Este estudio aborda la aplicación de la metodología Cross-Industry Standard Process for Data Mining (CRISP-DM) en el análisis de los gases disueltos en aceite de transformadores de potencia, siendo este un componente crítico en los sistemas eléctricos. La adopción de este método estructurado de seis fases permitió evaluar de forma integral la condición de las unidades de transformación del sistema eléctrico ecuatoriano a partir del análisis de datos de inversión y expansión del sector, así como del estudio de 1 099 perfiles DGA (Dissolved Gas Analysis) obtenidos de una población de 153 transformadores ubicados en las distintas regiones de Ecuador continental. Los hallazgos descritos en este trabajo tienen el potencial de mejorar significativamente las estrategias y políticas de inversión y mantenimiento. Además, se propone la adopción de técnicas de automatización en el proceso de clasificación DGA, utilizando modelos de aprendizaje supervisado para potenciar la confiabilidad y eficiencia del servicio público de energía. Los resultados sugieren que este enfoque no solo mejora el diagnóstico dentro de las actividades de mantenimiento, sino que también proporciona una base sólida para trazar una hoja de ruta hacia una gestión predictiva de los activos, lo que se traduce en una mejora sustancial de la confiabilidad del sistema eléctrico nacional.
This study investigates the performance of four control strategies—Proportional-Derivative (PD), Feedforward-Feedback PD (FF-FB PD), Linear Quadratic Regulator (LQR), and Feedforward-Feedback LQR (FF-FB LQR)—implemented on a robotic transfemoral prosthesis. The performance metrics, including overshoot, settling time, trajectory tracking accuracy, and torque requirements, were evaluated using simulation models. The results indicate that the FF-FB LQR controller demonstrated superior performance, achieving the lowest overshoot (4.98%) and near-zero trajectory tracking error. All controllers required approximately 8.6 Nm of torque, suggesting consistent energy requirements across strategies despite their performance differences. The LQR controller exhibited the best stability, minimizing overshoot and improving overall system response. These findings highlight the advantages of feedforward-feedback control strategies, particularly the FF-FB LQR, for controlling robotic transfemoral prostheses with enhanced stability and accuracy.
With this project, the control of a transfemoral prosthesis was developed by means of a neural network. The mechatronics methodology based on the V-model was used. This document first presents an analysis of the situation of physical-motor disabilities in Ecuador and reviews some prosthesis designs to propose a local solution. The requirements that were proposed, were used as inputs for the development of the conceptual and specific design. The data acquisition system was established to perform the measurements of muscle activity and position by means of an accelerometer. A neural network consisting of a 3-input layer with 12 neurons in the hidden layer and one neuron in the output layer was trained. The output of the network corresponds to the position that the prosthesis must reach, and the developed system has a 20% error in the prediction of the flexion and extension movements of the leg. En el presente proyecto se desarrolló el control de una prótesis transfemoral por medio de una red neuronal. Se utilizó la metodología mecatrónica basada en el modelo en V. Primero se presenta un análisis a la situación de discapacidades físico-motoras de Ecuador y se revisan algunos diseños de prótesis para proponer una solución local. Se plantearon requerimientos que se usaron como insumos para el desarrollo del diseño conceptual y específico. Se estableció el sistema de adquisición de datos para realizar las mediciones de actividad muscular y posición por medio de un acelerómetro. Se entrenó una red neuronal formada de una capa de 3 entradas con 12 neuronas en la capa oculta y una neurona en la capa de salida. La salida de la red corresponde a la posición a la que debe llegar la prótesis, y el sistema desarrollado tiene un error del 20% en la predicción de los movimientos de flexión y extensión de la pierna.
The aim of this study was the integration of model-based design and Wavelet transform techniques for filtering surface electromyography (sEMG) signals. In the first stage the noises and interferences that disturb sEMG signals were analyzed to implement a digital filter in a low-cost embedded system that filters these signals. It was shown that the noises and interferences are caused by various sources. Sources of interference and noise can be divided into internal and external. Internal noise is caused by the electrodes, EMG signals of other muscles, and noise associated with the functioning of other organs such as the heart or stomach. The external noises are due to the electrical environment, the most prominent of which is the direct interference of the power hum, produced by the incorrect grounding of other devices and electromotors. For the analysis of the digital filter, sEMG signals from the biceps muscle were used when the elbow joint was at rest and during flexion and extension movements. Signals from 10 participants who did not have any atrophies or pathologies in the muscle were considered for this stage. Denoising of sEMG signals was performed using different wavelets; the smallest error was observed when using the biorthogonal wavelet 3/5 of level 6 with the soft thresholding method. The wavelet filter was implemented using the V-model, and the Processor in The Loop (PIL) tests helped to determine the characteristics of the embedded system where the digital filter was implemented. The digital filter code was implemented on an ESP32 board due to its processing speed of 328 ms.