísmicas, alteraciones en la dinámica de fluidos, cambios en los mecanismos de generación de grietas, entre otras situaciones, que pueden modificar la distribución de probabilidad de los registros sísmicos (cambios de concepto), y por tanto, incrementar la no estacionariedad de estas señales. Durante las últimas dos décadas, en las áreas de Aprendizaje Automático y Reconocimiento de Patrones se han desarrollado múltiples técnicas y herramientas aplicadas a enfoques de representación y clasificación de sismos volcánicos, entre las cuales destacan las redes neuronales, las máquinas de vectores de soporte, los modelos ocultos de Markov, entre otros, enmarcados (incluso) en contexto muy actuales como el Aprendizaje Profundo. En general, los estudios hallados al respecto en el estado del arte muestran resultados optimistas; sin embargo, se detalla que éstos son consecuencia de configuraciones experimentales restrictivas que disminuyen la complejidad del problema de clasificación planteado; una condición común es el uso de datos procedentes de periodos cortos de registro y poco representativos de la actividad volcánica. Esta limitación simula un entorno estacionario donde los modelos predictivos tradicionales funcionan eficazmente, pero que van en detrimento al actuar por un tiempo prolongado cuando los cambios de concepto se hacen evidentes. Siendo notable la necesidad de disponer de sistemas automáticos de clasificación que satisfagan las ``condiciones realistas'' del problema, como requerimiento esencial en la vigilancia volcánica, en esta tesis se propone el desarrollo de una metodología de reconocimiento de patrones sísmicos, a partir de registros de eventos volcánicos, que considere la adaptación de la clasificación a entornos y condiciones realistas y cambiantes. Para ello, se diseñó un modelo de clasificación centrado en el área del aprendizaje adaptativo y basado en aprendizaje incremental (aún no explorados en datos sísmicos), con el cual se trata el paradigma del cambio del concepto, de tal manera que algunas propiedades como la recurrencia continua de datos adquiridos, la naturaleza multiclase de los registros, los efectos geológicos y las restricciones de generalización en la clasificación, sean contempladas, aprovechadas y eventualmente contrarrestadas al momento de hacer la clasificación automática de los sismos (Texto tomado de la fuente)
Planificación de trayectorias óptimas de robots móviles empleando el mapeo de celdas
(Universidad Nacional de Colombia, 2023) Grisales Ramírez, Efraín; Osorio Londoño, Gustavo Adolfo; Grisales Ramírez, Efraín [0009000814296566]; Percepción y Control Inteligente (Pci)
El problema de planificación de trayectorias es de importancia dentro de la robótica móvil, puesto que es indispensable generar caminos de referencia que eviten obstáculos y de fácil seguimiento por los sistemas autónomos. La solución dada por los algoritmos de planificación puede ser de diferente naturaleza, aumentando o disminuyendo la complejidad computacional. En este documento se propone un método para la planificación de trayectorias de robots móviles empleando el algoritmo de mapeo de celdas (AMC), el cual contiene las restricciones cinemáticas y dinámicas del robot, así como las del entorno, permitiendo generar trayectos de fácil seguimiento por el sistema. También contiene la información necesaria con el fin de dar solución al problema de optimización multiobjetivo. El algoritmo consiste en dividir el espacio de estados en celdas y luego simular la evolución del sistema dinámico del móvil para cada condición inicial; con esto se generan grafos que contienen la información de las conexiones entre celdas en términos de distancia, tiempo, esfuerzo de control y energía. Con esta información se solucionan los problemas de optimización simple o multiobjetivo, según sea la función de costo empleada para encontrar el camino más corto entre dos puntos con el algoritmo de Dijkstra. Además, en este documento se introducen algunos conceptos de control óptimo que serán empleados en la planificación de trayectorias de robots móviles. Estos conceptos se aplican inicialmente al doble integrador (DI) en tiempo continuo, debido a que éste emula el comportamiento de un robot sin masa. Posteriormente se implementa el AMC para solucionar los mismos problemas de optimización en forma discreta para el movimiento en 1 y 2 dimensiones, y se extiende el concepto a la aplicación en robots móviles de guiado diferencial, así como para el Robotino de FESTO. (Texto tomado de la fuente)
Deep learning framework with enhanced interpretability for classification of motor imagery tasks
(Universidad Nacional de Colombia, 2022) Collazos Huertas, Diego Fabian; Castellanos-Dominguez, German; Collazos Huertas, Diego Fabian [https://scienti.minciencias.gov.co/cvlac/visualizador/generarCurriculoCv.do?cod_rh=0000017335]; Collazos Huertas, Diego Fabian [D.F Collazos-Huertas]; Collazos Huertas, Diego Fabian [0002-0434-3444]; Collazos Huertas, Diego Fabian [https://www.researchgate.net/profile/Diego-Collazos]; Grupo de Control y Procesamiento Digital de Señales
Deep learning (DL) allows models composed of multiple processing layers to learn representations of data with several levels of abstraction. These methods have improved state-of-the-art tasks like speech recognition, visual object identification, and many other fields. Regarding electroencephalographic (EEG) signals analysis, especially for the Motor Imagery (MI) paradigm, the availability of large data sets and advances in machine learning have led to the deployment of DL architectures, allowing the understanding of the information that may contain for brain functionality. However, these models suffer some limitations in practice: i) often DL models not integrate properly EEG spatial information with extracted time-frequency features, ii) the resulting inter and intra-subject variability, along with frequently available small datasets, significantly decreases the performance of EEG-based MI systems, and iii) DL models are treated as “black boxes” lacking physiological interpretability. In this Ph.D. thesis proposal, we pretend to solve these issues i) developing a Deep&Wide learning methodology using multi-view feature extraction, ii) proposing a coupling information strategy based on transfer learning including subject’s clinical data, and iii) developing a relevance analysis methodology that allows improving the interpretability of neural responses. The detailed methodology and its respective execution plan (schedule) to carry out these objectives are further described. In addition, we list the available computational resources necessary for the proposed implementation (Texto tomado de la fuente)
Discriminating mental states by extracting relevant spatial patterns under non-stationary and subject-independent constraints
(Universidad Nacional de Colombia, 2023) Velásquez Martínez, Luisa Fernanda; Castellanos Dominguez, Cesar Germán; https://scholar.google.com.co/citations?user=kzOD4RQAAAAJ&hl=en; Grupo de Control y Procesamiento Digital de Señales
Evaluation of brain dynamics elicited by motor imagery (MI) tasks can contribute to clinical and learning applications. In this work, we propose four specific improvements for brain motor intention response analysis based on EEG recordings by considering the nonstationarity, nonlinearity of brain signals, inter- and intra-subject variability, aimed to provide physiological interpretability and the distintiveness between subjects neural response. Firstly, to build up the subject-level feature framework, a common representational space, is proposed that encodes the electrode (spatial) contribution, evolving through time and frequency domains. Three feature extraction methods were compared, providing insight into the possible limitations. Secondly, we present an Entropy-based method, termed \textit{VQEnt}, for estimation of ERD/S using quantized stochastic patterns as a symbolic space, aiming to improve their discriminability and physiological interpretability. The proposed method builds the probabilistic priors by assessing the Gaussian similarity between the input measured data and their reduced vector-quantized representation. The validating results of a bi-class imagine task database (left and right hand) prove that \textit{VQEnt} holds symbols that encode several neighboring samples, providing similar or even better accuracy than the other baseline sample-based algorithms of Entropy estimation. Besides, the performed ERD/S time-series are close enough to the trajectories extracted by the variational percentage of EEG signal power and fulfill the physiological MI paradigm. In BCI literate individuals, the \textit{VQEnt} estimator presents the most accurate outcomes at a lower amount of electrodes placed in the sensorimotor cortex so that reduced channel set directly involved with the MI paradigm is enough to discriminate between tasks, providing an accuracy similar to the performed by the whole electrode set. Thirdly, multi-subject analysis is to make inferences on the group/population level about the properties of MI brain activity. However, intrinsic neurophysiological variability of neural dynamics poses a challenge for devising efficient MI systems. Here, we develop a \textit{time-frequency} model for estimating the spatial relevance of common neural activity across subjects employing an introduced statistical thresholding rule. In deriving multi-subject spatial maps, we present a comparative analysis of three feature extraction methods: \textit{Common Spatial Patterns}, \textit{Functional Connectivity}, and \textit{Event-Related De/Synchronization}. In terms of interpretability, we evaluate the effectiveness in gathering MI data from collective populations by introducing two assumptions: \textit{i}) Non-linear assessment of the similarity between multi-subject data originating the subject-level dynamics; \textit{ii}) Assessment of time-varying brain network responses according to the ranking of individual accuracy performed in distinguishing distinct motor imagery tasks (left-hand versus right-hand). The obtained validation results indicate that the estimated collective dynamics differently reflect the flow of sensorimotor cortex activation, providing new insights into the evolution of MI responses. Lastly, we develop a data-driven estimator, termed {Deep Regression Network} (DRN), which jointly extracts and performs the regression analysis in order to assess the efficiency of the individual brain networks in practicing MI tasks. The proposed double-stage estimator initially learns a pool of deep patterns, extracted from the input data, in order to feed a neural regression model, allowing for infering the distinctiveness between subject assemblies having similar variability. The results, which were obtained on real-world MI data, prove that the DRN estimator fosters pre-training neural desynchronization and initial training synchronization to predict the bi-class accuracy response, thus providing a better understanding of the Brain--Computer Interface inefficiency of subjects. (Texto tomado de la fuente)
Toma de decisiones para la evaluación y priorización de proyectos verdes
(Universidad Nacional de Colombia, 2022) Castrillón Gómez, Julián Andrés; Olivar-Tost, Gerard; Valencia Calvo, Johnny; Angulo Garcia, Fabiola; Castrillón Gómez, Julián Andrés; Castrillon Gomez, Julian Andres [0000-0002-1598-137X]; Castrillón Gómez, Julián Andrés; Abc Dynamics
El presente documento de tesis doctoral, propone un enfoque de análisis sistémico y con este un resumen metodológico, para comprender el comportamiento ambiental, social y económico de una comunidad y, con base en ello, seleccionar y dar prioridad a los proyectos verdes (que propenden por la sostenibilidad) que puedan brindarle un mayor beneficio. Para este fin, se presenta una propuesta metodológica conformada por tres etapas. La primera, consistente en un grupo de ecuaciones diferenciales ordinarias, que conforman un modelo en dinámica de sistemas, cuyo objetivo es reunir las principales características de un sistema social sostenible, fundamentado en el uso del agua y el uso del suelo. La segunda, consistente en una serie de talleres participativos con la comunidad, cuyo objetivo es reunir información acerca de las necesidades, prioridades, expectativas, entre otros, que sus habitantes tienen en cuanto a las principales problemáticas relacionadas con el uso del agua y el uso del suelo. Y la tercera, consiste en una serie de ecuaciones y operaciones matriciales, que conforman el denominado proceso analítico de red, para jerarquizar los proyectos verdes. A lo largo del documento, se mostrará que el fundamento de la tesis se basa en el aprovechamiento de la dinámica de sistemas y los sistemas dinámicos para la creación de escenarios futuros; el modelado participativo, para involucrar a la ciudadanía, aprovechar su conocimiento y conocer sus necesidades y; el proceso ANP, para aplicar un método matemático que permita reducir el sesgo en las decisiones y seleccionar el mejor proyecto con base en los criterios y prioridades definidas. Además, se proponen dos casos de estudio con el fin de evidenciar los diferentes comportamientos del modelo cuando se varían los valores de los parámetros del sistema. Logrando así, que exista una sensibilidad en la selección de proyectos de acuerdo con la comunidad a la cual va dirigida, facilitando la toma de decisiones según cada caso. Hasta ahora, no se han encontrado en la literatura reportes sobre este uso conjunto de diferentes métodos para la selección de proyectos. Como primer resultado importante, se ha encontrado que, de acuerdo a la dinámica de cada comunidad, los tomadores de decisiones pueden preferir uno u otro proyecto y no sesgarse por la elección de aquel que consideran mejor a la luz de las propuestas presentadas. En este sentido, la planificación y el estudio de la dinámica de las comunidades frente a posibles escenarios, implica que la dinámica en la toma de decisiones podría verse alterada de acuerdo con las necesidades y prioridades que la misma comunidad da a sus problemáticas. Desde este trabajo, se propone entonces que, basados en una metodología donde se vinculan la dinámica de sistemas y el modelado participativo como métodos evaluativos y, el proceso analítico de red como método de priorización, se aprovechen como facilitadores en la toma de decisiones, respaldados en la pertinencia matemática de cada uno de ellos. (Texto tomado de la fuente)

Examinar

Envíos recientes