Thesis Desarrollo de un modelo de clasificación de fallas en imágenes de electroluminiscencia basado en arquitectura CNN
Loading...
Date
2023
Authors
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Mecánica
Departament
Campus
Campus Santiago San Joaquín
Abstract
Una de las razones fundamentales que impulsaron la realización de este proyecto de tesis radica en la necesidad de contar con personal especializado para llevar a cabo la clasificación de celdas en la inspección visual y no destructiva de los módulos fotovoltaicos mediante el ensayo de electroluminiscencia. Esto se convierte en un incremento en los costos del tiempo empleado y en los procesos. Por lo que se quiere automatizar la clasificación de celdas.
Si bien algunos estudios han explorado la problemática de automatización de clasificación, es todavía escasa la investigación que aborda esta clasificación en multiclases. Por esta razón, se desarrolló un modelo de clasificación de distintos modos de fallas en imágenes de electroluminiscencia basado en redes neuronales convolucionales (CNN).
Inicialmente, se desarrolló un algoritmo para segmentar y extraer celdas de imágenes de módulos fotovoltaicos. A continuación, se generó una base de datos a partir de diversas fuentes. Posteriormente, se llevó a cabo un estudio de diferentes arquitecturas de redes neuronales convolucionales utilizando la base de datos generada, con el objetivo de crear un modelo de red neuronal convolucional basado en el mejor rendimiento de estas arquitecturas. Las simulaciones realizadas utilizando el modelo CNN demostraron la efectividad de esta herramienta para predecir la clasificación del estado de la celda. Además, se destaca que dicha herramienta es práctica, sencilla y de fácil comprensión.
One of the fundamental reasons that prompted the realization of this project’s thesis lies in the need to have specialized personnel to carry out the cell classification in the visual and non-destructive inspection of the photovoltaic modules through the electroluminescence test. This translates into an increase in employee time and process costs. Therefore, it requires to automate cell sorting. Although some studies have explored the problem of classification automation, the research that addresses this classification in multi-classes is still scarce. For this reason, a classification model of different failure modes in electroluminescence images based on convolutional neural networks (CNN) was developed. Initially, an algorithm was developed to segment and extract cells of images of photovoltaic modules. A database was then generated from various sources. Subsequently, a study of different convolutional neural network architectures was carried out using the generated database, with the purpose of creating a convolutional neural network (CNN) model based on the best performance of these architectures. The simulations performed using the CNN model proved the effectiveness of this tool in predicting cell state classification. In addition, it is highlighted that this tool is practical, simple and easy to understand.
One of the fundamental reasons that prompted the realization of this project’s thesis lies in the need to have specialized personnel to carry out the cell classification in the visual and non-destructive inspection of the photovoltaic modules through the electroluminescence test. This translates into an increase in employee time and process costs. Therefore, it requires to automate cell sorting. Although some studies have explored the problem of classification automation, the research that addresses this classification in multi-classes is still scarce. For this reason, a classification model of different failure modes in electroluminescence images based on convolutional neural networks (CNN) was developed. Initially, an algorithm was developed to segment and extract cells of images of photovoltaic modules. A database was then generated from various sources. Subsequently, a study of different convolutional neural network architectures was carried out using the generated database, with the purpose of creating a convolutional neural network (CNN) model based on the best performance of these architectures. The simulations performed using the CNN model proved the effectiveness of this tool in predicting cell state classification. In addition, it is highlighted that this tool is practical, simple and easy to understand.
Description
Keywords
Redes neuronales, Fotovoltaico, Base de datos, Efectividad organizacional