Tesis Doctorado en Ingeniería Eléctrica y Electrónica

URI permanente para esta colecciónhttps://repositorio.uninorte.edu.co/handle/10584/5686

Examinar

Mostrando 1 - 15 de 15

Synthesis of abnormal ECG signals using generative neural networks
(2026-08-11) Barbosa Casanova, Geiner Giovanny
Cardiovascular diseases (CVDs) represent the leading cause of mortality worldwide, with a disproportionately severe impact on low- and middle-income regions due to limited diagnostic infrastructure and a scarcity of early detection programs. Automated electrocardiogram (ECG) classification algorithms driven by deep learning (DL) offer a non-invasive, highly scalable solution to enhance cardiac diagnostic capabilities. However, the deployment of these diagnostic support tools is fundamentally hindered by the pervasive challenge of dataset imbalance, wherein critical abnormal cardiac conditions are heavily outnumbered by normal sinus rhythms, leading to biased algorithmic training and suboptimal sensitivity toward minority classes. This doctoral dissertation addresses this critical limitation by proposing, implementing, and rigorously validating advanced Generative Adversarial Network (GAN) frameworks designed for high-fidelity synthetic data augmentation. The research explores two distinct methodological paradigms: one-dimensional (1D) time-domain signals and two-dimensional (2D) time–frequency spectrogram representations. The first core phase of this research evaluates the time-domain paradigm through a systematic, comparative analysis of eight distinct GAN architectures: Basic GAN, Deep Convolutional GAN (DCGAN),Wasserstein GAN with Gradient Penalty (WGANGP), Deep Convolutional WGAN-GP (DCWGAN-GP), Conditional GAN (CGAN), Conditional DCGAN (CDCGAN), Conditional WGAN-GP (CWGAN-GP), and Conditional DCWGAN-GP (CDCWGAN-GP). Trained on a preprocessed version of the MIT-BIH Arrhythmia Database, the generative performance of these models is quantified using a comprehensive multi-metric evaluation framework encompassing morphological fidelity (Euclidean Distance and Dynamic Time Warping) and statistical similarity (Maximum Mean Discrepancy). Downstream validation via a proof-of-concept 1DCNN classifier demonstrates that augmenting the minority classes with simples generated by the optimized WGAN-GP architecture significantly outperforms traditional data-balancing methodologies—such as the Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling (ADASYN), and Random Undersampling (RUS)—maximizing the F1-Score without introducing morphological degradation or overfitting.
Modelado de degradación de baterías de litio para segunda vida en microrredes aisladas
(2026-07-23) Olivero Ortiz, Víctor José
La transición energética y la creciente disponibilidad de baterías de iones de litio retiradas de vehículos eléctricos han impulsado el interés por reutilizarlas en aplicaciones estacionarias. En las zonas no interconectadas (ZNI) de Colombia, donde el suministro eléctrico depende de microrredes aisladas y el recurso solar es abundante, estas baterías de segunda vida ofrecen una alternativa de almacenamiento de bajo costo. Sin embargo, su aprovechamiento seguro exige modelar con precisión su degradación y gestionar la marcada heterogeneidad de celdas envejecidas bajo condiciones tropicales. Esta tesis aborda ese problema mediante cuatro contribuciones articuladas. Primero, una revisión estructurada de la literatura identifica las familias de modelos de degradación empleadas para estimar el estado de salud (SOH, State of Health) y la vida útil remanente (RUL, Remaining Useful Life), y concluye que los enfoques híbridos físico-informados con componente bayesiano equilibran mejor precisión, robustez y cuantificación de incertidumbre. Segundo, se diseña y valida por simulación un algoritmo heterogéneo adaptativo de carga y descarga (HA-ACD) que gestiona activamente la dispersión en capacidad, tensión y resistencia interna de un banco de ocho celdas 18650 en serie, evitando la sobredescarga de las celdas más débiles. Tercero, se desarrolla un sistema de gestión de baterías (BMS, Battery Management System) consciente de la degradación, que integra un estimador de SOH basado en regresión por procesos gaussianos (GPR, Gaussian Process Regression) con núcleo Matérn 5/2 y un factor proactivo de reducción de corriente; el estimador alcanza un error de 1,63 % en SOH y elimina por completo las violaciones de la envolvente de operación segura. Cuarto, el modelo se valida sobre los conjuntos de datos NASA PCoE y CALCE CS2, y se aplica a un caso de estudio de microrred aislada que proyecta la vida útil del banco bajo distintos escenarios de operación. Los resultados muestran que acoplar la estimación probabilística de SOH con un control adaptativo extiende la vida útil aprovechable de las baterías de segunda vida sin comprometer la seguridad, y ofrecen una base cuantitativa para su despliegue en microrredes aisladas.
Artificial intelligence based system for mixed operation microgrids for non-interconnected areas
(Universidad del Norte, 2025) Cardozo Sarmiento, Darwin; Pardo González, Mauricio
The increasing penetration of renewable energy resources in isolated and weakly interconnected electrical systems has introduced significant operational challenges, particularly in ensuring supply continuity, minimizing power fluctuations, and improving energy efficiency under highly variable generation conditions. In this context, this doctoral thesis addresses the operational management—rather than the planning—of a small-scale microgrid corresponding to node 611, focusing on real-time decision-making strategies supported by predictive analytics. The proposed work presents the design, integration, and validation of an intelligent energy management system (EMS) enhanced by short-term photovoltaic generation forecasting based on lightweight machine learning and deep learning models. The forecasting framework evaluates multiple architectures, including multilayer perceptron (MLP), convolutional neural networks (CNN), long short-term memory networks (LSTM), and ensemblebased approaches, using standard performance metrics such as root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R²). The selected predictive model is subsequently embedded into the operational logic of the EMS to inform state transitions and power flow decisions. The EMS operates as a finite-state machine, where transitions are governed by real-time measurements, predicted renewable generation, battery state of charge, and load demand. This strategy aims to reduce operational perturbations perceived by end users, improve the utilization of renewable resources, and decrease dependency on the main grid during critical operating conditions. The system is implemented and evaluated through detailed simulations, considering realistic operating constraints and time scales consistent with microgrid control applications. Although Raspberry Pi platforms are referenced as representative lowcost controllers, the proposed approach is intentionally formulated to be applicable to a broader class of single-board computing (SBC) platforms suitable for embedded and edge-based energy management applications. Simulation results demonstrate that the integration of predictive models into the EMS contributes to smoother operational behavior, improved power balance, and enhanced resilience against renewable generation variability. The contributions of this thesis include: (i) a predictive-enhanced operational EMS tailored for small-scale microgrids, (ii) a comparative evaluation of lightweight forecasting models suitable for real-time deployment, and (iii) a reproducible framework that bridges data-driven prediction and controloriented energy management. The findings support the feasibility of incorporating artificial intelligence techniques into embedded microgrid controllers to improve operational performance without increasing system complexity or hardware requirements.
Chirp-based localization method in lorawan
(Universidad del Norte, 2022) Márquez Cantillo, Luz Enith; Calle Torres, María Gabriela
Internet de las cosas requiere conectividad de millones de dispositivos. La tecnología LoRaWAN nace para conectar dispositivos a largas distancias. Dentro de los retos de esta tecnología se encuentra la mejora de los métodos que permiten la ubicación de los dispositivos conectados sin utilizar GNSS/GPS. En el presente trabajo inicialmente se analiza el desempeño de LoRa en canales AWGN utilizando el método de la envolvente compleja. Después, se estudia el alcance de señales LoRa y su robustez ante altos niveles de interferencia con experimentos en largas distancias. Luego, se estudian distintos algoritmos de localización basados en aprendizaje automático (ML) y se adopta una técnica de aumento de datos para RSS. Más adelante, se evalúa el efecto de obstaculizar un enlace LoRa con un cuerpo humano y se determina cómo se deterioran tanto el RSS como la precisión en el método de localización. Por último, se presenta un método de estimación de distancia basado en la forma de onda de la señal chirp, propia del sistema LoRaWAN. Con el retardo calculado a partir de la forma de onda se alimenta un algoritmo fingerprinting para probar diversos modelos de aprendizaje automático de localización. El método propuesto permite una mejora respecto a la literatura existente en las estimaciones de la distancia entre un nodo y un gateway. El método propuesto tiene potencial para la localización, alcanzando en un entorno simulado una precisión cercana a la de los dispositivos GNSS/GPS, sin requerir hardware adicional o aumento en la complejidad de los nodos.
Machine Learning-based system to predict the performance of exploration algorithms for mobile robots
(Universidad del Norte, 2021) Caballero Tovar, Liesle Yail; Percybrooks Bolívar, Winston Spencer
The thesis presented here proposes a Machine Learning based method with optimized neural networks that evaluates exploration algorithms for mobile robots based on the prediction of a performance metric. Two applications around the prediction of the performance of exploration algorithms are implemented here. The first consists of a system that compares and selects the most appropriate exploration algorithm according to the experimental environment, without implementing or simulating additional experiments when other different experimental conditions are required to evaluate. The second application deals with the problem of determining a useful energy budget for a mobile robot in a given environment without having to carry out experimental measures for every possible exploration task. The method of prediction using Machine Learning with optimization strategies Hill Climbing-Random Restart or Tabu List shows an improvement in predictor performance. With these optimization strategies the percentage relative absolute error for Random Walk, Random Walk Without Step Back and Q_Learning exploration algorithms, it was reduced by 92.47%, 90.41% and 87.38% respectively. An experimental energy consumption was measured and compared with the prediction of our model. A success rate of 96.09% was obtained, suggesting an adequate energy predictor, which could be useful for energy budgeting in actual mobile robot applications. Many recent applications of Machine Learning focus on the use of big datasets for training, however, there are many interesting problems, where the available datasets are small. Our method based on Machine Learning techniques has generated excellent results for small databases in general.
IoT-LPWAN-TinyML geolocation system for human voice detection in ground natural disasters
(Universidad del Norte, 2025) Soto Vergel, Angelo Joseph; Vélez Díaz, Juan Carlos; Amaya Mier, René Alejandro
A disaster is a discrete event, which can be caused by nature or by man and which exceeds local resources to respond adequately to its occurrence. Its consequences and impact directly affect society in all spheres, both economically and in physical and mental health. The management of such an event is immersed in a two-stage cycle called pre-disaster and post-disaster. The pre-disaster stage includes prevention, mitigation, and preparedness phases, and the post-disaster stage includes response, recovery, and reconstruction phases. Although all these phases are necessary and important, humanitarian organizations are usually more present in the response phase as it involves their greatest challenge; since it demands immediate activities such as the search and rescue of survivors, debris management, food supply, first aid, restoration of basic facilities and processing of site information. All these activities are part of the logistics processes in humanitarian response where time is of vital importance. In the disaster response phase, humanitarian response logistics is a specific process that focuses the execution of its activities on speed, profitability, and efficiency with the main objective of saving lives and helping vulnerable people. Historically, research in the disaster response phase has been directed at the urgency to find survivors. This emphasis on reducing the time of the search process has the potential to save lives by increasing the probability of finding victims with vital signs. Thus, with urgency and speed as the main drivers, solutions to reduce the detection and location time of human lives have become critical areas to consider. In this context, the speed of delivery of information about the affected population at the scene of the event is the most important asset. Therefore, the main question is how to know where the survivors area. Therefore, this doctoral thesis proposes to develop a geolocation system for the detection of survivors in a geophysical natural disaster through voice processing. The system will operate under the paradigm of the Internet of Things (IoT) whose connectivity between devices will be enabled by a Low-Power Wide Area Network (LPWAN). The LPWAN network will operate with the Long-Range Wide Area Network (LoRaWAN) protocol given its characteristics of long-range, low cost, low power consumption, and resistance to interference. For the nodes, a Microcontroller (µC) with a microphone as a sensor will be used, which will implement intelligent techniques for the recognition of human voice patterns in the noise present in the environment. For this, the node will be trained under the TinyML paradigm which seeks to enable Machine Learning (ML) applications in devices with limited computational resources. Furthermore, this thesis presents the following contributions: (i) the GDIMS framework, a novel system architecture integrating IoT, LPWAN, and TinyML for human life detection in disaster scenarios. This conceptual framework provides a foundation for designing scalable, low-power networks capable of continuous, low-latency monitoring, addressing critical limitations in prior works. (ii) The thesis introduces AtomicVAD, a Voice Activity Detection (VAD) model optimized for deployment on µC in resource-constrained, disaster-prone environments. AtomicVAD leverages Generalized Growing Cosine Unit (GGCU) as activation functions, significantly reducing the model’s size without losing performance, enabling efficient on-site detection without reliance on high-bandwidth or high-power resources. (iii) Field tests and practical validations were conducted, including a case study based on a disaster in Gramalote, Colombia—a rural area severely affected by landslides—refining the system’s operational parameters and demonstrating its applicability in diverse acoustic and environmental contexts. These findings support the viability of combining TinyML and IoT capabilities for resilient, scalable life detection systems in remote or disaster-affected areas.
Caracterización de la hemodinámica vascular en la zona de anclaje proximal de endoprótesis de aorta torácica usando procesamiento de imágenes de TAC y dinámica de fluidos computacional
(Universidad del Norte, 2024) Tello Portillo, Juan Pablo; Vélez Díaz, Juan Carlos; Cadena Bonfanti, Alberto
Se desarrolla un método para la simulación del flujo sanguíneo en la aorta para pacientes con reparación endovascular.  El método contempla el estudio y la planificación de cada una de las etapas, iniciando con la adquisición de las imágenes de TAC y finalizando con la caracterización del fluido en zonas del anclaje proximal de la endoprótesis. En la adquisición de las imágenes se emplea un protocolo de sincronización con el latido del corazón, con la finalidad de obtener imágenes susceptibles de evaluación y que en las etapas siguientes, se consigan modelos tridimensionales de las arterias que sean geométricamente lo más cercanas a las estructuras anatómicas reales. Se continúa con el procesamiento de las imágenes, donde se prueban diferentes métodos clásicos de filtrado y con base en medidas de calidad de imágenes, se elige el filtro de mejor desempeño. Posteriormente se realiza la segmentación y reconstrucción tridimensional de la aorta, para luego editarla y adecuarla a la etapa de simulación del fluido. El método propuesto concluye con la simulación del flujo sanguíneo y el análisis de los fenómenos hemodinámicos (perfiles de velocidad, distribuciones de presión y tensión de cizallamiento de la pared - WSS) en posiciones específicas de dos secciones transversales de la aorta torácica localizadas antes y después de la bifurcación del fluido. El método tiene como fin caracterizar el comportamiento del flujo sanguíneo e identificar patrones hemodinámicos que proporcionen información útil y ayuden al personal médico en la detección temprana de complicaciones clínicas en los pacientes de manera personalizada.
BCI applications based on artificial intelligence oriented to deep learning techniques
(Universidad del Norte, 2023) Cancino Suárez, Sandra Liliana; Schettini, Norelli; Delgado Saa, Jaime Fernando
A Brain-Computer Interface, BCI, can decode the brain signals corresponding to the intentions of individuals who have lost neuromuscular connection, to reestablish communication to control external devices. To this aim, BCI acquires brain signals as Electroencephalography (EEG) or Electrocorticography (ECoG), uses signal processing techniques and extracts features to train classifiers for providing proper control instructions. BCI development has increased in the last decades, improving its performance through the use of different signal processing techniques for feature extraction and artificial intelligence approaches for classification, such as deep learning-oriented classifiers. All of these can assure more accurate assistive systems but also can enable an analysis of the learning process of signal characteristics for the classification task. Initially, this work proposes the use of a priori knowledge and a correlation measure to select the most discriminative ECoG signal electrodes. Then, signals are processed using spatial filtering and three different types of temporal filtering, followed by a classifier made of stacked autoencoders and a softmax layer to discriminate between ECoG signals from two types of visual stimuli. Results show that the average accuracy obtained is 97% (+/- 0.02%), which is similar to state-of-the-art techniques, nevertheless, this method uses minimal prior physiological and an automated statistical technique to select some electrodes to train the classifier. Also, this work presents classifier analysis, figuring out which are the most relevant signal features useful for visual stimuli classification. The features and physiological information such as the brain areas involved are compared. Finally, this research uses Convolutional Neural Networks (CNN) or Convnets to classify 5 categories of motor tasks EEG signals. Movement-related cortical potentials (MRCPs) are used as a priori information to improve the processing of time-frequency representation of EEG signals. Results show an increase of more than 25% in average accuracy compared to a state-of-the-art method that uses the same database. In addition, an analysis of CNN or ConvNets filters and feature maps is done to and the most relevant signal characteristics that can help classify the five types of motor tasks.
Probabilistic Graphical Models for ERP-Based Brain Computer Interfaces
(Universidad del Norte, 2022) Cristancho Cuervo, Jorge Humberto; Ripoll Solano, Lácides Antonio; Delgado Saa, Jaime Fernando
An event related potential (ERP) is an electrical potential recorded from the nervous system of humans or other animals. An ERP is observed after the presentation of a stimulus. Some examples of the ERPs are P300, N400, among others. Although ERPs are used very often in neuroscience, its generation is not yet well understood and different theories have been proposed to explain the phenomena. ERPs could be generated due to changes in the alpha rhythm, an internal neural control that reset the ongoing oscillations in the brain, or separate and distinct additive neuronal phenomena. When different repetitions of the same stimuli are averaged, a coherence addition of the oscillations is obtained which explain the increase in amplitude in the signals. Two ERPs are mostly studied: N400 and P300. N400 signals arise when a subject tries to make semantic operations that support neural circuits for explicit memory. N400 potentials have been observed mostly in the rhinal cortex. P300 signals are related to attention and memory operations. When a new stimulus appears, a P300 ERP (named P3a) is generated in the frontal lobe. In contrast, when a subject perceives an expected stimulus, a P300 ERP (named P3b) is generated in the temporal – parietal areas. This implicates P3a and P3b are related, suggesting a circuit pathway between the frontal and temporal–parietal regions, whose existence has not been verified. Un potencial relacionado con un evento (ERP) es un potencial eléctrico registrado en el sistema nervioso de los seres humanos u otros animales. Un ERP se observa tras la presentación de un estímulo. Aunque los ERPs se utilizan muy a menudo en neurociencia, su generación aún no se entiende bien y se han propuesto diferentes teorías para explicar el fenómeno. Una interfaz cerebro-computador (BCI) es un sistema de comunicación en el que los mensajes o las órdenes que un sujeto envía al mundo exterior proceden de algunas señales cerebrales en lugar de los nervios y músculos periféricos. La BCI utiliza ritmos sensorimotores o señales ERP, por lo que se necesita un clasificador para distinguir entre los estímulos correctos y los incorrectos. En este trabajo, proponemos utilizar modelos probabilísticos gráficos para el modelado de la dinámica temporal y espacial de las señales cerebrales con aplicaciones a las BCIs. Los modelos gráficos han sido seleccionados por su flexibilidad y capacidad de incorporar información previa. Esta flexibilidad se ha utilizado anteriormente para modelar únicamente la dinámica temporal. Esperamos que el modelo refleje algunos aspectos del funcionamiento del cerebro relacionados con los ERPs, al incluir información espacial y temporal.
Synthesis of normal and abnormal heart sounds using Generative Adversarial Networks
(Universidad del Norte, 2022) Narváez Rosado, Pedro Juan; Percybrooks Bolívar, Winston Spencer
En esta tesis doctoral se presentan diferentes métodos propuestos para el análisis y síntesis de sonidos cardíacos normales y anormales, logrando los siguientes aportes al estado del arte: i) Se implementó un algoritmo basado en la transformada wavelet empírica (EWT) y la energía promedio normalizada de Shannon (NASE) para mejorar la etapa de segmentación automática de los sonidos cardíacos; ii) Se implementaron diferentes técnicas de extracción de características para las señales cardíacas utilizando los coeficientes cepstrales de frecuencia Mel (MFCC), los coeficientes de predicción lineal (LPC) y los valores de potencia. Además, se probaron varios modelos de Machine Learning para la clasificación automática de sonidos cardíacos normales y anormales; iii) Se diseñó un modelo basado en Redes Adversarias Generativas (GAN) para generar sonidos cardíacos sintéticos normales. Además, se implementa un algoritmo de eliminación de ruido utilizando EWT, lo que permite una disminución en la cantidad de épocas y el costo computacional que requiere el modelo GAN; iv) Finalmente, se propone un modelo basado en la arquitectura GAN, que consiste en refinar señales cardíacas sintéticas obtenidas por un modelo matemático con características de señales cardíacas reales. Este modelo se ha denominado FeaturesGAN y no requiere una gran base de datos para generar diferentes tipos de sonidos cardíacos. Cada uno de estos aportes fueron validados con diferentes métodos objetivos y comparados con trabajos publicados en el estado del arte, obteniendo resultados favorables.
Diseño de algoritmos de aprendizaje profundo basados en números complejos para la identificación de patrones anómalos en bioseñales humanas
(Universidad del Norte, 2023) Jojoa Acosta, Mario Fernando; Percybrooks Bolívar, Winston Spencer; García-zapirain Soto, Begoña
El objetivo es contribuir a la clasificación de patrones anómalos en bioseñales utilizando un enfoque novedoso. Nos enfocamos específicamente en el melanoma y en los soplos cardíacos. Usamos un estudio comparativo de dos redes de convolución en los dominios numéricos Complejo y Real. La idea es obtener un enfoque poderoso para construir sistemas portátiles para la detección temprana de enfermedades. Se eligieron dos estructuras algorítmicas similares para que no haya un sesgo determinado por la cantidad de parámetros a entrenar. Se utilizaron tres conjuntos de datos clínicos, ISIC 2017, PH2 y Pascal, para llevar a cabo los experimentos. Se realizaron pruebas de hipótesis de comparación de medias para asegurar la objetividad estadística en las conclusiones. En todos los casos, las redes de valor complejo presentaron un desempeño superior para las métricas de Precisión, Recall, F1 Score, Accuracy y Specificity en la detección de anomalías asociadas. El mejor clasificador basado en números complejos obtenido en el espacio ROC presenta una distancia euclidiana de 0,26127 con respecto al clasificador ideal, a diferencia del mejor clasificador basado en números reales, cuya distancia euclidiana al ideal es de 0,36022 para la misma tarea de detección de melanoma. La superioridad del 27,46 % en esta métrica, como en las otras reportadas en este trabajo, sugiere que las redes de valores complejos tienen una mayor capacidad para extraer características para una discriminación más eficiente en el conjunto de datos.
Adaptive methodology based on computational intelligence for time series modeling
(Universidad del Norte, 2021) Jiménez Mares, Jamer René; Quintero Monroy, Christian Giovanny
Time series processes are important in several sectors like marketing, transport, energy, telecommunications, etc. Time series forecasting tasks can help in operative and strategic tasks. Several conventional and non-conventional techniques as ARIMA models, artificial neural networks (ANN), support vector machines (SVM), Regression Tree Ensembles (RTE) or combinations of them have been used for time series modeling. The implementation of this type of techniques provides support in time series modeling, however, normally the trained models may lose performance due to the dynamic behavior of the phenomena. A methodology capable to assess the performance and maintenance of the models is necessary to guarantee the automatic adaptability in each case. Hence, in this research an adaptive methodology based on computational intelligence for time series modeling is proposed. In this case, an Auditor is developed, which allows identifying when a model must be retrained or updated before losing forecast performance. Furthermore, when the retrain process is not achieving a better performance, a new metric is proposed to choose which time series modeling technique is included in the knowledge base. The intelligent system allows building the time series model automatically, considering exogenous variables such as weather, calendar and statistical transformations to group and simplify the number of models required.
Adaptive protection in active distribution networks using local information
(Universidad del Norte, 2022) Marín Quintero, Juan Guillermo; Vélez Díaz, Juan Carlos; Orozco Henao, Cesar Augusto
This thesis proposes two adaptive protection approaches to protect MG/ADN in an autonomous way and without robust communication. The first approach presents an adaptive protection method based on an intelligent fault detector, which uses local measurements. Additionally, this solution was implemented in an online grid, where it used data-driven models running on Jetson Nano-system intended to run machine/deep learning loads at the edge. The second solution presents a decentralized adaptive protection scheme and introduces a data-driven and communication-less approach. The present solution uses an Artificial Neural Network to train Intelligent Electronic Devices as fault classifiers and brings backup protection to adjacent devices; also, it uses a cuckoo search metaheuristic to its quasi-optimal adjustment. The approaches were validated on several modified IEEE test feeders such as IEEE 13, IEEE 34, and IEEE 123. The results of the adaptive protection scheme show values of accuracy above 96% and dependability of 99%. In addition, the solution shows a correlation between the location and the combination of features and hyper-parameters. The implementation of the fault detector model into a physical low voltage network located at Universidad del Norte Colombia showed outstanding results. This test network is based on the IEEE-13 Node Test Feeder scaled to 220V.
Developing a remote sensing system based on X-band radar technology for coastal morphodynamics study
(Universidad del Norte, 2020) Navarro Ariza, Wendy Paola; Vélez Díaz, Juan Carlos; Orfila F., Alejandro
New data processing techniques are proposed for the assessment of scopes and limitations from radar-derived sea state parameters, coastline evolution and water depth estimates. Most of the raised research is focused on Colombian Caribbean coast and the Western Mediterranean Sea. First, a novel procedure to mitigate shadowing in radar images is proposed. The method compensates distortions introduced by the radar acquisition process and the power decay of the radar signal along range applying image enhancement techniques through a couple of pre-processing steps based on filtering and interpolation. Results reveal that the proposed methodology reproduces with high accuracy the sea state parameters in nearshore areas. The improvement resulting from the proposed method is assessed in a coral reef barrier, introducing a completely novel use for X-Band radar in coastal environments. So far, wave energy dissipation on a coral reef barrier has been studied by a few in-situ sensors placed in a straight line, perpendicular to the coastline, but never been described using marine radars. In this context, marine radar images are used to describe prominent features of coral reefs, including the delineation of reef morphological structure, wave energy dissipation and wave transformation processes in the lagoon of San Andres Island barrier-reef system. Results show that reef attenuates incident waves by approximately 75% due to both frictional and wave breaking dissipation, with an equivalent bottom roughness of 0.20 m and a wave friction factor of 0.18. These parameters are comparable with estimates reported in other shallow coral reef lagoons as well as at meadow canopies, obtained using in-situ measurements of wave parameters.
Wearable-based human activity recognition: from a healthcare application to a kinetic energy harvesting approach
(Universidad del Norte, 2020) Manjarrés Córdoba, José Elías; Pardo González, Mauricio
Wearable technology is changing society by becoming an essential component of daily life. Human activity recognition (HAR) is one of the most prominent research areas where wearable devices play a key role. The first major contribution to the field in this dissertation is a smart physical work- load tracking system that combines wearable-based HAR and heart rate tracking. The proposed system employs a concept from ergonomics, the Frimat’s method, to compute the physical workload from heart rate measurements within a specified time window. This dissertation includes a case of study where tracking of an individual over the course of 20 days corroborates the ability of the system to assess adaptation to an exercise routine. The second and third contributions of this dissertation point to KEH in wearable environments. The second contribution is an energy logger for wrist-worn systems, with the purpose of tracking energy generation in KEH systems during daily activities. Thus, it is possible to determine if the harvested energy is enough to power a conventional wearable device. The proposed system computes the harvested energy using the characteristics of the objective load, which in this case is a battery charger. I carried out experiments with multiple subjects to examine the generation capabilities of a commercial harvester under the conditions of human motion. This study provides insights of the performance and limitations of kinetic harvesters as battery chargers. The third contribution is a KEH-based HAR system using deep learning, data augmentation and transfer learning to outperform existing classification approaches in the KEH domain. The proposed architecture comprises convolutional neural networks (CNN) and long short-term memory networks (LSTM), which has been demonstrated to outperform other architectures found in the literature. Since deep learning classifiers require large amounts of data, and KEH datasets are limited in size, this thesis also includes the proposal of three data augmentation methods to synthesize KEH signals simulating new users. Finally, transfer learning is employed to build a system that maintains performance independent of device location or the subject wearing the device.

Examinar

Envíos recientes